{ "schema_version": 1, "description": "Canonical metadata for all research domains; generated pages must not maintain paper tables.", "papers": [ { "domain": "agent-research", "key": "ace-data", "title": "What Makes Good Agentic Data? An ACE Lens on Data Generation for LLM Agents", "paper_url": "https://arxiv.org/abs/2608.27260", "detail_path": "agent-research/2608.27260-ace-data/README.md", "topic": [ "Agent 数据质量" ], "first_author": "Xingshan Zeng", "first_author_affiliation": "Huawei Technologies Co., Ltd", "published": "2026-08-27", "code": null, "adapter": "ace-data" }, { "domain": "agent-research", "key": "adavdr", "title": "AdaVDR: Adaptive Tool Use and Reflection for Video Deep Research", "paper_url": "https://arxiv.org/abs/2608.25559", "detail_path": "agent-research/2608.25559-adavdr/README.md", "topic": [ "视频研究自适应工具与反思" ], "first_author": "Xintong Zhang", "first_author_affiliation": "Accio Team, Alibaba Group", "published": "2026-08-26", "code": "https://github.com/Accio-Lab/AdaVDR", "adapter": "adavdr" }, { "domain": "agent-research", "key": "agent-g2", "title": "Agent-G²: Gaussian Guidance for Agentic Reinforcement Learning", "paper_url": "https://arxiv.org/abs/2608.23318", "detail_path": "agent-research/2608.23318-agent-g2/README.md", "topic": [ "高斯 guidance Agent RL" ], "first_author": "Zixuan Wang", "first_author_affiliation": "Baidu", "published": "2026-08-24", "code": "https://github.com/ZJU-REAL/Agent-G2", "adapter": "agent-g2" }, { "domain": "agent-research", "key": "agent-lightning", "title": "Agent Lightning", "paper_url": "https://arxiv.org/abs/2508.03680", "detail_path": "agent-research/2508.03680-agent-lightning/README.md", "topic": [ "Agent RL" ], "first_author": "Xufang Luo", "first_author_affiliation": "Microsoft Research", "published": "2025-08-05", "code": "https://github.com/microsoft/agent-lightning", "adapter": "agent-lightning" }, { "domain": "agent-research", "key": "agent-opsd", "title": "AgentOPSD", "paper_url": "https://arxiv.org/abs/2608.05987", "detail_path": "agent-research/2608.05987-agent-opsd/README.md", "topic": [ "递归 turn 信用" ], "first_author": "Zi-Han Wang", "first_author_affiliation": "Tsinghua University", "published": "2026-08-06", "code": "https://github.com/ZethWang/AgentOPSD", "adapter": "agent-opsd" }, { "domain": "agent-research", "key": "agent-r1", "title": "Agent-R1", "paper_url": "https://arxiv.org/abs/2511.14460", "detail_path": "agent-research/2511.14460-agent-r1/README.md", "topic": [ "Agentic RL 基础设施" ], "first_author": "Mingyue Cheng", "first_author_affiliation": "University of Science and Technology of China", "published": "2025-11-18", "code": "https://github.com/AgentR1/Agent-R1", "adapter": "agent-r1" }, { "domain": "agent-research", "key": "agent0", "title": "Agent0", "paper_url": "https://arxiv.org/abs/2511.16043", "detail_path": "agent-research/2511.16043-agent0/README.md", "topic": [ "零数据多 Agent" ], "first_author": "Peng Xia", "first_author_affiliation": "University of North Carolina at Chapel Hill", "published": "2025-11-20", "code": "未发现官方代码", "adapter": "agent0" }, { "domain": "agent-research", "key": "agentx", "title": "AgentX: Towards Agent-Driven Self-Iteration of Industrial Recommender Systems", "paper_url": "https://arxiv.org/abs/2606.26859", "detail_path": "agent-research/2606.26859-agentx/README.md", "topic": [ "研究自动化", "工业推荐 Agent" ], "first_author": "AgentX Team", "first_author_affiliation": "Kuaishou", "published": "2026-06-26", "code": null, "adapter": "agentx" }, { "domain": "agent-research", "key": "ahead", "title": "AHEAD: Agentic Hints for Effective Agent Development", "paper_url": "https://arxiv.org/abs/2608.24114", "detail_path": "agent-research/2608.24114-ahead/README.md", "topic": [ "环境反馈提示训练" ], "first_author": "Xiaolong Jin", "first_author_affiliation": "AWS AI Labs / Purdue University", "published": "2026-08-25", "code": null, "adapter": "ahead" }, { "domain": "agent-research", "key": "art", "title": "ART", "paper_url": "https://arxiv.org/abs/2303.09014", "detail_path": "agent-research/2303.09014-art/README.md", "topic": [ "自动工具推理" ], "first_author": "Bhargavi Paranjape", "first_author_affiliation": "University of Washington", "published": "2023-03-16", "code": "https://github.com/bhargaviparanjape/language-programmes", "adapter": "art" }, { "domain": "agent-research", "key": "atomrec", "title": "AtomRec: Evolving Atomic Memory for Agentic Recommendation", "paper_url": "https://arxiv.org/abs/2609.04882", "detail_path": "agent-research/2609.04882-atomrec/README.md", "topic": [ "Agentic recommendation memory" ], "first_author": "Peiyu Hu", "first_author_affiliation": "Xi'an Jiaotong-Liverpool University", "published": "2026-09-04", "code": null, "adapter": "atomrec" }, { "domain": "agent-research", "key": "auso", "title": "AUSO: Action-Level Unified Skill Optimization from Internalization to Utilization", "paper_url": "https://arxiv.org/abs/2608.21292", "detail_path": "agent-research/2608.21292-auso/README.md", "topic": [ "动作级技能优化" ], "first_author": "Huizu Lin", "first_author_affiliation": "University of Science and Technology of China", "published": "2026-08-21", "code": "https://github.com/JordanSancholhz/Action-Skill", "adapter": "auso" }, { "domain": "agent-research", "key": "autogen", "title": "AutoGen", "paper_url": "https://arxiv.org/abs/2308.08155", "detail_path": "agent-research/2308.08155-autogen/README.md", "topic": [ "多 Agent" ], "first_author": "Qingyun Wu", "first_author_affiliation": "Microsoft Research", "published": "2023-08-16", "code": "https://github.com/microsoft/autogen", "adapter": "autogen" }, { "domain": "agent-research", "key": "autosaddler", "title": "AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces", "paper_url": "https://arxiv.org/abs/2608.23041", "detail_path": "agent-research/2608.23041-autosaddler/README.md", "topic": [ "Harness 自动优化" ], "first_author": "Sungho Park", "first_author_affiliation": "Microsoft / POSTECH", "published": "2026-08-24", "code": null, "adapter": "autosaddler" }, { "domain": "agent-research", "key": "cam-df", "title": "CAM-DF", "paper_url": "https://arxiv.org/abs/2607.27083", "detail_path": "agent-research/2607.27083-cam-df/README.md", "topic": [ "成本感知工具停止" ], "first_author": "Yicheng Feng", "first_author_affiliation": "Peking University", "published": "2026-07-29", "code": "未发现官方代码", "adapter": "cam-df" }, { "domain": "agent-research", "key": "camel", "title": "CAMEL", "paper_url": "https://arxiv.org/abs/2303.17760", "detail_path": "agent-research/2303.17760-camel/README.md", "topic": [ "多 Agent 协作" ], "first_author": "Guohao Li", "first_author_affiliation": "King Abdullah University of Science and Technology", "published": "2023-03-31", "code": "https://github.com/camel-ai/camel", "adapter": "camel" }, { "domain": "agent-research", "key": "caskg", "title": "CaSKG: Counterfactual-Causal Skill Graphs for Scalable Agent Skill Retrieval", "paper_url": "https://arxiv.org/abs/2608.25500", "detail_path": "agent-research/2608.25500-caskg/README.md", "topic": [ "反事实因果技能图" ], "first_author": "Zhiyuan Li", "first_author_affiliation": "Jilin University / Ant Group", "published": "2026-08-26", "code": "https://github.com/ZhiyuanLi218/Caskg", "adapter": "caskg" }, { "domain": "agent-research", "key": "cast", "title": "CAST", "paper_url": "https://arxiv.org/abs/2607.25308", "detail_path": "agent-research/2607.25308-cast/README.md", "topic": [ "Agentic RL / turn-level credit" ], "first_author": "Yu Wang", "first_author_affiliation": "University of Science and Technology of China", "published": "2026-07-28", "code": "https://github.com/Wloner0809/CAST", "adapter": "cast" }, { "domain": "agent-research", "key": "cera-moa", "title": "CERA-MoA: Co-Evolving Router and Agents for Mixture-of-Agents", "paper_url": "https://arxiv.org/abs/2609.18779", "detail_path": "agent-research/2609.18779-cera-moa/README.md", "topic": [ "多 Agent", "路由与协同进化" ], "first_author": "Jiaxuan Jiang", "first_author_affiliation": "IIIS, Tsinghua University", "published": "2026-09-16", "code": null, "adapter": "cera-moa", "priority": "P1" }, { "domain": "agent-research", "key": "cipo", "title": "Contextual Information Policy Optimization for Search Agents", "paper_url": "https://arxiv.org/abs/2608.06128", "detail_path": "agent-research/2608.06128-cipo/README.md", "topic": [ "搜索 Agent RL" ], "first_author": "Xingyu Guo", "first_author_affiliation": "Beihang University", "published": "2026-08-06", "code": null, "adapter": "cipo" }, { "domain": "agent-research", "key": "cobra-skills", "title": "COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization", "paper_url": "https://arxiv.org/abs/2609.11682", "detail_path": "agent-research/2609.11682-cobra-skills/README.md", "topic": [ "技能进化", "Contextual bandit" ], "first_author": "Pingchen Lu", "first_author_affiliation": "The Chinese University of Hong Kong, Shenzhen", "published": "2026-09-10", "code": "https://github.com/Jerry-LuP/COBRA-Skills", "adapter": "cobra-skills", "priority": "P0" }, { "domain": "agent-research", "key": "codegrep", "title": "CodeGrep: An RL-Trained Retrieval Agent for LLM Coding Agents", "paper_url": "https://arxiv.org/abs/2608.05886", "detail_path": "agent-research/2608.05886-codegrep/README.md", "topic": [ "代码检索 Agent" ], "first_author": "Wuya Chen", "first_author_affiliation": "NetEase Guangzhou AI Lab", "published": "2026-08-06", "code": null, "adapter": "codegrep" }, { "domain": "agent-research", "key": "coevo-mem", "title": "CoEvo-Mem", "paper_url": "https://arxiv.org/abs/2608.01739", "detail_path": "agent-research/2608.01739-coevo-mem/README.md", "topic": [ "检索—记忆共进化" ], "first_author": "Bowen Ye", "first_author_affiliation": "论文未列机构", "published": "2026-08-03", "code": "未发现官方代码", "adapter": "coevo-mem" }, { "domain": "agent-research", "key": "coskill", "title": "CoSkill: Joint Reinforcement Learning of Reasoning and Meta-Skill Agents for Hierarchical Skill Evolution", "paper_url": "https://arxiv.org/abs/2609.04865", "detail_path": "agent-research/2609.04865-coskill/README.md", "topic": [ "Hierarchical skill coevolution" ], "first_author": "Jinyuan Feng", "first_author_affiliation": "Institute of Automation, Chinese Academy of Sciences", "published": "2026-09-04", "code": "https://github.com/jinyuan-cookie/CoSkill", "adapter": "coskill" }, { "domain": "agent-research", "key": "covemem", "title": "When Memory Takes Gradients: Collaborative Vector Memory for Agentic Recommender Systems", "paper_url": "https://arxiv.org/abs/2608.26895", "detail_path": "agent-research/2608.26895-covemem/README.md", "topic": [ "可训练协同记忆" ], "first_author": "Hanchong Chen", "first_author_affiliation": "Shenzhen Technology University", "published": "2026-08-27", "code": null, "adapter": "covemem" }, { "domain": "agent-research", "key": "critic", "title": "CRITIC", "paper_url": "https://arxiv.org/abs/2305.11738", "detail_path": "agent-research/2305.11738-critic/README.md", "topic": [ "工具反馈" ], "first_author": "Zhibin Gou", "first_author_affiliation": "Microsoft", "published": "2023-05-19", "code": "https://github.com/microsoft/ProphetNet/tree/master/CRITIC", "adapter": "critic" }, { "domain": "agent-research", "key": "deeprepro", "title": "DeepRepro: State-Aware Subplanning for Paper-to-Code Reproduction in Evolving Repositories", "paper_url": "https://arxiv.org/abs/2608.26557", "detail_path": "agent-research/2608.26557-deeprepro/README.md", "topic": [ "论文到代码状态子规划" ], "first_author": "Hongru Song", "first_author_affiliation": "Institute of Computing Technology, Chinese Academy of Sciences", "published": "2026-08-27", "code": "https://github.com/ruyisy/DeepRepro", "adapter": "deeprepro" }, { "domain": "agent-research", "key": "deepresearcher", "title": "DeepResearcher", "paper_url": "https://arxiv.org/abs/2504.03160", "detail_path": "agent-research/2504.03160-deepresearcher/README.md", "topic": [ "深度研究 RL" ], "first_author": "Yuxiang Zheng", "first_author_affiliation": "HKU", "published": "2025-04-04", "code": "https://github.com/GAIR-NLP/DeepResearcher", "adapter": "deepresearcher" }, { "domain": "agent-research", "key": "dependency-refinement", "title": "Dependency-Aware Trajectory Refinement for Efficient Multi-Turn Agent Fine-Tuning", "paper_url": "https://arxiv.org/abs/2609.18417", "detail_path": "agent-research/2609.18417-dependency-refinement/README.md", "topic": [ "轨迹精炼", "Agent 微调" ], "first_author": "Zhuo Chen", "first_author_affiliation": "ShanghaiTech University", "published": "2026-09-16", "code": "https://github.com/Alibaba-NLP/VLLM-KB", "adapter": "dependency-refinement", "priority": "P1" }, { "domain": "agent-research", "key": "ecdysis", "title": "Ecdysis: Efficient and Effective Training of Runtime Harnesses for LLM Agents", "paper_url": "https://arxiv.org/abs/2609.11677", "detail_path": "agent-research/2609.11677-ecdysis/README.md", "topic": [ "Harness 自进化", "失败诊断" ], "first_author": "Ruiqing Yue", "first_author_affiliation": "Chengdu Institute of Computer Applications, Chinese Academy of Sciences", "published": "2026-09-10", "code": "https://github.com/cuiyu-ai/Ecdysis", "adapter": "ecdysis", "priority": "P0" }, { "domain": "agent-research", "key": "envace", "title": "EnvACE", "paper_url": "https://arxiv.org/abs/2608.06197", "detail_path": "agent-research/2608.06197-envace/README.md", "topic": [ "环境 rehearsal / Agent RL" ], "first_author": "Zishan Xu", "first_author_affiliation": "Shanghai Jiao Tong", "published": "2026-08-06", "code": "https://github.com/Within-yao/EnvACE", "adapter": "envace" }, { "domain": "agent-research", "key": "evoharness-rl", "title": "EvoHarness-RL: Learning Self-Evolving Runtime Harness for Long-Horizon LLM Agents", "paper_url": "https://arxiv.org/abs/2608.05446", "detail_path": "agent-research/2608.05446-evoharness-rl/README.md", "topic": [ "Harness policy RL" ], "first_author": "Xuying Ning", "first_author_affiliation": "University of Illinois Urbana–Champaign / Meta AI", "published": "2026-08-05", "code": null, "adapter": "evoharness-rl" }, { "domain": "agent-research", "key": "evoskill-gui", "title": "Reflect, Revise, Reuse: Training-Free Skill Evolution for GUI Agents", "paper_url": "https://arxiv.org/abs/2609.17653", "detail_path": "agent-research/2609.17653-evoskill-gui/README.md", "topic": [ "GUI Agent", "技能进化" ], "first_author": "Bofan Chen", "first_author_affiliation": "Zhejiang University", "published": "2026-09-15", "code": "https://github.com/ZJU-REAL/EvoSkill-GUI", "adapter": "evoskill-gui", "priority": "P0" }, { "domain": "agent-research", "key": "feedback-scaffold", "title": "Environments as Scaffold: Enriching Feedback to Bootstrap Self-Evolving Agents in Long-Horizon Tasks", "paper_url": "https://arxiv.org/abs/2609.08404", "detail_path": "agent-research/2609.08404-feedback-scaffold/README.md", "topic": [ "环境反馈脚手架" ], "first_author": "Hongbang Yuan", "first_author_affiliation": "Fudan University", "published": "2026-09-08", "code": "https://github.com/HongbangYuan/EnvAsScaffold", "adapter": "feedback-scaffold" }, { "domain": "agent-research", "key": "fuse-evaluator", "title": "Verifiable Social Reasoning for LLM Assistants", "paper_url": "https://arxiv.org/abs/2609.17496", "detail_path": "agent-research/2609.17496-fuse-evaluator/README.md", "topic": [ "Agent 评测", "社会推理" ], "first_author": "Amir Taubenfeld", "first_author_affiliation": "Google Research / Hebrew University of Jerusalem / University of Cambridge", "published": "2026-09-15", "code": null, "adapter": "fuse-evaluator", "priority": "P0" }, { "domain": "agent-research", "key": "gaia", "title": "GAIA", "paper_url": "https://arxiv.org/abs/2311.12983", "detail_path": "agent-research/2311.12983-gaia/README.md", "topic": [ "通用 Agent 评测" ], "first_author": "Grégoire Mialon", "first_author_affiliation": "Meta AI", "published": "2023-11-21", "code": "https://huggingface.co/gaia-benchmark", "adapter": "gaia" }, { "domain": "agent-research", "key": "generative-agents", "title": "Generative Agents", "paper_url": "https://arxiv.org/abs/2304.03442", "detail_path": "agent-research/2304.03442-generative-agents/README.md", "topic": [ "记忆与反思" ], "first_author": "Joon Sung Park", "first_author_affiliation": "Stanford University", "published": "2023-04-07", "code": "https://github.com/joonspk-research/generative_agents", "adapter": "generative-agents" }, { "domain": "agent-research", "key": "gigpo", "title": "GiGPO", "paper_url": "https://arxiv.org/abs/2505.10978", "detail_path": "agent-research/2505.10978-gigpo/README.md", "topic": [ "Agent group credit" ], "first_author": "Lang Feng", "first_author_affiliation": "Nanyang Technological University", "published": "2025-05-16", "code": "未发现官方代码", "adapter": "gigpo" }, { "domain": "agent-research", "key": "grounded-memory", "title": "Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents", "paper_url": "https://arxiv.org/abs/2609.11060", "detail_path": "agent-research/2609.11060-grounded-memory/README.md", "topic": [ "Agent 记忆", "环境探测" ], "first_author": "Susheel Suresh", "first_author_affiliation": "Microsoft", "published": "2026-09-09", "code": null, "adapter": "grounded-memory", "priority": "P0" }, { "domain": "agent-research", "key": "grsd", "title": "Group-Reflective Self-Distillation", "paper_url": "https://arxiv.org/abs/2607.28076", "detail_path": "agent-research/2607.28076-grsd/README.md", "topic": [ "Agent group credit" ], "first_author": "Binbin Zheng", "first_author_affiliation": "Baidu Inc.", "published": "2026-07-30", "code": "https://github.com/BinbZheng1/GRSD", "adapter": "grsd" }, { "domain": "agent-research", "key": "gse", "title": "Learning Globally Reusable Skills for Coding Agents", "paper_url": "https://arxiv.org/abs/2608.06153", "detail_path": "agent-research/2608.06153-gse/README.md", "topic": [ "全局技能进化" ], "first_author": "Chen Yang", "first_author_affiliation": "Tianjin University", "published": "2026-08-06", "code": null, "adapter": "gse" }, { "domain": "agent-research", "key": "harness-bandit", "title": "HarnessBandit: Joint Learnability-Transferability Scheduling for Multi-Harness Agentic Reinforcement Learning", "paper_url": "https://arxiv.org/abs/2609.13739", "detail_path": "agent-research/2609.13739-harness-bandit/README.md", "topic": [ "Agent RL", "Harness 调度" ], "first_author": "Hongliang Wei", "first_author_affiliation": "Harbin Institute of Technology / Alibaba Cloud", "published": "2026-09-12", "code": null, "adapter": "harness-bandit", "priority": "P1" }, { "domain": "agent-research", "key": "harness-design-study", "title": "An Empirical Study of Harness Design for Coding Agents", "paper_url": "https://arxiv.org/abs/2609.20804", "detail_path": "agent-research/2609.20804-harness-design-study/README.md", "topic": [ "编程 Agent", "Harness 设计" ], "first_author": "Run-Ze Fan", "first_author_affiliation": "University of Massachusetts Amherst", "published": "2026-09-17", "code": null, "adapter": "harness-design-study", "priority": "P1" }, { "domain": "agent-research", "key": "harnesslens", "title": "Verify Smarter, Evolve Further: Efficient Harness Evolution through Behavior-Aware Verification", "paper_url": "https://arxiv.org/abs/2608.27311", "detail_path": "agent-research/2608.27311-harnesslens/README.md", "topic": [ "行为相关 Harness 验证" ], "first_author": "Jinghan Xu", "first_author_affiliation": "Fudan University", "published": "2026-08-27", "code": "https://github.com/jhxu5214/HarnessLens", "adapter": "harnesslens" }, { "domain": "agent-research", "key": "harnessopt-bench", "title": "HarnessOpt-Bench: Evaluating LLMs at Harness Optimization", "paper_url": "https://arxiv.org/abs/2608.06301", "detail_path": "agent-research/2608.06301-harnessopt-bench/README.md", "topic": [ "Harness 优化评测" ], "first_author": "Varun Ursekar", "first_author_affiliation": "Scale AI", "published": "2026-08-06", "code": null, "adapter": "harnessopt-bench" }, { "domain": "agent-research", "key": "hindsearch", "title": "HindSearch: Trajectory-Level Hindsight Critique for Search-Augmented Reinforcement Learning", "paper_url": "https://arxiv.org/abs/2608.01597", "detail_path": "agent-research/2608.01597-hindsearch/README.md", "topic": [ "搜索轨迹 hindsight" ], "first_author": "Haowei Liu", "first_author_affiliation": "Santa Clara University", "published": "2026-08-03", "code": "https://anonymous.4open.science/r/hindsearch-anon-EBDC", "adapter": "hindsearch" }, { "domain": "agent-research", "key": "hiskill", "title": "HiSkill", "paper_url": "https://arxiv.org/abs/2607.25853", "detail_path": "agent-research/2607.25853-hiskill/README.md", "topic": [ "Hierarchical skill memory" ], "first_author": "Yu Hao", "first_author_affiliation": "Beijing University of Posts and Telecommunications", "published": "2026-07-28", "code": "https://github.com/BUPT-GAMMA/HiSkill", "adapter": "hiskill" }, { "domain": "agent-research", "key": "hugginggpt", "title": "HuggingGPT", "paper_url": "https://arxiv.org/abs/2303.17580", "detail_path": "agent-research/2303.17580-hugginggpt/README.md", "topic": [ "专家模型编排" ], "first_author": "Yongliang Shen", "first_author_affiliation": "Zhejiang University", "published": "2023-03-30", "code": "https://github.com/microsoft/JARVIS", "adapter": "hugginggpt" }, { "domain": "agent-research", "key": "hymem", "title": "HyMem: Hierarchical Context Management for Long-Horizon Agents via Information Isolation", "paper_url": "https://arxiv.org/abs/2608.15703", "detail_path": "agent-research/2608.15703-hymem/README.md", "topic": [ "记忆", "长时序 Agent" ], "first_author": "XinQi Wang", "first_author_affiliation": "Institute of Automation, Chinese Academy of Sciences", "published": "2026-08-16", "code": null, "adapter": "hymem" }, { "domain": "agent-research", "key": "hyperagent", "title": "HyperAgent: Planning and Acting over Tool-Schema Hypergraphs for Tool-Use LLM Agents", "paper_url": "https://arxiv.org/abs/2608.02650", "detail_path": "agent-research/2608.02650-hyperagent/README.md", "topic": [ "工具规划", "超图" ], "first_author": "Zian Zhai", "first_author_affiliation": "University of New South Wales", "published": "2026-07-31", "code": null, "adapter": "hyperagent" }, { "domain": "agent-research", "key": "interactive-memory", "title": "Interactive Memory Learning for Long-Term Conversations", "paper_url": "https://arxiv.org/abs/2609.17088", "detail_path": "agent-research/2609.17088-interactive-memory/README.md", "topic": [ "长期记忆", "多 Agent 强化学习" ], "first_author": "Cai Ke", "first_author_affiliation": "Harbin Institute of Technology, Shenzhen / Pengcheng Laboratory", "published": "2026-09-15", "code": null, "adapter": "interactive-memory", "priority": "P1" }, { "domain": "agent-research", "key": "jit-agent", "title": "JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution", "paper_url": "https://arxiv.org/abs/2608.25593", "detail_path": "agent-research/2608.25593-jit-agent/README.md", "topic": [ "Harness 即时生成与进化" ], "first_author": "Guibin Zhang", "first_author_affiliation": "LV-NUS Lab", "published": "2026-08-26", "code": "https://github.com/bingreeky/JIT", "adapter": "jit-agent" }, { "domain": "agent-research", "key": "lats", "title": "LATS", "paper_url": "https://arxiv.org/abs/2310.04406", "detail_path": "agent-research/2310.04406-lats/README.md", "topic": [ "Agent 搜索" ], "first_author": "Andy Zhou", "first_author_affiliation": "University of Illinois Urbana-Champaign", "published": "2023-10-06", "code": "https://github.com/lapisrocks/LanguageAgentTreeSearch", "adapter": "lats" }, { "domain": "agent-research", "key": "legomem", "title": "LEGOMem", "paper_url": "https://arxiv.org/abs/2510.04851", "detail_path": "agent-research/2510.04851-legomem/README.md", "topic": [ "过程记忆" ], "first_author": "Dongge Han", "first_author_affiliation": "Microsoft Research", "published": "2025-10-06", "code": "未发现官方代码", "adapter": "legomem" }, { "domain": "agent-research", "key": "loongreflect", "title": "LoongReflect: Boosting Long-Horizon Reflection in Search Agents via Global Perspective Distillation", "paper_url": "https://arxiv.org/abs/2608.11967", "detail_path": "agent-research/2608.11967-loongreflect/README.md", "topic": [ "反思", "Agentic RL" ], "first_author": "Zhixin Zhang", "first_author_affiliation": "Authors did not disclose affiliation", "published": "2026-08-12", "code": null, "adapter": "loongreflect" }, { "domain": "agent-research", "key": "loop", "title": "LOOP", "paper_url": "https://arxiv.org/abs/2502.01600", "detail_path": "agent-research/2502.01600-loop/README.md", "topic": [ "长时程 Agent RL" ], "first_author": "Kevin Chen", "first_author_affiliation": "Apple", "published": "2025-02-03", "code": "https://github.com/apple/ml-loop", "adapter": "loop" }, { "domain": "agent-research", "key": "manta", "title": "MANTA: Multi-Agent Network Topology Adaptation for Self-Evolving Multi-Agent Systems", "paper_url": "https://arxiv.org/abs/2607.28527", "detail_path": "agent-research/2607.28527-manta/README.md", "topic": [ "多 Agent", "自进化" ], "first_author": "MaoXun Huang", "first_author_affiliation": "Cornell University", "published": "2026-07-30", "code": null, "adapter": "manta" }, { "domain": "agent-research", "key": "maple", "title": "MAPLE: Memory-Augmented Planning with Language and Evolution", "paper_url": "https://arxiv.org/abs/2609.11636", "detail_path": "agent-research/2609.11636-maple/README.md", "topic": [ "持久化优化规划" ], "first_author": "Kesheng Chen", "first_author_affiliation": "Harbin Institute of Technology, Shenzhen", "published": "2026-09-10", "code": "https://github.com/xin8coder/MAPLE", "adapter": "maple" }, { "domain": "agent-research", "key": "memento-skills", "title": "Memento-Skills", "paper_url": "https://arxiv.org/abs/2603.18743", "detail_path": "agent-research/2603.18743-memento-skills/README.md", "topic": [ "技能设计" ], "first_author": "Huichi Zhou", "first_author_affiliation": "Memento Team", "published": "2026-03-19", "code": "https://github.com/Memento-Teams/Memento-Skills", "adapter": "memento-skills" }, { "domain": "agent-research", "key": "memforest", "title": "MemForest: Efficient Agent Memory Management via EventTree Partitioning and Progressive Merging", "paper_url": "https://arxiv.org/abs/2609.08273", "detail_path": "agent-research/2609.08273-memforest/README.md", "topic": [ "事件树记忆压缩" ], "first_author": "Junxi Wang", "first_author_affiliation": "Shanghai Jiao Tong University", "published": "2026-09-08", "code": "https://github.com/Celina-love-sweet/MemForest", "adapter": "memforest" }, { "domain": "agent-research", "key": "memgpt", "title": "MemGPT", "paper_url": "https://arxiv.org/abs/2310.08560", "detail_path": "agent-research/2310.08560-memgpt/README.md", "topic": [ "虚拟上下文" ], "first_author": "Charles Packer", "first_author_affiliation": "University of California, Berkeley", "published": "2023-10-12", "code": "https://github.com/letta-ai/letta", "adapter": "memgpt" }, { "domain": "agent-research", "key": "memorycpt", "title": "MemoryCPT: An End-to-End Agent Memory Framework for Cost-Performance Trade-off", "paper_url": "https://arxiv.org/abs/2608.04843", "detail_path": "agent-research/2608.04843-memorycpt/README.md", "topic": [ "端到端 Agent 记忆" ], "first_author": "Songxin Lei", "first_author_affiliation": "Hong Kong University of Science and Technology / Tencent LIGHTSPEED STUDIOS", "published": "2026-08-05", "code": null, "adapter": "memorycpt" }, { "domain": "agent-research", "key": "memskill", "title": "MemSkill", "paper_url": "https://arxiv.org/abs/2602.02474", "detail_path": "agent-research/2602.02474-memskill/README.md", "topic": [ "记忆技能" ], "first_author": "Haozhen Zhang", "first_author_affiliation": "Nanyang Technological University", "published": "2026-02-02", "code": "https://github.com/ViktorAxelsen/MemSkill", "adapter": "memskill" }, { "domain": "agent-research", "key": "memtool", "title": "MemTool", "paper_url": "https://arxiv.org/abs/2507.21428", "detail_path": "agent-research/2507.21428-memtool/README.md", "topic": [ "工具记忆" ], "first_author": "Elias Lumer", "first_author_affiliation": "PricewaterhouseCoopers Commercial Technology and Innovation Office", "published": "2025-07-29", "code": "未发现官方代码", "adapter": "memtool" }, { "domain": "agent-research", "key": "metagpt", "title": "MetaGPT", "paper_url": "https://arxiv.org/abs/2308.00352", "detail_path": "agent-research/2308.00352-metagpt/README.md", "topic": [ "多 Agent 软件工程" ], "first_author": "Sirui Hong", "first_author_affiliation": "DeepWisdom", "published": "2023-08-01", "code": "https://github.com/FoundationAgents/MetaGPT", "adapter": "metagpt" }, { "domain": "agent-research", "key": "mrkl", "title": "MRKL", "paper_url": "https://arxiv.org/abs/2205.00445", "detail_path": "agent-research/2205.00445-mrkl/README.md", "topic": [ "神经符号路由" ], "first_author": "Ehud Karpas", "first_author_affiliation": "AI21 Labs", "published": "2022-05-01", "code": "未发现官方代码", "adapter": "mrkl" }, { "domain": "agent-research", "key": "mua-rl", "title": "MUA-RL", "paper_url": "https://arxiv.org/abs/2508.18669", "detail_path": "agent-research/2508.18669-mua-rl/README.md", "topic": [ "多轮用户 Agent RL" ], "first_author": "Weikang Zhao", "first_author_affiliation": "Meituan", "published": "2025-08-26", "code": "https://github.com/zzwkk/MUA-RL", "adapter": "mua-rl" }, { "domain": "agent-research", "key": "multi-harness-rl", "title": "What Does Multi-Harness RL Learn? Credit Assignment and Portability in Coding Agents", "paper_url": "https://arxiv.org/abs/2609.04518", "detail_path": "agent-research/2609.04518-multi-harness-rl/README.md", "topic": [ "Multi-harness RL audit" ], "first_author": "Chenqian Le", "first_author_affiliation": "New York University", "published": "2026-09-03", "code": null, "adapter": "multi-harness-rl" }, { "domain": "agent-research", "key": "ocsd", "title": "OCSD", "paper_url": "https://arxiv.org/abs/2608.04788", "detail_path": "agent-research/2608.04788-ocsd/README.md", "topic": [ "观测校准蒸馏" ], "first_author": "Yi Yang", "first_author_affiliation": "Nanjing University", "published": "2026-08-05", "code": "https://github.com/yiy1x/OCSD", "adapter": "ocsd" }, { "domain": "agent-research", "key": "openhands", "title": "OpenHands", "paper_url": "https://arxiv.org/abs/2407.16741", "detail_path": "agent-research/2407.16741-openhands/README.md", "topic": [ "通用软件 Agent" ], "first_author": "Xingyao Wang", "first_author_affiliation": "All-Hands-AI", "published": "2024-07-23", "code": "https://github.com/All-Hands-AI/OpenHands", "adapter": "openhands" }, { "domain": "agent-research", "key": "openloopevolve", "title": "OpenLoopEvolve: A Verifiable Self-Evolution Framework for Loop Policies in Long-Horizon Complex Tasks", "paper_url": "https://arxiv.org/abs/2608.09380", "detail_path": "agent-research/2608.09380-openloopevolve/README.md", "topic": [ "自进化", "策略治理" ], "first_author": "Siqi Wang", "first_author_affiliation": "Tsinghua University", "published": "2026-08-10", "code": "https://github.com/yoyoshikc/OpenLoopEvolve", "adapter": "openloopevolve" }, { "domain": "agent-research", "key": "os-shepherd", "title": "OSReward / OS-Shepherd", "paper_url": "https://arxiv.org/abs/2607.28609", "detail_path": "agent-research/2607.28609-osreward/README.md", "topic": [ "CUA Reward 评测" ], "first_author": "Qiushi Sun", "first_author_affiliation": "The University of Hong Kong", "published": "2026-07-30", "code": "https://os-copilot.github.io/OSReward-Home/", "adapter": "os-shepherd" }, { "domain": "agent-research", "key": "pal", "title": "PAL", "paper_url": "https://arxiv.org/abs/2211.10435", "detail_path": "agent-research/2211.10435-pal/README.md", "topic": [ "程序推理" ], "first_author": "Luyu Gao", "first_author_affiliation": "Carnegie Mellon University", "published": "2022-11-18", "code": "https://github.com/reasoning-machines/pal", "adapter": "pal" }, { "domain": "agent-research", "key": "pearl", "title": "PEARL", "paper_url": "https://arxiv.org/abs/2601.20439", "detail_path": "agent-research/2601.20439-pearl/README.md", "topic": [ "规划强化学习" ], "first_author": "Qihao Wang", "first_author_affiliation": "中国科学院信息工程研究所", "published": "2026-01-28", "code": "未发现官方代码", "adapter": "pearl" }, { "domain": "agent-research", "key": "planpo", "title": "PlanPO: Group Planning-Aware Policy Optimization for Multi-Turn Agentic LLMs", "paper_url": "https://arxiv.org/abs/2608.17289", "detail_path": "agent-research/2608.17289-planpo/README.md", "topic": [ "Agentic RL", "规划" ], "first_author": "Dayang Liang", "first_author_affiliation": "Xiamen University", "published": "2026-08-18", "code": null, "adapter": "planpo" }, { "domain": "agent-research", "key": "pmcoder", "title": "Coupling Planning with Episodic Memory in LLM Agents for Software Issue Resolution", "paper_url": "https://arxiv.org/abs/2608.06811", "detail_path": "agent-research/2608.06811-pmcoder/README.md", "topic": [ "代码 Agent", "记忆与规划" ], "first_author": "Jiahao Zhang", "first_author_affiliation": "Vanderbilt University", "published": "2026-08-07", "code": null, "adapter": "pmcoder" }, { "domain": "agent-research", "key": "procedural-graphs", "title": "Procedural Graphs: Self-Evolving Execution Structures for LLM Agents", "paper_url": "https://arxiv.org/abs/2609.09153", "detail_path": "agent-research/2609.09153-procedural-graphs/README.md", "topic": [ "自进化程序图" ], "first_author": "Yuxing Lu", "first_author_affiliation": "Google", "published": "2026-09-08", "code": null, "adapter": "procedural-graphs" }, { "domain": "agent-research", "key": "progrouter", "title": "ProgRouter: Online Progress-Guided Orchestration for Multi-Agent LLM Workflows under Quality-Cost Tradeoffs", "paper_url": "https://arxiv.org/abs/2608.25992", "detail_path": "agent-research/2608.25992-progrouter/README.md", "topic": [ "在线进展成本路由" ], "first_author": "Songyuan Li", "first_author_affiliation": "Aston University", "published": "2026-08-26", "code": null, "adapter": "progrouter" }, { "domain": "agent-research", "key": "prompts", "title": "PROMPTS: Performance Optimization via Multi-Agent Planning for LLM Training and Serving", "paper_url": "https://research.google/pubs/prompts-performance-optimization-via-multi-agent-planning-for-llm-training-and-serving/", "detail_path": "agent-research/mlsys2026-prompts/README.md", "topic": [ "训练服务优化", "多 Agent 性能诊断" ], "first_author": "Yuran Ding", "first_author_affiliation": "University of Maryland / Google", "published": "2026-05-18", "code": null, "adapter": "prompts", "priority": "P0" }, { "domain": "agent-research", "key": "ragen", "title": "RAGEN", "paper_url": "https://arxiv.org/abs/2504.20073", "detail_path": "agent-research/2504.20073-ragen/README.md", "topic": [ "多轮 Agent RL" ], "first_author": "Zihan Wang", "first_author_affiliation": "Northwestern", "published": "2025-04-24", "code": "https://github.com/RAGEN-AI/RAGEN", "adapter": "ragen" }, { "domain": "agent-research", "key": "react", "title": "ReAct", "paper_url": "https://arxiv.org/abs/2210.03629", "detail_path": "agent-research/2210.03629-react/README.md", "topic": [ "推理与行动" ], "first_author": "Shunyu Yao", "first_author_affiliation": "Princeton University", "published": "2022-10-06", "code": "https://github.com/ysymyth/ReAct", "adapter": "react" }, { "domain": "agent-research", "key": "redevoagent", "title": "RedEvoAgent: Automatic Red-Teaming Agent with Experience-Driven Skill Evolution", "paper_url": "https://arxiv.org/abs/2608.27439", "detail_path": "agent-research/2608.27439-redevoagent/README.md", "topic": [ "红队技能进化" ], "first_author": "Junjie Zhang", "first_author_affiliation": "City University of Hong Kong", "published": "2026-08-27", "code": null, "adapter": "redevoagent" }, { "domain": "agent-research", "key": "reflexion", "title": "Reflexion", "paper_url": "https://arxiv.org/abs/2303.11366", "detail_path": "agent-research/2303.11366-reflexion/README.md", "topic": [ "自我反思" ], "first_author": "Noah Shinn", "first_author_affiliation": "Northeastern University", "published": "2023-03-20", "code": "https://github.com/noahshinn/reflexion", "adapter": "reflexion" }, { "domain": "agent-research", "key": "repoatlas", "title": "RepoAtlas: Guiding Coding Agents via Evolving Multimodal Repository Views", "paper_url": "https://arxiv.org/abs/2609.16936", "detail_path": "agent-research/2609.16936-repoatlas/README.md", "topic": [ "编程 Agent", "多模态仓库视图" ], "first_author": "Yunxiang Zhang", "first_author_affiliation": "Beihang University", "published": "2026-09-15", "code": null, "adapter": "repoatlas", "priority": "P0" }, { "domain": "agent-research", "key": "retool", "title": "ReTool", "paper_url": "https://arxiv.org/abs/2504.11536", "detail_path": "agent-research/2504.11536-retool/README.md", "topic": [ "推理中工具调用" ], "first_author": "Jiazhan Feng", "first_author_affiliation": "ByteDance Seed", "published": "2025-04-15", "code": "未发现官方代码", "adapter": "retool" }, { "domain": "agent-research", "key": "rewoo", "title": "ReWOO", "paper_url": "https://arxiv.org/abs/2305.18323", "detail_path": "agent-research/2305.18323-rewoo/README.md", "topic": [ "解耦规划" ], "first_author": "Binfeng Xu", "first_author_affiliation": "Microsoft Research", "published": "2023-05-29", "code": "https://github.com/billxbf/ReWOO", "adapter": "rewoo" }, { "domain": "agent-research", "key": "rtpo", "title": "RTPO: Reverse-Turn Policy Optimization for Stabilizing Agentic RL Training", "paper_url": "https://arxiv.org/abs/2608.18682", "detail_path": "agent-research/2608.18682-rtpo/README.md", "topic": [ "Agentic RL", "credit assignment" ], "first_author": "Yugu Li", "first_author_affiliation": "Adelaide University", "published": "2026-08-19", "code": null, "adapter": "rtpo" }, { "domain": "agent-research", "key": "sage", "title": "SAGE", "paper_url": "https://arxiv.org/abs/2512.17102", "detail_path": "agent-research/2512.17102-sage/README.md", "topic": [ "技能库强化学习" ], "first_author": "Jiongxiao Wang", "first_author_affiliation": "University of Wisconsin–Madison", "published": "2025-12-18", "code": "未发现官方代码", "adapter": "sage" }, { "domain": "agent-research", "key": "sapo", "title": "SAPO: Single-Rollout Autoregressive Policy Optimization for Agentic Reinforcement Learning", "paper_url": "https://arxiv.org/abs/2608.19842", "detail_path": "agent-research/2608.19842-sapo/README.md", "topic": [ "Agentic RL", "credit assignment" ], "first_author": "Dayang Liang", "first_author_affiliation": "Xiamen University", "published": "2026-08-20", "code": null, "adapter": "sapo" }, { "domain": "agent-research", "key": "saycan", "title": "SayCan", "paper_url": "https://arxiv.org/abs/2204.01691", "detail_path": "agent-research/2204.01691-saycan/README.md", "topic": [ "具身规划" ], "first_author": "Michael Ahn", "first_author_affiliation": "Robotics at Google", "published": "2022-04-04", "code": "https://github.com/google-research/google-research/tree/master/saycan", "adapter": "saycan" }, { "domain": "agent-research", "key": "sciencebuddy", "title": "ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents", "paper_url": "https://arxiv.org/abs/2609.17523", "detail_path": "agent-research/2609.17523-sciencebuddy/README.md", "topic": [ "科学 Agent", "递归自进化" ], "first_author": "Shuhan Xue", "first_author_affiliation": "Gen-Verse research collaboration", "published": "2026-09-15", "code": "https://github.com/Gen-Verse/ScienceBuddy", "adapter": "sciencebuddy", "priority": "P1" }, { "domain": "agent-research", "key": "search-r1", "title": "Search-R1", "paper_url": "https://arxiv.org/abs/2503.09516", "detail_path": "agent-research/2503.09516-search-r1/README.md", "topic": [ "搜索 Agent RL" ], "first_author": "Bowen Jin", "first_author_affiliation": "University of Illinois Urbana-Champaign", "published": "2025-03-12", "code": "https://github.com/PeterGriffinJin/Search-R1", "adapter": "search-r1" }, { "domain": "agent-research", "key": "searchatlas", "title": "SearchAtlas: Analyzing Agentic Search Strategies via Evidential Query Graphs", "paper_url": "https://arxiv.org/abs/2609.10901", "detail_path": "agent-research/2609.10901-searchatlas/README.md", "topic": [ "搜索 Agent", "过程审计" ], "first_author": "Jiacheng Sang", "first_author_affiliation": "Duke University", "published": "2026-09-09", "code": "https://github.com/DukeNLP/SearchAtlas", "adapter": "searchatlas", "priority": "P1" }, { "domain": "agent-research", "key": "searl", "title": "SEARL", "paper_url": "https://arxiv.org/abs/2604.07791", "detail_path": "agent-research/2604.07791-searl/README.md", "topic": [ "策略—工具图共进化" ], "first_author": "Xinshun Feng", "first_author_affiliation": "Shanghai AI Laboratory", "published": "2026-04-09", "code": "未发现官方代码", "adapter": "searl" }, { "domain": "agent-research", "key": "seed", "title": "SEED", "paper_url": "https://arxiv.org/abs/2607.14777", "detail_path": "agent-research/2607.14777-seed/README.md", "topic": [ "Agentic RL / hindsight skill" ], "first_author": "Jinyang Wu", "first_author_affiliation": "Tsinghua University", "published": "2026-07-16", "code": "https://github.com/jinyangwu/SEED", "adapter": "seed" }, { "domain": "agent-research", "key": "self-refine", "title": "Self-Refine", "paper_url": "https://arxiv.org/abs/2303.17651", "detail_path": "agent-research/2303.17651-self-refine/README.md", "topic": [ "自我迭代" ], "first_author": "Aman Madaan", "first_author_affiliation": "Carnegie Mellon University", "published": "2023-03-30", "code": "https://github.com/madaan/self-refine", "adapter": "self-refine" }, { "domain": "agent-research", "key": "silr", "title": "SiLR: Structure-Preserving Admission and Process Reward for LLM Tool Agents", "paper_url": "https://arxiv.org/abs/2609.04629", "detail_path": "agent-research/2609.04629-silr/README.md", "topic": [ "Structure-preserving verifier and reward" ], "first_author": "Chenyu Zhou", "first_author_affiliation": "Institute of Science Tokyo", "published": "2026-09-03", "code": null, "adapter": "silr" }, { "domain": "agent-research", "key": "sinkflex-rl", "title": "Efficient Reinforcement Learning for Long-Horizon Tool-Use Agentic Tasks", "paper_url": "https://arxiv.org/abs/2608.10357", "detail_path": "agent-research/2608.10357-sinkflex-rl/README.md", "topic": [ "Agentic RL / efficient long context" ], "first_author": "Zelei Cheng", "first_author_affiliation": "Capital One AI Foundations", "published": "2026-08-11", "code": null, "adapter": "sinkflex-rl" }, { "domain": "agent-research", "key": "skill-retention", "title": "When Synthetic Data Hurts: On Catastrophic Forgetting in Skill Retrieval for LLM Agents", "paper_url": "https://arxiv.org/abs/2609.10750", "detail_path": "agent-research/2609.10750-skill-retention/README.md", "topic": [ "技能检索", "灾难性遗忘" ], "first_author": "Syed Shariyar Murtaza", "first_author_affiliation": "Manulife", "published": "2026-09-09", "code": "https://github.com/manulife-ai/emnlp2026/", "adapter": "skill-retention", "priority": "P1" }, { "domain": "agent-research", "key": "skillforge", "title": "SkillForge: Automated Skill Discovery and Refinement for Tool-Using Agents", "paper_url": "https://arxiv.org/abs/2608.24747", "detail_path": "agent-research/2608.24747-skillforge/README.md", "topic": [ "可验证技能锻造" ], "first_author": "Shidong Yang", "first_author_affiliation": "AMAP, Alibaba Group", "published": "2026-08-25", "code": null, "adapter": "skillforge" }, { "domain": "agent-research", "key": "skillrise", "title": "SkillRise", "paper_url": "https://arxiv.org/abs/2607.26784", "detail_path": "agent-research/2607.26784-skillrise/README.md", "topic": [ "跨任务技能进化" ], "first_author": "Zhiyuan Yao", "first_author_affiliation": "Zhejiang University", "published": "2026-07-29", "code": "https://github.com/Within-yao/SkillRise", "adapter": "skillrise" }, { "domain": "agent-research", "key": "smith", "title": "SMITH: Self-Improving Tool-Using Agents through Multi-Aspect Verification", "paper_url": "https://arxiv.org/abs/2608.24571", "detail_path": "agent-research/2608.24571-smith/README.md", "topic": [ "多维验证工具自进化" ], "first_author": "Zhi Rui Tam", "first_author_affiliation": "Appier AI Research / National Taiwan University", "published": "2026-08-25", "code": "https://github.com/appier-research/smith", "adapter": "smith" }, { "domain": "agent-research", "key": "spade", "title": "SPADE: Self-Play in Adaptive Synthetic Executable Environments", "paper_url": "https://arxiv.org/abs/2608.19197", "detail_path": "agent-research/2608.19197-spade/README.md", "topic": [ "Agentic RL", "环境生成" ], "first_author": "Bo Liu", "first_author_affiliation": "University of Washington", "published": "2026-08-19", "code": "https://github.com/spade-rl/spade", "adapter": "spade" }, { "domain": "agent-research", "key": "spo-plus-plus", "title": "SPO++: Stabilizing Asynchronous Agentic Reinforcement Learning via Measure-Theoretic Token Correction", "paper_url": "https://arxiv.org/abs/2608.24870", "detail_path": "agent-research/2608.24870-spo-plus-plus/README.md", "topic": [ "异步单流 Agent RL" ], "first_author": "Kai Ruan", "first_author_affiliation": "Renmin University of China", "published": "2026-08-25", "code": null, "adapter": "spo-plus-plus" }, { "domain": "agent-research", "key": "spt", "title": "SPT: Skills as Pre-Training Data for Agentic Language Models", "paper_url": "https://arxiv.org/abs/2608.26563", "detail_path": "agent-research/2608.26563-spt/README.md", "topic": [ "Agent 技能预训练" ], "first_author": "Yufei Sun", "first_author_affiliation": "Beijing University of Posts and Telecommunications", "published": "2026-08-27", "code": null, "adapter": "spt" }, { "domain": "agent-research", "key": "state2state", "title": "State2State: Environment-Derived Mid-Training for LLM Agents", "paper_url": "https://arxiv.org/abs/2608.04934", "detail_path": "agent-research/2608.04934-state2state/README.md", "topic": [ "环境派生中训练" ], "first_author": "Xuanyu Lei", "first_author_affiliation": "Tsinghua University AIR / Alibaba Group", "published": "2026-08-05", "code": "https://github.com/THUNLP-MT/State2State", "adapter": "state2state" }, { "domain": "agent-research", "key": "steppo", "title": "StepPO", "paper_url": "https://arxiv.org/abs/2604.18401", "detail_path": "agent-research/2604.18401-steppo/README.md", "topic": [ "Step-aligned Agent RL" ], "first_author": "Daoyu Wang", "first_author_affiliation": "University of Science and Technology of China", "published": "2026-04-20", "code": "未发现官方代码", "adapter": "steppo" }, { "domain": "agent-research", "key": "swe-agent", "title": "SWE-agent", "paper_url": "https://arxiv.org/abs/2405.15793", "detail_path": "agent-research/2405.15793-swe-agent/README.md", "topic": [ "软件工程 ACI" ], "first_author": "John Yang", "first_author_affiliation": "Princeton University", "published": "2024-05-06", "code": "https://github.com/SWE-agent/SWE-agent", "adapter": "swe-agent" }, { "domain": "agent-research", "key": "swe-prime", "title": "SWE-Prime: Fewer Trajectories, Better Performance", "paper_url": "https://arxiv.org/abs/2608.27449", "detail_path": "agent-research/2608.27449-swe-prime/README.md", "topic": [ "软件轨迹筛选" ], "first_author": "Dewu Zheng", "first_author_affiliation": "Sun Yat-sen University", "published": "2026-08-27", "code": null, "adapter": "swe-prime" }, { "domain": "agent-research", "key": "t1-terminal-rl", "title": "T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks", "paper_url": "https://arxiv.org/abs/2609.11042", "detail_path": "agent-research/2609.11042-t1-terminal-rl/README.md", "topic": [ "Agent RL", "长程终端任务" ], "first_author": "Junyao Yang", "first_author_affiliation": "National University of Singapore / Tencent", "published": "2026-09-09", "code": null, "adapter": "t1-terminal-rl", "priority": "P1" }, { "domain": "agent-research", "key": "tapo", "title": "TAPO", "paper_url": "https://arxiv.org/abs/2607.27973", "detail_path": "agent-research/2607.27973-tapo/README.md", "topic": [ "Agent RL" ], "first_author": "Cong Li", "first_author_affiliation": "Peking University", "published": "2026-07-30", "code": "未发现官方代码", "adapter": "tapo" }, { "domain": "agent-research", "key": "toolbench", "title": "ToolBench", "paper_url": "https://arxiv.org/abs/2305.16504", "detail_path": "agent-research/2305.16504-toolbench/README.md", "topic": [ "工具指令与评测" ], "first_author": "Qiantong Xu", "first_author_affiliation": "Tsinghua University", "published": "2023-05-25", "code": "未发现官方代码", "adapter": "toolbench" }, { "domain": "agent-research", "key": "toolformer", "title": "Toolformer", "paper_url": "https://arxiv.org/abs/2302.04761", "detail_path": "agent-research/2302.04761-toolformer/README.md", "topic": [ "工具学习" ], "first_author": "Timo Schick", "first_author_affiliation": "Meta AI Research", "published": "2023-02-09", "code": "未发现官方代码", "adapter": "toolformer" }, { "domain": "agent-research", "key": "toolgrad", "title": "ToolGrad: Efficient Tool-Use Dataset Generation with Textual Gradients", "paper_url": "https://arxiv.org/abs/2508.04086", "detail_path": "agent-research/2508.04086-toolgrad/README.md", "topic": [ "工具学习", "数据生成" ], "first_author": "Zhongyi Zhou", "first_author_affiliation": "Google", "published": "2025-08-06", "code": "https://github.com/zhongyi-zhou/toolgrad", "adapter": "toolgrad", "priority": "P0" }, { "domain": "agent-research", "key": "toollift", "title": "ToolLIFT: Lifting Tool-Specific Trajectories into Function-Level Graphs for Generalizable Tool Planning", "paper_url": "https://arxiv.org/abs/2608.03468", "detail_path": "agent-research/2608.03468-toollift/README.md", "topic": [ "工具规划", "图结构" ], "first_author": "Xiuhui You", "first_author_affiliation": "Authors did not disclose affiliation", "published": "2026-08-04", "code": null, "adapter": "toollift" }, { "domain": "agent-research", "key": "toolrl", "title": "ToolRL", "paper_url": "https://arxiv.org/abs/2504.13958", "detail_path": "agent-research/2504.13958-toolrl/README.md", "topic": [ "工具强化学习" ], "first_author": "Cheng Qian", "first_author_affiliation": "University of Illinois Urbana-Champaign", "published": "2025-04-16", "code": "未发现官方代码", "adapter": "toolrl" }, { "domain": "agent-research", "key": "topas", "title": "TOPAS: Workflow-Aware Prefix-State Scheduling for Multi-Agent LLM Serving", "paper_url": "https://arxiv.org/abs/2608.25523", "detail_path": "agent-research/2608.25523-topas/README.md", "topic": [ "工作流 Prefix-State 调度" ], "first_author": "Hongqiu Ni", "first_author_affiliation": "University of Science and Technology of China", "published": "2026-08-26", "code": null, "adapter": "topas" }, { "domain": "agent-research", "key": "traceml", "title": "TraceML: An Empirical Analysis of Human-Agent Planning in Machine Learning Development", "paper_url": "https://arxiv.org/abs/2608.26086", "detail_path": "agent-research/2608.26086-traceml/README.md", "topic": [ "ML 开发轨迹规划" ], "first_author": "Jiarui Yan", "first_author_affiliation": "Carnegie Mellon University", "published": "2026-08-26", "code": "https://huggingface.co/datasets/jerryyan/TraceML", "adapter": "traceml" }, { "domain": "agent-research", "key": "trca", "title": "TRCA: Transition-wise Rubric Credit Assignment for Long-horizon LLM Agents", "paper_url": "https://arxiv.org/abs/2608.16156", "detail_path": "agent-research/2608.16156-trca/README.md", "topic": [ "Agentic RL", "过程奖励" ], "first_author": "Huan Zhang", "first_author_affiliation": "Authors did not disclose affiliation", "published": "2026-08-17", "code": null, "adapter": "trca" }, { "domain": "agent-research", "key": "tree-of-thoughts", "title": "Tree of Thoughts", "paper_url": "https://arxiv.org/abs/2305.10601", "detail_path": "agent-research/2305.10601-tree-of-thoughts/README.md", "topic": [ "推理搜索" ], "first_author": "Shunyu Yao", "first_author_affiliation": "Princeton University", "published": "2023-05-17", "code": "https://github.com/princeton-nlp/tree-of-thought-llm", "adapter": "tree-of-thoughts" }, { "domain": "agent-research", "key": "turn-opd", "title": "TurnOPD", "paper_url": "https://arxiv.org/abs/2607.05804", "detail_path": "agent-research/2607.05804-turn-opd/README.md", "topic": [ "Agentic OPD / rollout budgeting" ], "first_author": "Yuhang Zhou", "first_author_affiliation": "Academic author team", "published": "2026-07-07", "code": "未发现官方代码", "adapter": "turn-opd" }, { "domain": "agent-research", "key": "u-mem", "title": "U-Mem", "paper_url": "https://arxiv.org/abs/2602.22406", "detail_path": "agent-research/2602.22406-u-mem/README.md", "topic": [ "主动记忆" ], "first_author": "Xinle Wu", "first_author_affiliation": "National University of Singapore", "published": "2026-02-25", "code": "https://anonymous.4open.science/r/code-release-456D/", "adapter": "u-mem" }, { "domain": "agent-research", "key": "unimem", "title": "UniMem", "paper_url": "https://arxiv.org/abs/2607.26017", "detail_path": "agent-research/2607.26017-unimem/README.md", "topic": [ "Continual agent memory" ], "first_author": "Siyu Xia", "first_author_affiliation": "CASIA", "published": "2026-07-28", "code": "未发现官方代码", "adapter": "unimem" }, { "domain": "agent-research", "key": "vag", "title": "When Self-Evolution Backfires: Pre-Commit Gating against Skill Contamination in LLM Agents", "paper_url": "https://arxiv.org/abs/2608.05810", "detail_path": "agent-research/2608.05810-vag/README.md", "topic": [ "技能进化安全" ], "first_author": "Linfang Shang", "first_author_affiliation": "论文未列机构", "published": "2026-08-06", "code": null, "adapter": "vag" }, { "domain": "agent-research", "key": "vermem", "title": "VerMem", "paper_url": "https://arxiv.org/abs/2608.03137", "detail_path": "agent-research/2608.03137-vermem/README.md", "topic": [ "可验证统一记忆" ], "first_author": "Xiaolong Sun", "first_author_affiliation": "Sun Yat-sen University", "published": "2026-08-04", "code": "https://github.com/Sun-SYSU-24/VerMem", "adapter": "vermem" }, { "domain": "agent-research", "key": "voyager", "title": "Voyager", "paper_url": "https://arxiv.org/abs/2305.16291", "detail_path": "agent-research/2305.16291-voyager/README.md", "topic": [ "终身学习" ], "first_author": "Guanzhi Wang", "first_author_affiliation": "NVIDIA", "published": "2023-05-25", "code": "https://github.com/MineDojo/Voyager", "adapter": "voyager" }, { "domain": "agent-research", "key": "webagent-r1", "title": "WebAgent-R1", "paper_url": "https://arxiv.org/abs/2505.16421", "detail_path": "agent-research/2505.16421-webagent-r1/README.md", "topic": [ "网页 Agent RL" ], "first_author": "Zhepei Wei", "first_author_affiliation": "University of Virginia", "published": "2025-05-22", "code": "https://github.com/weizhepei/WebAgent-R1", "adapter": "webagent-r1" }, { "domain": "agent-research", "key": "webgpt", "title": "WebGPT", "paper_url": "https://arxiv.org/abs/2112.09332", "detail_path": "agent-research/2112.09332-webgpt/README.md", "topic": [ "浏览问答" ], "first_author": "Reiichiro Nakano", "first_author_affiliation": "OpenAI", "published": "2021-12-17", "code": "https://huggingface.co/datasets/openai/webgpt_comparisons", "adapter": "webgpt" }, { "domain": "foundation-models", "key": "adadsf", "title": "Adaptive Depth Sparse Framework: Similarity-Driven Resource Allocation for Pre-Trained LLMs", "paper_url": "https://arxiv.org/abs/2607.21291", "detail_path": "reproductions/2607.21291-adadsf/README.md", "topic": [ "llm-architecture", "efficient-inference", "token-routing" ], "first_author": null, "first_author_affiliation": "Huawei ACS Lab / Southern University of Science and Technology", "published": "2026-07-23", "code": null, "adapter": { "adapter_key": "adadsf", "arxiv_id": "2607.21291", "title": "Adaptive Depth Sparse Framework: Similarity-Driven Resource Allocation for Pre-Trained LLMs", "paper_url": "https://arxiv.org/abs/2607.21291", "track": "llm", "organization": "Huawei ACS Lab / Southern University of Science and Technology", "published": "2026-07-23", "code_url": null, "topics": [ "llm-architecture", "efficient-inference", "token-routing" ], "local_code_dir": "src/auto_research/reproductions/adadsf", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同一个 96-d、4-layer dense teacher", "metrics": [ "variants.adadsf_80.perplexity", "variants.dense_teacher.perplexity", "variants.uniform_mod_80.perplexity" ], "default_seeds": [ 42 ], "budget": "steps=70", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "agentkv", "title": "AgentKV: Phase-Aware KV Eviction for Agentic LLMs", "paper_url": "https://arxiv.org/abs/2609.14872", "detail_path": "foundation-models/2609.14872-agentkv/README.md", "topic": [ "Agent 推理", "KV Cache" ], "first_author": "Taowen Tony Liu", "first_author_affiliation": "University of Cambridge", "published": "2026-09-14", "code": "https://github.com/LiuTaowen-Tony/agentkv", "adapter": "agentkv", "priority": "P1" }, { "domain": "foundation-models", "key": "alibi", "title": "Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation", "paper_url": "https://arxiv.org/abs/2108.12409", "detail_path": "reproductions/2108.12409-alibi/README.md", "topic": [ "position-encoding", "length-extrapolation", "attention" ], "first_author": null, "first_author_affiliation": "University of Washington / Meta AI", "published": "2021-08-27", "code": "https://github.com/ofirpress/attention_with_linear_biases", "adapter": { "adapter_key": "alibi", "arxiv_id": "2108.12409", "title": "Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation", "paper_url": "https://arxiv.org/abs/2108.12409", "track": "llm", "organization": "University of Washington / Meta AI", "published": "2021-08-27", "code_url": "https://github.com/ofirpress/attention_with_linear_biases", "topics": [ "position-encoding", "length-extrapolation", "attention" ], "local_code_dir": "src/auto_research/reproductions/alibi", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同预算 llamamodern，实验组为 alibi", "metrics": [ "baseline.composite_loss", "baseline.final_loss", "baseline.fitness", "baseline.initial_loss", "baseline.instruction_loss", "baseline.lm_loss", "baseline.muon_orthogonality_error", "baseline.parameters", "baseline.perplexity", "baseline.post_training_loss", "baseline.primary", "baseline.public_composite", "baseline.reward_bearing_group_rate", "method.architecture_stats.train_short_test_long", "method.composite_loss", "method.final_loss", "method.fitness", "method.initial_loss", "method.instruction_loss", "method.lm_loss", "method.muon_orthogonality_error", "method.parameters", "method.perplexity", "method.post_training_loss", "method.primary", "method.public_composite", "method.reward_bearing_group_rate", "paper_results.memory_percent", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=12", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "aspire", "title": "ASPIRE: Asynchronous Batch Self-Speculative Decoding", "paper_url": "https://arxiv.org/abs/2609.17943", "detail_path": "foundation-models/2609.17943-aspire/README.md", "topic": [ "推测解码", "推理效率" ], "first_author": "Amir Ziashahabi", "first_author_affiliation": "University of Southern California", "published": "2026-09-16", "code": null, "adapter": "aspire", "priority": "P1", "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/aspire-a100-20260919.json" }, { "domain": "foundation-models", "key": "autonomy-heads", "title": "Autonomy-of-Heads: Data-Free Sparse Attention from Frozen Query-Key Geometry", "paper_url": "https://arxiv.org/abs/2608.06849", "detail_path": "reproductions/2608.06849-autonomy-heads/README.md", "topic": [ "注意力与长上下文", "稀疏注意力" ], "first_author": null, "first_author_affiliation": "Institute of Computing Technology, Chinese Academy of Sciences", "published": "2026-08-07", "code": null, "adapter": { "adapter_key": "autonomy-heads", "arxiv_id": "2608.06849", "title": "Autonomy-of-Heads: Data-Free Sparse Attention from Frozen Query-Key Geometry", "paper_url": "https://arxiv.org/abs/2608.06849", "track": "llm", "organization": "Institute of Computing Technology, Chinese Academy of Sciences", "published": "2026-08-07", "code_url": null, "topics": [ "注意力与长上下文", "稀疏注意力" ], "local_code_dir": "src/auto_research/reproductions/autonomy_heads", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "deterministic long-context public mini-suite" ], "baseline": "recent-window attention", "metrics": [ "accuracy", "relative cost" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "awq", "title": "AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration", "paper_url": "https://arxiv.org/abs/2306.00978", "detail_path": "reproductions/2306.00978-awq/README.md", "topic": [ "inference-serving", "quantization" ], "first_author": null, "first_author_affiliation": "MIT / NVIDIA / Harvard / SJTU", "published": "2023-06-01", "code": "https://github.com/mit-han-lab/llm-awq", "adapter": { "adapter_key": "awq", "arxiv_id": "2306.00978", "title": "AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration", "paper_url": "https://arxiv.org/abs/2306.00978", "track": "llm", "organization": "MIT / NVIDIA / Harvard / SJTU", "published": "2023-06-01", "code_url": "https://github.com/mit-han-lab/llm-awq", "topics": [ "inference-serving", "quantization" ], "local_code_dir": "src/auto_research/reproductions/awq", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2 activation calibration" ], "baseline": "round-to-nearest W4，实验组为 activation-aware equivalent scaling W4，只改变论文核心机制", "metrics": [ "baseline.output_mse", "method.output_mse", "relative.output_mse_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "bakron", "title": "BaKron: Efficient Quantization with Kronecker-Factored Hessians", "paper_url": "https://arxiv.org/abs/2608.06291", "detail_path": "reproductions/2608.06291-bakron/README.md", "topic": [ "quantization", "hessian", "inference-efficiency" ], "first_author": null, "first_author_affiliation": "University of California, San Diego", "published": "2026-08-06", "code": null, "adapter": { "adapter_key": "bakron", "arxiv_id": "2608.06291", "title": "BaKron: Efficient Quantization with Kronecker-Factored Hessians", "paper_url": "https://arxiv.org/abs/2608.06291", "track": "llm", "organization": "University of California, San Diego", "published": "2026-08-06", "code_url": null, "topics": [ "quantization", "hessian", "inference-efficiency" ], "local_code_dir": "src/auto_research/reproductions/bakron", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "deterministic public-style mini-suite" ], "baseline": "去掉论文特有机制、其余数据切分与预算相同的 matched control", "metrics": [ "bakron.reported_objective" ], "default_seeds": [ 42 ], "budget": "steps=160", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "beaconkv", "title": "BeaconKV: Key-Value Cache Compression Guided by Beacon Queries for Efficient Large Reasoning Model Inference", "paper_url": "https://arxiv.org/abs/2609.04971", "detail_path": "reproductions/2609.04971-beaconkv/README.md", "topic": [ "kv-cache", "long-context", "reasoning", "inference-serving" ], "first_author": null, "first_author_affiliation": "Hanyang University", "published": "2026-09-04", "code": null, "adapter": { "adapter_key": "beaconkv", "arxiv_id": "2609.04971", "title": "BeaconKV: Key-Value Cache Compression Guided by Beacon Queries for Efficient Large Reasoning Model Inference", "paper_url": "https://arxiv.org/abs/2609.04971", "track": "llm", "organization": "Hanyang University", "published": "2026-09-04", "code_url": null, "topics": [ "kv-cache", "long-context", "reasoning", "inference-serving" ], "local_code_dir": "src/auto_research/reproductions/beaconkv", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "public-domain long-context probe" ], "baseline": "equal-budget recent-token KV retention", "metrics": [ "attention-output cosine", "retained tokens", "peak GPU memory" ], "default_seeds": [ 42, 43, 44 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cuda" ], "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/beaconkv-a100-20260907.json", "online_evidence": [], "selection_exception": null, "evolve_operators": [ "attention:beacon-query-kv" ] } }, { "domain": "foundation-models", "key": "blip2", "title": "BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models", "paper_url": "https://arxiv.org/abs/2301.12597", "detail_path": "reproductions/2301.12597-blip2/README.md", "topic": [ "multimodal", "q-former", "frozen-encoders" ], "first_author": null, "first_author_affiliation": "Salesforce Research", "published": "2023-01-30", "code": "https://github.com/salesforce/LAVIS", "adapter": { "adapter_key": "blip2", "arxiv_id": "2301.12597", "title": "BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models", "paper_url": "https://arxiv.org/abs/2301.12597", "track": "llm", "organization": "Salesforce Research", "published": "2023-01-30", "code_url": "https://github.com/salesforce/LAVIS", "topics": [ "multimodal", "q-former", "frozen-encoders" ], "local_code_dir": "src/auto_research/reproductions/blip2", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "Fashion-MNIST object QA" ], "baseline": "linear mean-pooled connector", "metrics": [ "test_accuracy", "visual_dependency_delta", "visual_tokens" ], "default_seeds": [ 42 ], "budget": "steps=120", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "blt", "title": "Byte Latent Transformer: Patches Scale Better Than Tokens", "paper_url": "https://arxiv.org/abs/2412.09871", "detail_path": "reproductions/2412.09871-blt/README.md", "topic": [ "tokenizer-free", "byte-model", "dynamic-patching", "architecture" ], "first_author": null, "first_author_affiliation": "Meta FAIR", "published": "2024-12-13", "code": "https://github.com/facebookresearch/blt", "adapter": { "adapter_key": "blt", "arxiv_id": "2412.09871", "title": "Byte Latent Transformer: Patches Scale Better Than Tokens", "paper_url": "https://arxiv.org/abs/2412.09871", "track": "llm", "organization": "Meta FAIR", "published": "2024-12-13", "code_url": "https://github.com/facebookresearch/blt", "topics": [ "tokenizer-free", "byte-model", "dynamic-patching", "architecture" ], "local_code_dir": "src/auto_research/reproductions/blt", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同预算 llamamodern，实验组为 blt", "metrics": [ "baseline.composite_loss", "baseline.final_loss", "baseline.fitness", "baseline.initial_loss", "baseline.instruction_loss", "baseline.lm_loss", "baseline.muon_orthogonality_error", "baseline.parameters", "baseline.perplexity", "baseline.post_training_loss", "baseline.primary", "baseline.public_composite", "baseline.reward_bearing_group_rate", "method.architecture_stats.expanded_back_to_byte_targets", "method.architecture_stats.observed_patch_boundary_rate", "method.composite_loss", "method.final_loss", "method.fitness", "method.initial_loss", "method.instruction_loss", "method.lm_loss", "method.muon_orthogonality_error", "method.parameters", "method.perplexity", "method.post_training_loss", "method.primary", "method.public_composite", "method.reward_bearing_group_rate", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=12", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "c2kv", "title": "C²KV: Compressed and Composable KV Cache Reuse for Efficient LLM Inference", "paper_url": "https://arxiv.org/abs/2607.17715", "detail_path": "reproductions/2607.17715-c2kv/README.md", "topic": [ "推理与系统效率", "KV cache" ], "first_author": null, "first_author_affiliation": "Shanghai Jiao Tong University", "published": "2026-07-20", "code": "https://github.com/s7a9/C2KV", "adapter": { "adapter_key": "c2kv", "arxiv_id": "2607.17715", "title": "C²KV: Compressed and Composable KV Cache Reuse for Efficient LLM Inference", "paper_url": "https://arxiv.org/abs/2607.17715", "track": "llm", "organization": "Shanghai Jiao Tong University", "published": "2026-07-20", "code_url": "https://github.com/s7a9/C2KV", "topics": [ "推理与系统效率", "KV cache" ], "local_code_dir": "src/auto_research/reproductions/c2kv", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "deterministic long-context public mini-suite" ], "baseline": "recent-window attention", "metrics": [ "accuracy", "relative cost" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "clip", "title": "Learning Transferable Visual Models From Natural Language Supervision", "paper_url": "https://arxiv.org/abs/2103.00020", "detail_path": "reproductions/2103.00020-clip/README.md", "topic": [ "multimodal", "contrastive-pretraining" ], "first_author": null, "first_author_affiliation": "OpenAI", "published": "2021-02-26", "code": "https://github.com/openai/CLIP", "adapter": { "adapter_key": "clip", "arxiv_id": "2103.00020", "title": "Learning Transferable Visual Models From Natural Language Supervision", "paper_url": "https://arxiv.org/abs/2103.00020", "track": "llm", "organization": "OpenAI", "published": "2021-02-26", "code_url": "https://github.com/openai/CLIP", "topics": [ "multimodal", "contrastive-pretraining" ], "local_code_dir": "src/auto_research/reproductions/clip", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "Fashion-MNIST image-text pairs" ], "baseline": "uniform label retrieval", "metrics": [ "test_accuracy", "visual_dependency_delta" ], "default_seeds": [ 42 ], "budget": "steps=150", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "conv-llm", "title": "Convolution for Large Language Models", "paper_url": "https://arxiv.org/abs/2607.18413", "detail_path": "reproductions/2607.18413-conv-llm/README.md", "topic": [ "llm-architecture", "attention", "depthwise-convolution" ], "first_author": null, "first_author_affiliation": "Huawei / Peking University / Tsinghua University", "published": "2026-07-20", "code": null, "adapter": { "adapter_key": "conv-llm", "arxiv_id": "2607.18413", "title": "Convolution for Large Language Models", "paper_url": "https://arxiv.org/abs/2607.18413", "track": "llm", "organization": "Huawei / Peking University / Tsinghua University", "published": "2026-07-20", "code_url": null, "topics": [ "llm-architecture", "attention", "depthwise-convolution" ], "local_code_dir": "src/auto_research/reproductions/conv_llm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "相同 96-d、3-layer、60-step Transformer", "metrics": [ "baseline.parameters", "baseline.perplexity", "baseline.test_loss", "method.parameters", "method.perplexity", "method.test_loss", "perplexity_reduction_percent" ], "default_seeds": [ 42 ], "budget": "steps=60", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "criticl", "title": "CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes", "paper_url": "https://arxiv.org/abs/2608.27455", "detail_path": "foundation-models/2608.27455-criticl/README.md", "topic": [ "弱模型失败模式 ICL" ], "first_author": "Yufan Wu", "first_author_affiliation": "The Ohio State University", "published": "2026-08-27", "code": null, "adapter": "criticl" }, { "domain": "foundation-models", "key": "dart", "title": "DART: Decoded Attention over Recurrent States for Efficient Long-Context Sequence Modeling", "paper_url": "https://arxiv.org/abs/2608.02032", "detail_path": "reproductions/2608.02032-dart/README.md", "topic": [ "网络架构", "注意力与长上下文" ], "first_author": null, "first_author_affiliation": "Zhejiang University", "published": "2026-08-03", "code": null, "adapter": { "adapter_key": "dart", "arxiv_id": "2608.02032", "title": "DART: Decoded Attention over Recurrent States for Efficient Long-Context Sequence Modeling", "paper_url": "https://arxiv.org/abs/2608.02032", "track": "llm", "organization": "Zhejiang University", "published": "2026-08-03", "code_url": null, "topics": [ "网络架构", "注意力与长上下文" ], "local_code_dir": "src/auto_research/reproductions/dart", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "deterministic long-context public mini-suite" ], "baseline": "recent-window attention", "metrics": [ "accuracy", "relative cost" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "data-mixing-laws", "title": "Data Mixing Laws: Optimizing Data Mixtures by Predicting Language Modeling Performance", "paper_url": "https://arxiv.org/abs/2403.16952", "detail_path": "reproductions/2403.16952-data-mixing-laws/README.md", "topic": [ "pretraining-data", "scaling-law", "data-mixture" ], "first_author": null, "first_author_affiliation": "University of Cambridge / Shanghai AI Laboratory", "published": "2024-03-25", "code": "https://github.com/yegcjs/mixinglaws", "adapter": { "adapter_key": "data-mixing-laws", "arxiv_id": "2403.16952", "title": "Data Mixing Laws: Optimizing Data Mixtures by Predicting Language Modeling Performance", "paper_url": "https://arxiv.org/abs/2403.16952", "track": "llm", "organization": "University of Cambridge / Shanghai AI Laboratory", "published": "2024-03-25", "code_url": "https://github.com/yegcjs/mixinglaws", "topics": [ "pretraining-data", "scaling-law", "data-mixture" ], "local_code_dir": "src/auto_research/reproductions/data_mixing_laws", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2 + public narrative domain" ], "baseline": "均匀数据混合，实验组为 data-mixing-laws", "metrics": [ "baseline.validation_loss", "method.validation_loss", "relative.validation_loss_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "data-orchestra", "title": "DataOrchestra: Learning to Orchestrate Per-Example Curation of Pretraining Data", "paper_url": "https://arxiv.org/abs/2607.24717", "detail_path": "reproductions/2607.24717-data-orchestra/README.md", "topic": [ "llm-pretraining", "data-curation", "data-cleaning", "learned-orchestration" ], "first_author": null, "first_author_affiliation": "Fudan University / Shanghai Jiao Tong University / SII-GAIR", "published": "2026-07-27", "code": "https://github.com/GAIR-NLP/DataOrchestra", "adapter": { "adapter_key": "data-orchestra", "arxiv_id": "2607.24717", "title": "DataOrchestra: Learning to Orchestrate Per-Example Curation of Pretraining Data", "paper_url": "https://arxiv.org/abs/2607.24717", "track": "llm", "organization": "Fudan University / Shanghai Jiao Tong University / SII-GAIR", "published": "2026-07-27", "code_url": "https://github.com/GAIR-NLP/DataOrchestra", "topics": [ "llm-pretraining", "data-curation", "data-cleaning", "learned-orchestration" ], "local_code_dir": "src/auto_research/reproductions/data_orchestra", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "在 WikiText-2 raw 文本上训练同一个 208,872 参数 Llama-style LM", "metrics": [ "orchestrator_training.final_loss", "orchestrator_training.held_out_action_accuracy", "orchestrator_training.held_out_operation_accuracy", "orchestrator_training.initial_loss", "relative.perplexity_reduction_vs_raw_percent", "setup.same_tokens_steps_optimizer_architecture", "variants.DataOrchestra.final_loss", "variants.DataOrchestra.initial_loss", "variants.DataOrchestra.loss", "variants.DataOrchestra.perplexity", "variants.Raw.final_loss", "variants.Raw.initial_loss", "variants.Raw.loss", "variants.Raw.perplexity", "variants.Static cleaner.final_loss", "variants.Static cleaner.initial_loss", "variants.Static cleaner.loss", "variants.Static cleaner.perplexity" ], "default_seeds": [ 42 ], "budget": "steps=55", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "dblast", "title": "DBLast: Dependent Block Drafting for Stochastic Speculative Decoding", "paper_url": "https://arxiv.org/abs/2608.05448", "detail_path": "reproductions/2608.05448-dblast/README.md", "topic": [ "speculative-decoding", "inference-efficiency", "stochastic-decoding" ], "first_author": null, "first_author_affiliation": "Huawei Technologies Canada", "published": "2026-08-05", "code": null, "adapter": { "adapter_key": "dblast", "arxiv_id": "2608.05448", "title": "DBLast: Dependent Block Drafting for Stochastic Speculative Decoding", "paper_url": "https://arxiv.org/abs/2608.05448", "track": "llm", "organization": "Huawei Technologies Canada", "published": "2026-08-05", "code_url": null, "topics": [ "speculative-decoding", "inference-efficiency", "stochastic-decoding" ], "local_code_dir": "src/auto_research/reproductions/dblast", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "deterministic public-style mini-suite" ], "baseline": "去掉论文特有机制、其余数据切分与预算相同的 matched control", "metrics": [ "dblast.reported_objective" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "distillcache", "title": "DistillCache: KL-Guided Adaptive KV-Cache Eviction for Memory-Efficient LLM Inference", "paper_url": "https://arxiv.org/abs/2608.08878", "detail_path": "reproductions/2608.08878-distillcache/README.md", "topic": [ "推理与系统效率", "KV cache" ], "first_author": null, "first_author_affiliation": "Oklahoma State University", "published": "2026-08-09", "code": null, "adapter": { "adapter_key": "distillcache", "arxiv_id": "2608.08878", "title": "DistillCache: KL-Guided Adaptive KV-Cache Eviction for Memory-Efficient LLM Inference", "paper_url": "https://arxiv.org/abs/2608.08878", "track": "llm", "organization": "Oklahoma State University", "published": "2026-08-09", "code_url": null, "topics": [ "推理与系统效率", "KV cache" ], "local_code_dir": "src/auto_research/reproductions/distillcache", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "deterministic long-context public mini-suite" ], "baseline": "recent-window attention", "metrics": [ "accuracy", "relative cost" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "doremi", "title": "DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining", "paper_url": "https://arxiv.org/abs/2305.10429", "detail_path": "reproductions/2305.10429-doremi/README.md", "topic": [ "pretraining-data", "data-mixture", "group-dro" ], "first_author": null, "first_author_affiliation": "Stanford University / Google Research", "published": "2023-05-17", "code": "https://github.com/sangmichaelxie/doremi", "adapter": { "adapter_key": "doremi", "arxiv_id": "2305.10429", "title": "DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining", "paper_url": "https://arxiv.org/abs/2305.10429", "track": "llm", "organization": "Stanford University / Google Research", "published": "2023-05-17", "code_url": "https://github.com/sangmichaelxie/doremi", "topics": [ "pretraining-data", "data-mixture", "group-dro" ], "local_code_dir": "src/auto_research/reproductions/doremi", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2 + public narrative domain" ], "baseline": "均匀数据混合，实验组为 doremi", "metrics": [ "baseline.validation_loss", "method.validation_loss", "paper_results.few_shot_accuracy_points", "relative.validation_loss_percent", "stages.worst_excess_loss_updates" ], "default_seeds": [ 42 ], "budget": "steps=80", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "dqwen35", "title": "dQwen3.5: Adapting Hybrid Language Models into Bidirectional Diffusion Language Models", "paper_url": "https://arxiv.org/abs/2609.20751", "detail_path": "foundation-models/2609.20751-dqwen35/README.md", "topic": [ "扩散语言模型", "网络架构" ], "first_author": "Anton Xue", "first_author_affiliation": "University of Texas at Austin", "published": "2026-09-17", "code": null, "checkpoint": "https://huggingface.co/UT-IFML/dQwen3.5-9B-Base", "adapter": "dqwen35", "priority": "P1", "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/dqwen35-a100-20260919.json" }, { "domain": "foundation-models", "key": "echo", "title": "ECHO: Early-layer Collaborative Hierarchical Orchestration with Bonus Logits in Speculative Decoding", "paper_url": "https://arxiv.org/abs/2609.17241", "detail_path": "foundation-models/2609.17241-echo/README.md", "topic": [ "推测解码", "推理效率" ], "first_author": "Ziyang Ma", "first_author_affiliation": "Wuhan University", "published": "2026-09-15", "code": "https://github.com/whucs21Mzy/ECHO", "adapter": "echo", "priority": "P0", "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/echo-a100-20260916.json" }, { "domain": "foundation-models", "key": "engram", "title": "Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models", "paper_url": "https://arxiv.org/abs/2601.07372", "detail_path": "reproductions/2601.07372-engram/README.md", "topic": [ "llm-architecture", "conditional-memory", "sparsity" ], "first_author": null, "first_author_affiliation": "DeepSeek", "published": "2026-01-12", "code": "https://github.com/deepseek-ai/Engram", "adapter": { "adapter_key": "engram", "arxiv_id": "2601.07372", "title": "Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models", "paper_url": "https://arxiv.org/abs/2601.07372", "track": "llm", "organization": "DeepSeek", "published": "2026-01-12", "code_url": "https://github.com/deepseek-ai/Engram", "topics": [ "llm-architecture", "conditional-memory", "sparsity" ], "local_code_dir": "src/auto_research/reproductions/engram", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同预算 llamamodern", "metrics": [ "baseline.lm_loss", "baseline.parameters", "baseline.perplexity", "method.lm_loss", "method.parameters", "method.perplexity", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=30", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "frames-on-demand", "title": "Caption-once, Frames-on-Demand: Visual-Need Routing for Budget-Aware Agentic Long Video Understanding", "paper_url": "https://arxiv.org/abs/2609.11899", "detail_path": "foundation-models/2609.11899-frames-on-demand/README.md", "topic": [ "长视频理解", "视觉预算路由" ], "first_author": "Weitong Cai", "first_author_affiliation": "Queen Mary University of London", "published": "2026-09-10", "code": null, "adapter": "frames-on-demand", "priority": "P1" }, { "domain": "foundation-models", "key": "gas", "title": "Generation as Auxiliary Supervision: Enhancing Visual Understanding at Zero Inference Overhead via Decoupled Embedding Prediction", "paper_url": "https://arxiv.org/abs/2608.12209", "detail_path": "reproductions/2608.12209-gas/README.md", "topic": [ "multimodal", "generation-guided-training", "next-embedding-prediction" ], "first_author": null, "first_author_affiliation": "ByteDance", "published": "2026-08-12", "code": null, "adapter": { "adapter_key": "gas", "arxiv_id": "2608.12209", "title": "Generation as Auxiliary Supervision: Enhancing Visual Understanding at Zero Inference Overhead via Decoupled Embedding Prediction", "paper_url": "https://arxiv.org/abs/2608.12209", "track": "llm", "organization": "ByteDance", "published": "2026-08-12", "code_url": null, "topics": [ "multimodal", "generation-guided-training", "next-embedding-prediction" ], "local_code_dir": "src/auto_research/reproductions/gas", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "Fashion-MNIST object QA" ], "baseline": "understanding-only matched trunk", "metrics": [ "test_accuracy", "visual_dependency_delta", "deployed_parameter_overhead_percent" ], "default_seeds": [ 42 ], "budget": "steps=120", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "gated-attention", "title": "Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free", "paper_url": "https://arxiv.org/abs/2505.06708", "detail_path": "reproductions/2505.06708-gated-attention/README.md", "topic": [ "llm-architecture", "attention-gating", "long-context" ], "first_author": null, "first_author_affiliation": "Qwen / Alibaba", "published": "2025-05-10", "code": "https://github.com/qiuzh20/gated_attention", "adapter": { "adapter_key": "gated-attention", "arxiv_id": "2505.06708", "title": "Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free", "paper_url": "https://arxiv.org/abs/2505.06708", "track": "llm", "organization": "Qwen / Alibaba", "published": "2025-05-10", "code_url": "https://github.com/qiuzh20/gated_attention", "topics": [ "llm-architecture", "attention-gating", "long-context" ], "local_code_dir": "src/auto_research/reproductions/gated_attention", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同预算 llamamodern", "metrics": [ "baseline.lm_loss", "baseline.parameters", "baseline.perplexity", "method.lm_loss", "method.output_gate_below_0_5", "method.output_gate_mean", "method.parameters", "method.perplexity", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=30", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "gaugequant", "title": "GaugeQuant: Online Learning of Quantization-Optimal Bases from LLM Symmetries", "paper_url": "https://arxiv.org/abs/2607.20757", "detail_path": "reproductions/2607.20757-gaugequant/README.md", "topic": [ "llm-training", "quantization", "w4a4" ], "first_author": null, "first_author_affiliation": "University of Cambridge", "published": "2026-07-22", "code": "https://github.com/MPedraBento/gauge-quant", "adapter": { "adapter_key": "gaugequant", "arxiv_id": "2607.20757", "title": "GaugeQuant: Online Learning of Quantization-Optimal Bases from LLM Symmetries", "paper_url": "https://arxiv.org/abs/2607.20757", "track": "llm", "organization": "University of Cambridge", "published": "2026-07-22", "code_url": "https://github.com/MPedraBento/gauge-quant", "topics": [ "llm-training", "quantization", "w4a4" ], "local_code_dir": "src/auto_research/reproductions/gaugequant", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "浮点 llamamodern", "metrics": [ "baseline.lm_loss", "baseline.parameters", "baseline.perplexity", "method.lm_loss", "method.parameters", "method.perplexity", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=30", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "gliclass", "title": "GLiClass: Generalist Lightweight Model for Sequence Classification Tasks", "paper_url": "https://arxiv.org/abs/2508.07662", "detail_path": "foundation-models/2508.07662-gliclass/README.md", "topic": [ "动态候选分类", "System One 相邻方法" ], "first_author": "Ihor Stepanov", "first_author_affiliation": "Knowledgator", "published": "2025-08-11", "code": "https://github.com/Knowledgator/GLiClass", "adapter": "system-one:bilinear-ce", "priority": "P1" }, { "domain": "foundation-models", "key": "gqa", "title": "GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints", "paper_url": "https://arxiv.org/abs/2305.13245", "detail_path": "reproductions/2305.13245-gqa/README.md", "topic": [ "attention", "kv-cache", "serving", "architecture" ], "first_author": null, "first_author_affiliation": "Google Research", "published": "2023-05-22", "code": null, "adapter": { "adapter_key": "gqa", "arxiv_id": "2305.13245", "title": "GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints", "paper_url": "https://arxiv.org/abs/2305.13245", "track": "llm", "organization": "Google Research", "published": "2023-05-22", "code_url": null, "topics": [ "attention", "kv-cache", "serving", "architecture" ], "local_code_dir": "src/auto_research/reproductions/gqa", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同预算 llamamodern，实验组为 gqa", "metrics": [ "baseline.composite_loss", "baseline.final_loss", "baseline.fitness", "baseline.initial_loss", "baseline.instruction_loss", "baseline.lm_loss", "baseline.muon_orthogonality_error", "baseline.parameters", "baseline.perplexity", "baseline.post_training_loss", "baseline.primary", "baseline.public_composite", "baseline.reward_bearing_group_rate", "method.architecture_stats.kv_cache_reduction_x", "method.architecture_stats.kv_heads", "method.architecture_stats.query_heads", "method.composite_loss", "method.final_loss", "method.fitness", "method.initial_loss", "method.instruction_loss", "method.lm_loss", "method.muon_orthogonality_error", "method.parameters", "method.perplexity", "method.post_training_loss", "method.primary", "method.public_composite", "method.reward_bearing_group_rate", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=12", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "gzip-sparse-attention", "title": "Parameter-free Adaptive Sparse Attention via Compression-Based Content Selection", "paper_url": "https://arxiv.org/abs/2607.21752", "detail_path": "reproductions/2607.21752-gzip-sparse-attention/README.md", "topic": [ "llm", "sparse-attention", "long-context", "compression" ], "first_author": null, "first_author_affiliation": "Pennsylvania State University", "published": "2026-07-23", "code": null, "adapter": { "adapter_key": "gzip-sparse-attention", "arxiv_id": "2607.21752", "title": "Parameter-free Adaptive Sparse Attention via Compression-Based Content Selection", "paper_url": "https://arxiv.org/abs/2607.21752", "track": "llm", "organization": "Pennsylvania State University", "published": "2026-07-23", "code_url": null, "topics": [ "llm", "sparse-attention", "long-context", "compression" ], "local_code_dir": "src/auto_research/reproductions/gzip_sparse_attention", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2 raw UTF-8 bytes" ], "baseline": "相同 132,032 参数、相同初始化、WikiText-2 原始 bytes、256 context 和 120 steps 的 BigBird fixed mask", "metrics": [ "gzip-sparse-attention.reported_objective" ], "default_seeds": [ 42 ], "budget": "steps=120", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "hilp", "title": "Hierarchical Latent Prediction for Language Models", "paper_url": "https://arxiv.org/abs/2608.05806", "detail_path": "reproductions/2608.05806-hilp/README.md", "topic": [ "pretraining", "latent-prediction", "multi-token-prediction", "speculative-decoding" ], "first_author": null, "first_author_affiliation": "University of Texas at Austin", "published": "2026-08-06", "code": null, "adapter": { "adapter_key": "hilp", "arxiv_id": "2608.05806", "title": "Hierarchical Latent Prediction for Language Models", "paper_url": "https://arxiv.org/abs/2608.05806", "track": "llm", "organization": "University of Texas at Austin", "published": "2026-08-06", "code_url": null, "topics": [ "pretraining", "latent-prediction", "multi-token-prediction", "speculative-decoding" ], "local_code_dir": "src/auto_research/reproductions/hilp", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "deterministic public-style mini-suite" ], "baseline": "去掉论文特有机制、其余数据切分与预算相同的 matched control", "metrics": [ "hilp.reported_objective" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "hymba", "title": "Hymba: A Hybrid-head Architecture for Small Language Models", "paper_url": "https://arxiv.org/abs/2411.13676", "detail_path": "reproductions/2411.13676-hymba/README.md", "topic": [ "hybrid-attention-ssm", "small-language-model", "architecture" ], "first_author": null, "first_author_affiliation": "NVIDIA", "published": "2024-11-20", "code": "https://github.com/NVlabs/hymba", "adapter": { "adapter_key": "hymba", "arxiv_id": "2411.13676", "title": "Hymba: A Hybrid-head Architecture for Small Language Models", "paper_url": "https://arxiv.org/abs/2411.13676", "track": "llm", "organization": "NVIDIA", "published": "2024-11-20", "code_url": "https://github.com/NVlabs/hymba", "topics": [ "hybrid-attention-ssm", "small-language-model", "architecture" ], "local_code_dir": "src/auto_research/reproductions/hymba", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同预算 llamamodern，实验组为 hymba", "metrics": [ "baseline.composite_loss", "baseline.final_loss", "baseline.fitness", "baseline.initial_loss", "baseline.instruction_loss", "baseline.lm_loss", "baseline.muon_orthogonality_error", "baseline.parameters", "baseline.perplexity", "baseline.post_training_loss", "baseline.primary", "baseline.public_composite", "baseline.reward_bearing_group_rate", "method.architecture_stats.hybrid_gate_mean", "method.architecture_stats.parallel_attention_ssm", "method.composite_loss", "method.final_loss", "method.fitness", "method.initial_loss", "method.instruction_loss", "method.lm_loss", "method.muon_orthogonality_error", "method.parameters", "method.perplexity", "method.post_training_loss", "method.primary", "method.public_composite", "method.reward_bearing_group_rate", "paper_results.accuracy_vs_llama32_3b_percent", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=12", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "kvmem", "title": "KVMem: Virtualizing Million-Token Agent Workspaces on a Consumer GPU", "paper_url": "https://arxiv.org/abs/2609.04852", "detail_path": "reproductions/2609.04852-kvmem/README.md", "topic": [ "agent-memory", "kv-cache", "long-context", "inference-serving" ], "first_author": null, "first_author_affiliation": "Shanghai University of Finance and Economics", "published": "2026-09-04", "code": "https://github.com/kvmem/kvmem-qw3", "adapter": { "adapter_key": "kvmem", "arxiv_id": "2609.04852", "title": "KVMem: Virtualizing Million-Token Agent Workspaces on a Consumer GPU", "paper_url": "https://arxiv.org/abs/2609.04852", "track": "llm", "organization": "Shanghai University of Finance and Economics", "published": "2026-09-04", "code_url": "https://github.com/kvmem/kvmem-qw3", "topics": [ "agent-memory", "kv-cache", "long-context", "inference-serving" ], "local_code_dir": "src/auto_research/reproductions/kvmem", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "public-domain long-context probe" ], "baseline": "equal-budget recent-context compaction", "metrics": [ "attention-output cosine", "materialized blocks", "peak GPU memory" ], "default_seeds": [ 42, 43, 44 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cuda" ], "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/kvmem-a100-20260907.json", "online_evidence": [], "selection_exception": null, "evolve_operators": [ "memory:query-conditioned-paged-kv" ] } }, { "domain": "foundation-models", "key": "llava", "title": "Visual Instruction Tuning", "paper_url": "https://arxiv.org/abs/2304.08485", "detail_path": "reproductions/2304.08485-llava/README.md", "topic": [ "multimodal", "instruction-tuning" ], "first_author": null, "first_author_affiliation": "University of Wisconsin-Madison / Microsoft Research / Columbia University", "published": "2023-04-17", "code": "https://github.com/haotian-liu/LLaVA", "adapter": { "adapter_key": "llava", "arxiv_id": "2304.08485", "title": "Visual Instruction Tuning", "paper_url": "https://arxiv.org/abs/2304.08485", "track": "llm", "organization": "University of Wisconsin-Madison / Microsoft Research / Columbia University", "published": "2023-04-17", "code_url": "https://github.com/haotian-liu/LLaVA", "topics": [ "multimodal", "instruction-tuning" ], "local_code_dir": "src/auto_research/reproductions/llava", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "Fashion-MNIST object QA" ], "baseline": "linear mean-pooled connector", "metrics": [ "test_accuracy", "visual_dependency_delta" ], "default_seeds": [ 42 ], "budget": "steps=120", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "lngram-v2", "title": "Lngram v2: Latent N-Gram Memory with Interpretable Discrete Representations", "paper_url": "https://arxiv.org/abs/2609.03426", "detail_path": "reproductions/2609.03426-lngram-v2/README.md", "topic": [ "multimodal-foundation-model", "conditional-memory", "discrete-routing", "model-architecture" ], "first_author": null, "first_author_affiliation": "Beijing University of Posts and Telecommunications / Kuaishou Technology", "published": "2026-09-03", "code": null, "adapter": { "adapter_key": "lngram-v2", "arxiv_id": "2609.03426", "title": "Lngram v2: Latent N-Gram Memory with Interpretable Discrete Representations", "paper_url": "https://arxiv.org/abs/2609.03426", "track": "llm", "organization": "Beijing University of Posts and Telecommunications / Kuaishou Technology", "published": "2026-09-03", "code_url": null, "topics": [ "multimodal-foundation-model", "conditional-memory", "discrete-routing", "model-architecture" ], "local_code_dir": "src/auto_research/reproductions/lngram_v2", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "SmolVLM2-256M-Video-Instruct checkpoint + deterministic RGB geometry probe" ], "baseline": "same checkpoint without latent memory branch", "metrics": [ "finite output", "route diversity", "sink selectivity", "activation memory" ], "default_seeds": [ 42, 43, 44 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cuda" ], "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/lngram-v2-a100-20260906.json", "online_evidence": [], "selection_exception": null, "evolve_operators": [ "memory:latent-ngram-gqa" ] } }, { "domain": "foundation-models", "key": "looped-latent-attention", "title": "Looped Latent Attention: Cross-Loop KV Compression for Looped Transformers", "paper_url": "https://arxiv.org/abs/2607.15456", "detail_path": "reproductions/2607.15456-looped-latent-attention/README.md", "topic": [ "llm-architecture", "kv-cache", "looped-transformer" ], "first_author": null, "first_author_affiliation": "University of Maryland / Meta AI", "published": "2026-07-16", "code": null, "adapter": { "adapter_key": "looped-latent-attention", "arxiv_id": "2607.15456", "title": "Looped Latent Attention: Cross-Loop KV Compression for Looped Transformers", "paper_url": "https://arxiv.org/abs/2607.15456", "track": "llm", "organization": "University of Maryland / Meta AI", "published": "2026-07-16", "code_url": null, "topics": [ "llm-architecture", "kv-cache", "looped-transformer" ], "local_code_dir": "src/auto_research/reproductions/looped_latent_attention", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "两层 llamamodern", "metrics": [ "baseline.lm_loss", "baseline.parameters", "baseline.perplexity", "method.lm_loss", "method.parameters", "method.perplexity", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=30", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "loopspec", "title": "LoopSpec: Pipelined Self-Speculative Decoding for Looped Transformers", "paper_url": "https://arxiv.org/abs/2609.17184", "detail_path": "foundation-models/2609.17184-loopspec/README.md", "topic": [ "推测解码", "循环 Transformer" ], "first_author": "SangLyul Cho", "first_author_affiliation": "Seoul National University / KAIST", "published": "2026-09-15", "code": null, "adapter": "loopspec", "priority": "P1" }, { "domain": "foundation-models", "key": "macro", "title": "MACRO: Markov Chain Routing of Transformer Layers", "paper_url": "https://arxiv.org/abs/2608.05872", "detail_path": "reproductions/2608.05872-macro/README.md", "topic": [ "network-architecture", "dynamic-routing", "inference-efficiency" ], "first_author": null, "first_author_affiliation": "Heinrich Heine University Düsseldorf", "published": "2026-08-06", "code": "https://github.com/Batorskq/MACRO", "adapter": { "adapter_key": "macro", "arxiv_id": "2608.05872", "title": "MACRO: Markov Chain Routing of Transformer Layers", "paper_url": "https://arxiv.org/abs/2608.05872", "track": "llm", "organization": "Heinrich Heine University Düsseldorf", "published": "2026-08-06", "code_url": "https://github.com/Batorskq/MACRO", "topics": [ "network-architecture", "dynamic-routing", "inference-efficiency" ], "local_code_dir": "src/auto_research/reproductions/macro", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "deterministic public-style mini-suite" ], "baseline": "去掉论文特有机制、其余数据切分与预算相同的 matched control", "metrics": [ "macro_accuracy", "paper_results.average_accuracy_percent", "sequential_accuracy" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "mamba", "title": "Mamba: Linear-Time Sequence Modeling with Selective State Spaces", "paper_url": "https://arxiv.org/abs/2312.00752", "detail_path": "reproductions/2312.00752-mamba/README.md", "topic": [ "llm-architecture", "state-space-model", "classic" ], "first_author": null, "first_author_affiliation": "Carnegie Mellon University / Princeton University", "published": "2023-12-01", "code": "https://github.com/state-spaces/mamba", "adapter": { "adapter_key": "mamba", "arxiv_id": "2312.00752", "title": "Mamba: Linear-Time Sequence Modeling with Selective State Spaces", "paper_url": "https://arxiv.org/abs/2312.00752", "track": "llm", "organization": "Carnegie Mellon University / Princeton University", "published": "2023-12-01", "code_url": "https://github.com/state-spaces/mamba", "topics": [ "llm-architecture", "state-space-model", "classic" ], "local_code_dir": "src/auto_research/reproductions/mamba", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "tiny Transformer", "metrics": [ "baseline.perplexity", "method.perplexity" ], "default_seeds": [ 42 ], "budget": "steps=30", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "medusa", "title": "Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads", "paper_url": "https://arxiv.org/abs/2401.10774", "detail_path": "reproductions/2401.10774-medusa/README.md", "topic": [ "inference-serving", "decoding" ], "first_author": null, "first_author_affiliation": "Together AI / Princeton University / University of Illinois Urbana-Champaign", "published": "2024-01-19", "code": "https://github.com/FasterDecoding/Medusa", "adapter": { "adapter_key": "medusa", "arxiv_id": "2401.10774", "title": "Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads", "paper_url": "https://arxiv.org/abs/2401.10774", "track": "llm", "organization": "Together AI / Princeton University / University of Illinois Urbana-Champaign", "published": "2024-01-19", "code_url": "https://github.com/FasterDecoding/Medusa", "topics": [ "inference-serving", "decoding" ], "local_code_dir": "src/auto_research/reproductions/medusa", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "single-head greedy，实验组为 three future heads + tree verification，只改变论文核心机制", "metrics": [ "baseline.backbone_calls", "method.backbone_calls", "relative.backbone_call_reduction_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "memory-grafting", "title": "Memory Grafting: Scaling Language Model Pre-training via Offline Conditional Memory", "paper_url": "https://arxiv.org/abs/2605.20948", "detail_path": "reproductions/2605.20948-memory-grafting/README.md", "topic": [ "llm-pretraining", "conditional-memory", "offline-knowledge-transfer" ], "first_author": null, "first_author_affiliation": "Tsinghua University / Microsoft Research Asia", "published": "2026-05-20", "code": null, "adapter": { "adapter_key": "memory-grafting", "arxiv_id": "2605.20948", "title": "Memory Grafting: Scaling Language Model Pre-training via Offline Conditional Memory", "paper_url": "https://arxiv.org/abs/2605.20948", "track": "llm", "organization": "Tsinghua University / Microsoft Research Asia", "published": "2026-05-20", "code_url": null, "topics": [ "llm-pretraining", "conditional-memory", "offline-knowledge-transfer" ], "local_code_dir": "src/auto_research/reproductions/memory_grafting", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同结构普通 Transformer", "metrics": [ "paper_benchmark_average.Memory Grafting", "relative.ppl_reduction_vs_engram_percent", "relative.ppl_reduction_vs_transformer_percent", "setup.memory_capacity", "variants.Engram.loss", "variants.Engram.perplexity", "variants.Memory Grafting.exact_hit_rate", "variants.Memory Grafting.loss", "variants.Memory Grafting.perplexity", "variants.Memory Grafting.trainable_parameters", "variants.Transformer.loss", "variants.Transformer.perplexity" ], "default_seeds": [ 42 ], "budget": "steps=45", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "mhc", "title": "mHC: Manifold-Constrained Hyper-Connections", "paper_url": "https://arxiv.org/abs/2512.24880", "detail_path": "reproductions/2512.24880-mhc/README.md", "topic": [ "llm-architecture", "hyper-connections", "training-stability" ], "first_author": null, "first_author_affiliation": "DeepSeek-AI", "published": "2025-12-31", "code": null, "adapter": { "adapter_key": "mhc", "arxiv_id": "2512.24880", "title": "mHC: Manifold-Constrained Hyper-Connections", "paper_url": "https://arxiv.org/abs/2512.24880", "track": "llm", "organization": "DeepSeek-AI", "published": "2025-12-31", "code_url": null, "topics": [ "llm-architecture", "hyper-connections", "training-stability" ], "local_code_dir": "src/auto_research/reproductions/mhc", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同 token、优化器和主干宽度的普通 Transformer", "metrics": [ "relative.ppl_reduction_vs_hc_percent", "relative.ppl_reduction_vs_transformer_percent", "variants.HC.loss", "variants.HC.perplexity", "variants.HC.spectral_norm_max", "variants.Transformer.loss", "variants.Transformer.perplexity", "variants.mHC.loss", "variants.mHC.perplexity", "variants.mHC.spectral_norm_max" ], "default_seeds": [ 42 ], "budget": "steps=45", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "minimax-sparse-attention", "title": "MiniMax Sparse Attention", "paper_url": "https://arxiv.org/abs/2606.13392", "detail_path": "reproductions/2606.13392-minimax-sparse-attention/README.md", "topic": [ "llm", "sparse-attention", "gqa", "long-context" ], "first_author": null, "first_author_affiliation": "MiniMax", "published": "2026-06-11", "code": "https://github.com/MiniMax-AI/MSA", "adapter": { "adapter_key": "minimax-sparse-attention", "arxiv_id": "2606.13392", "title": "MiniMax Sparse Attention", "paper_url": "https://arxiv.org/abs/2606.13392", "track": "llm", "organization": "MiniMax", "published": "2026-06-11", "code_url": "https://github.com/MiniMax-AI/MSA", "topics": [ "llm", "sparse-attention", "gqa", "long-context" ], "local_code_dir": "src/auto_research/reproductions/minimax_sparse_attention", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "dense GQA，实验组为带 index 辅助损失的 MiniMax block-sparse GQA", "metrics": [ "baseline.attention_pair_ratio", "baseline.perplexity", "baseline.seconds", "baseline.validation_loss", "method.attention_pair_ratio", "method.perplexity", "method.seconds", "method.validation_loss", "relative.attention_pairs_percent", "relative.measured_training_seconds_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=36", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "mllmclip", "title": "MLLMCLIP: Feature-Level Distillation of MLLM for Robust Vision-Language Representations", "paper_url": "https://arxiv.org/abs/2608.25575", "detail_path": "reproductions/2608.25575-mllmclip/README.md", "topic": [ "multimodal-foundation-model", "feature-distillation", "compositionality", "cka" ], "first_author": null, "first_author_affiliation": "KAIST / Sony AI", "published": "2026-08-26", "code": null, "adapter": { "adapter_key": "mllmclip", "arxiv_id": "2608.25575", "title": "MLLMCLIP: Feature-Level Distillation of MLLM for Robust Vision-Language Representations", "paper_url": "https://arxiv.org/abs/2608.25575", "track": "llm", "organization": "KAIST / Sony AI", "published": "2026-08-26", "code_url": null, "topics": [ "multimodal-foundation-model", "feature-distillation", "compositionality", "cka" ], "local_code_dir": "src/auto_research/reproductions/mllmclip", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "POPE adversarial / COCO val2014 checkpoint subset", "MovieLens-1M diagnostic proxy" ], "baseline": "frozen CLIP features with a validation-selected ridge projection", "metrics": [ "linear CKA", "neighbor overlap@5", "training loss", "peak GPU memory" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=40", "device_capabilities": [ "cuda" ], "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/mllmclip-a100-20260828.json", "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "moba", "title": "MoBA: Mixture of Block Attention for Long-Context LLMs", "paper_url": "https://arxiv.org/abs/2502.13189", "detail_path": "reproductions/2502.13189-moba/README.md", "topic": [ "sparse-attention", "long-context", "block-routing" ], "first_author": null, "first_author_affiliation": "Moonshot AI", "published": "2025-02-18", "code": "https://github.com/MoonshotAI/MoBA", "adapter": { "adapter_key": "moba", "arxiv_id": "2502.13189", "title": "MoBA: Mixture of Block Attention for Long-Context LLMs", "paper_url": "https://arxiv.org/abs/2502.13189", "track": "llm", "organization": "Moonshot AI", "published": "2025-02-18", "code_url": "https://github.com/MoonshotAI/MoBA", "topics": [ "sparse-attention", "long-context", "block-routing" ], "local_code_dir": "src/auto_research/reproductions/moba", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同预算 llamamodern，实验组为 moba", "metrics": [ "baseline.composite_loss", "baseline.final_loss", "baseline.fitness", "baseline.initial_loss", "baseline.instruction_loss", "baseline.lm_loss", "baseline.muon_orthogonality_error", "baseline.parameters", "baseline.perplexity", "baseline.post_training_loss", "baseline.primary", "baseline.public_composite", "baseline.reward_bearing_group_rate", "method.architecture_stats.block_size", "method.architecture_stats.selected_block_fraction", "method.composite_loss", "method.final_loss", "method.fitness", "method.initial_loss", "method.instruction_loss", "method.lm_loss", "method.muon_orthogonality_error", "method.parameters", "method.perplexity", "method.post_training_loss", "method.primary", "method.public_composite", "method.reward_bearing_group_rate", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=12", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "mobius-rope", "title": "Anti-Periodic Positional Encoding: Möbius Boundary Conditions Make In-Context Retrieval Reliable", "paper_url": "https://arxiv.org/abs/2607.21405", "detail_path": "reproductions/2607.21405-mobius-rope/README.md", "topic": [ "llm-architecture", "positional-encoding", "long-context" ], "first_author": null, "first_author_affiliation": "Independent researcher", "published": "2026-07-23", "code": null, "adapter": { "adapter_key": "mobius-rope", "arxiv_id": "2607.21405", "title": "Anti-Periodic Positional Encoding: Möbius Boundary Conditions Make In-Context Retrieval Reliable", "paper_url": "https://arxiv.org/abs/2607.21405", "track": "llm", "organization": "Independent researcher", "published": "2026-07-23", "code_url": null, "topics": [ "llm-architecture", "positional-encoding", "long-context" ], "local_code_dir": "src/auto_research/reproductions/mobius_rope", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2 + synthetic single-needle retrieval" ], "baseline": "standard RoPE", "metrics": [ "hybrid_mobius_rope.far_needle_accuracy", "hybrid_mobius_rope.needle_accuracy", "hybrid_mobius_rope.perplexity", "relative.far_needle_accuracy_points", "relative.needle_accuracy_points", "relative.perplexity_reduction_percent", "standard_rope.far_needle_accuracy", "standard_rope.needle_accuracy", "standard_rope.perplexity" ], "default_seeds": [ 42 ], "budget": "steps=90", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "muon", "title": "Muon is Scalable for LLM Training", "paper_url": "https://arxiv.org/abs/2502.16982", "detail_path": "reproductions/2502.16982-muon/README.md", "topic": [ "llm-training", "optimizer", "training-efficiency" ], "first_author": null, "first_author_affiliation": "Moonshot AI / UCLA", "published": "2025-02-24", "code": "https://github.com/MoonshotAI/Moonlight", "adapter": { "adapter_key": "muon", "arxiv_id": "2502.16982", "title": "Muon is Scalable for LLM Training", "paper_url": "https://arxiv.org/abs/2502.16982", "track": "llm", "organization": "Moonshot AI / UCLA", "published": "2025-02-24", "code_url": "https://github.com/MoonshotAI/Moonlight", "topics": [ "llm-training", "optimizer", "training-efficiency" ], "local_code_dir": "src/auto_research/reproductions/muon", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "matched control without the muon mechanism", "metrics": [ "baseline.lm_loss", "baseline.parameters", "baseline.perplexity", "method.lm_loss", "method.orthogonality_error", "method.parameters", "method.perplexity", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=30", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "musec", "title": "Musec: MomentUm SpEctral Clipping for Stable Muon-type Training", "paper_url": "https://arxiv.org/abs/2609.11655", "detail_path": "foundation-models/2609.11655-musec/README.md", "topic": [ "优化器", "谱裁剪" ], "first_author": "Zhuanghua Liu", "first_author_affiliation": "National University of Singapore", "published": "2026-09-10", "code": null, "adapter": "musec", "priority": "P1" }, { "domain": "foundation-models", "key": "naju", "title": "Naju: A Native Discrete State-Space Model with Independent Retention and Writing for Long-Sequence Memory", "paper_url": "https://arxiv.org/abs/2607.21000", "detail_path": "reproductions/2607.21000-naju/README.md", "topic": [ "llm-architecture", "state-space-model", "long-context" ], "first_author": null, "first_author_affiliation": "Independent researchers", "published": "2026-07-23", "code": null, "adapter": { "adapter_key": "naju", "arxiv_id": "2607.21000", "title": "Naju: A Native Discrete State-Space Model with Independent Retention and Writing for Long-Sequence Memory", "paper_url": "https://arxiv.org/abs/2607.21000", "track": "llm", "organization": "Independent researchers", "published": "2026-07-23", "code_url": null, "topics": [ "llm-architecture", "state-space-model", "long-context" ], "local_code_dir": "src/auto_research/reproductions/naju", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同 token、optimizer、70 steps 的 96-d、3-layer Transformer", "metrics": [ "naju.loss", "naju.perplexity", "relative.perplexity_reduction_percent", "transformer.loss", "transformer.perplexity" ], "default_seeds": [ 42 ], "budget": "steps=70", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "native-sparse-attention", "title": "Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention", "paper_url": "https://arxiv.org/abs/2502.11089", "detail_path": "reproductions/2502.11089-native-sparse-attention/README.md", "topic": [ "llm-architecture", "sparse-attention", "long-context" ], "first_author": null, "first_author_affiliation": "DeepSeek", "published": "2025-02-16", "code": null, "adapter": { "adapter_key": "native-sparse-attention", "arxiv_id": "2502.11089", "title": "Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention", "paper_url": "https://arxiv.org/abs/2502.11089", "track": "llm", "organization": "DeepSeek", "published": "2025-02-16", "code_url": null, "topics": [ "llm-architecture", "sparse-attention", "long-context" ], "local_code_dir": "src/auto_research/reproductions/native_sparse_attention", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同维度、层数、token、AdamW 和 30-step 预算的 llamamodern", "metrics": [ "baseline.lm_loss", "baseline.parameters", "baseline.perplexity", "method.attention_edge_fraction", "method.compression_block", "method.lm_loss", "method.local_window", "method.parameters", "method.perplexity", "method.selected_blocks", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=30", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "oda", "title": "On-Demand Attention: Efficient Long-Context Decoding with Learned Recall", "paper_url": "https://arxiv.org/abs/2609.20734", "detail_path": "foundation-models/2609.20734-oda/README.md", "topic": [ "注意力与长上下文", "推理效率" ], "first_author": "Paper author team", "first_author_affiliation": "Southern University of Science and Technology", "published": "2026-09-17", "code": null, "adapter": "oda", "priority": "P0", "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/oda-a100-20260919.json" }, { "domain": "foundation-models", "key": "olmpool-long-context", "title": "Cracks in the Foundation: Seemingly Minor Architectural Choices Impact Long Context Extension", "paper_url": "https://arxiv.org/abs/2608.10296", "detail_path": "reproductions/2608.10296-olmpool-long-context/README.md", "topic": [ "注意力与长上下文", "网络架构" ], "first_author": null, "first_author_affiliation": "Ai2", "published": "2026-08-10", "code": "https://github.com/allenai/olmpool/", "adapter": { "adapter_key": "olmpool-long-context", "arxiv_id": "2608.10296", "title": "Cracks in the Foundation: Seemingly Minor Architectural Choices Impact Long Context Extension", "paper_url": "https://arxiv.org/abs/2608.10296", "track": "llm", "organization": "Ai2", "published": "2026-08-10", "code_url": "https://github.com/allenai/olmpool/", "topics": [ "注意力与长上下文", "网络架构" ], "local_code_dir": "src/auto_research/reproductions/olmpool_long_context", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "deterministic long-context public mini-suite" ], "baseline": "recent-window attention", "metrics": [ "accuracy", "relative cost" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "omnikvquant", "title": "OmniKVQuant: KV Cache Quantization for Omni-LLMs", "paper_url": "https://arxiv.org/abs/2609.11582", "detail_path": "foundation-models/2609.11582-omnikvquant/README.md", "topic": [ "多模态推理", "KV Cache 量化" ], "first_author": "Suho Yoo", "first_author_affiliation": "KAIST", "published": "2026-09-10", "code": "https://github.com/kaistmm/OmniKVQuant", "adapter": "omnikvquant", "priority": "P0" }, { "domain": "foundation-models", "key": "open-1b-audit", "title": "OPEN-1B: A Fully Auditable Training Run", "paper_url": "https://arxiv.org/abs/2609.17380", "detail_path": "foundation-models/2609.17380-open-1b-audit/README.md", "topic": [ "可复现训练", "训练审计" ], "first_author": "John Donaghy", "first_author_affiliation": "Gensyn", "published": "2026-09-15", "code": "https://github.com/gensyn-ai/open-transformers", "code_extra": "https://github.com/gensyn-ai/pretraining-audit-cli", "adapter": "open-1b-audit", "priority": "P1" }, { "domain": "foundation-models", "key": "open-language-model", "title": "OpenLanguageModel: Readable and Composable Small-Language-Model Pretraining for Education and Research", "paper_url": "https://arxiv.org/abs/2607.16669", "detail_path": "reproductions/2607.16669-open-language-model/README.md", "topic": [ "pretraining-infrastructure", "composable-architecture", "small-language-model", "evolve" ], "first_author": null, "first_author_affiliation": "Indian Institute of Technology Madras", "published": "2026-07-18", "code": "https://github.com/openlanguagemodel/openlanguagemodel", "adapter": { "adapter_key": "open-language-model", "arxiv_id": "2607.16669", "title": "OpenLanguageModel: Readable and Composable Small-Language-Model Pretraining for Education and Research", "paper_url": "https://arxiv.org/abs/2607.16669", "track": "llm", "organization": "Indian Institute of Technology Madras", "published": "2026-07-18", "code_url": "https://github.com/openlanguagemodel/openlanguagemodel", "topics": [ "pretraining-infrastructure", "composable-architecture", "small-language-model", "evolve" ], "local_code_dir": "src/auto_research/reproductions/open_language_model", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同预算 LLaMA-modern，实验组为 olmcomposable", "metrics": [ "baseline.composite_loss", "baseline.final_loss", "baseline.fitness", "baseline.initial_loss", "baseline.instruction_loss", "baseline.lm_loss", "baseline.muon_orthogonality_error", "baseline.parameters", "baseline.perplexity", "baseline.post_training_loss", "baseline.primary", "baseline.public_composite", "baseline.reward_bearing_group_rate", "method.architecture_stats.ordinary_pytorch_modules", "method.composite_loss", "method.final_loss", "method.fitness", "method.initial_loss", "method.instruction_loss", "method.lm_loss", "method.muon_orthogonality_error", "method.parameters", "method.perplexity", "method.post_training_loss", "method.primary", "method.public_composite", "method.reward_bearing_group_rate", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=30", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "pace-vlm", "title": "PACE: A Unified Condense-and-Extract Paradigm for Fast VLM Inference", "paper_url": "https://arxiv.org/abs/2608.27206", "detail_path": "reproductions/2608.27206-pace-vlm/README.md", "topic": [ "multimodal-foundation-model", "visual-token-compression", "inference-serving" ], "first_author": null, "first_author_affiliation": "Sun Yat-sen University", "published": "2026-08-27", "code": "https://github.com/jjL357/PACE", "adapter": { "adapter_key": "pace-vlm", "arxiv_id": "2608.27206", "title": "PACE: A Unified Condense-and-Extract Paradigm for Fast VLM Inference", "paper_url": "https://arxiv.org/abs/2608.27206", "track": "llm", "organization": "Sun Yat-sen University", "published": "2026-08-27", "code_url": "https://github.com/jjL357/PACE", "topics": [ "multimodal-foundation-model", "visual-token-compression", "inference-serving" ], "local_code_dir": "src/auto_research/reproductions/pace_vlm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "RealWorldQA checkpoint subset" ], "baseline": "Qwen2.5-VL at the original visual-token budget", "metrics": [ "answer accuracy", "visual tokens", "latency", "peak GPU memory" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cuda" ], "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/pace-vlm-a30-20260829.json", "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "penelope", "title": "Penelope: Localized Latent Recurrence for Efficient Structured Reasoning", "paper_url": "https://arxiv.org/abs/2607.25915", "detail_path": "reproductions/2607.25915-penelope/README.md", "topic": [ "llm-architecture", "latent-reasoning", "recurrence", "efficiency" ], "first_author": null, "first_author_affiliation": "Academic author team", "published": "2026-07-28", "code": null, "adapter": { "adapter_key": "penelope", "arxiv_id": "2607.25915", "title": "Penelope: Localized Latent Recurrence for Efficient Structured Reasoning", "paper_url": "https://arxiv.org/abs/2607.25915", "track": "llm", "organization": "Academic author team", "published": "2026-07-28", "code_url": null, "topics": [ "llm-architecture", "latent-reasoning", "recurrence", "efficiency" ], "local_code_dir": "src/auto_research/reproductions/penelope", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "llamamodern，实验组为 penelope", "metrics": [ "baseline.composite_loss", "baseline.final_loss", "baseline.fitness", "baseline.initial_loss", "baseline.instruction_loss", "baseline.lm_loss", "baseline.muon_orthogonality_error", "baseline.parameters", "baseline.perplexity", "baseline.post_training_loss", "baseline.primary", "baseline.public_composite", "baseline.reward_bearing_group_rate", "method.architecture_stats.full_decoder_reexecution", "method.architecture_stats.latent_recurrence_steps", "method.composite_loss", "method.final_loss", "method.fitness", "method.initial_loss", "method.instruction_loss", "method.lm_loss", "method.muon_orthogonality_error", "method.parameters", "method.perplexity", "method.post_training_loss", "method.primary", "method.public_composite", "method.reward_bearing_group_rate", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=12", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "persistent-recurrent-memory", "title": "Persistent Recurrent Memory Between Transformer Layers Improves Language Model Generalization", "paper_url": "https://arxiv.org/abs/2609.17251", "detail_path": "foundation-models/2609.17251-persistent-recurrent-memory/README.md", "topic": [ "网络架构", "持久记忆" ], "first_author": "Eduardo Novaes Hering", "first_author_affiliation": "FITec Labs / Ericsson São Paulo", "published": "2026-09-15", "code": null, "adapter": "persistent-recurrent-memory", "priority": "P1" }, { "domain": "foundation-models", "key": "physics-mm-pretraining", "title": "Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes", "paper_url": "https://arxiv.org/abs/2608.05000", "detail_path": "reproductions/2608.05000-physics-mm-pretraining/README.md", "topic": [ "多模态基础模型", "预训练与数据" ], "first_author": null, "first_author_affiliation": "Meta", "published": "2026-08-05", "code": null, "adapter": { "adapter_key": "physics-mm-pretraining", "arxiv_id": "2608.05000", "title": "Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes", "paper_url": "https://arxiv.org/abs/2608.05000", "track": "llm", "organization": "Meta", "published": "2026-08-05", "code_url": null, "topics": [ "多模态基础模型", "预训练与数据" ], "local_code_dir": "src/auto_research/reproductions/physics_mm_pretraining", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "deterministic long-context public mini-suite" ], "baseline": "recent-window attention", "metrics": [ "accuracy", "relative cost" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "ppl-factory", "title": "PPL-Factory: Task-Aware and Budget-Aware Data Selection from Language Modeling to Reasoning", "paper_url": "https://arxiv.org/abs/2607.18199", "detail_path": "reproductions/2607.18199-ppl-factory/README.md", "topic": [ "llm-training", "data-selection", "efficient-finetuning" ], "first_author": null, "first_author_affiliation": "McGill University", "published": "2026-07-20", "code": null, "adapter": { "adapter_key": "ppl-factory", "arxiv_id": "2607.18199", "title": "PPL-Factory: Task-Aware and Budget-Aware Data Selection from Language Modeling to Reasoning", "paper_url": "https://arxiv.org/abs/2607.18199", "track": "llm", "organization": "McGill University", "published": "2026-07-20", "code_url": null, "topics": [ "llm-training", "data-selection", "efficient-finetuning" ], "local_code_dir": "src/auto_research/reproductions/ppl_factory", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "20% 随机 block", "metrics": [ "perplexity_reduction_vs_random_percent", "variants.Easy-NLL.perplexity", "variants.Easy-NLL.test_loss", "variants.PPL-Factory.perplexity", "variants.PPL-Factory.test_loss", "variants.Random.perplexity", "variants.Random.test_loss" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "qevict", "title": "QEvict: Recoverable Quantized KV Eviction for Attention-Drift-Robust Long-Context Decoding", "paper_url": "https://arxiv.org/abs/2608.05326", "detail_path": "reproductions/2608.05326-qevict/README.md", "topic": [ "long-context", "kv-cache", "quantization", "serving" ], "first_author": null, "first_author_affiliation": "Indian Institute of Technology Roorkee", "published": "2026-08-05", "code": null, "adapter": { "adapter_key": "qevict", "arxiv_id": "2608.05326", "title": "QEvict: Recoverable Quantized KV Eviction for Attention-Drift-Robust Long-Context Decoding", "paper_url": "https://arxiv.org/abs/2608.05326", "track": "llm", "organization": "Indian Institute of Technology Roorkee", "published": "2026-08-05", "code_url": null, "topics": [ "long-context", "kv-cache", "quantization", "serving" ], "local_code_dir": "src/auto_research/reproductions/qevict", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "deterministic public-style mini-suite" ], "baseline": "去掉论文特有机制、其余数据切分与预算相同的 matched control", "metrics": [ "binary_evict_recall", "full_precision_slots", "qevict_recall" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "r4t", "title": "Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion", "paper_url": "https://arxiv.org/abs/2603.06397", "detail_path": "foundation-models/2603.06397-r4t/README.md", "topic": [ "生成式检索", "RL 编译" ], "first_author": "Pengcheng Jiang", "first_author_affiliation": "University of Illinois Urbana-Champaign / Google Research", "published": "2026-03-06", "code": null, "adapter": "r4t", "priority": "P0" }, { "domain": "foundation-models", "key": "random-attention", "title": "Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning", "paper_url": "https://arxiv.org/abs/2609.03430", "detail_path": "reproductions/2609.03430-random-attention/README.md", "topic": [ "kv-cache", "reasoning", "inference-serving", "compression" ], "first_author": null, "first_author_affiliation": "Salesforce AI Research / University of Illinois Urbana-Champaign", "published": "2026-09-03", "code": "https://github.com/SalesforceAIResearch/Random-Attention", "adapter": { "adapter_key": "random-attention", "arxiv_id": "2609.03430", "title": "Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning", "paper_url": "https://arxiv.org/abs/2609.03430", "track": "llm", "organization": "Salesforce AI Research / University of Illinois Urbana-Champaign", "published": "2026-09-03", "code_url": "https://github.com/SalesforceAIResearch/Random-Attention", "topics": [ "kv-cache", "reasoning", "inference-serving", "compression" ], "local_code_dir": "src/auto_research/reproductions/random_attention", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2 real-checkpoint subset" ], "baseline": "prompt-protected recent-token eviction at equal budget", "metrics": [ "attention-output cosine", "selection latency", "KV bytes", "peak GPU memory" ], "default_seeds": [ 42, 43, 44 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cuda" ], "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/random-attention-a100-20260906.json", "online_evidence": [], "selection_exception": null, "evolve_operators": [ "attention:prompt-protected-random-eviction" ] } }, { "domain": "foundation-models", "key": "rare", "title": "RARE: Decoupling Representation Steering from Expert Routing in Mixture-of-Experts Language Models", "paper_url": "https://arxiv.org/abs/2608.21236", "detail_path": "reproductions/2608.21236-rare/README.md", "topic": [ "llm-architecture", "mixture-of-experts", "representation-steering", "model-editing" ], "first_author": null, "first_author_affiliation": "Huazhong University of Science and Technology", "published": "2026-08-21", "code": null, "adapter": { "adapter_key": "rare", "arxiv_id": "2608.21236", "title": "RARE: Decoupling Representation Steering from Expert Routing in Mixture-of-Experts Language Models", "paper_url": "https://arxiv.org/abs/2608.21236", "track": "llm", "organization": "Huazhong University of Science and Technology", "published": "2026-08-21", "code_url": null, "topics": [ "llm-architecture", "mixture-of-experts", "representation-steering", "model-editing" ], "local_code_dir": "src/auto_research/reproductions/rare", "fidelity": "core_mechanism", "evaluation_tier": "l1_mechanism", "datasets": [ "deterministic MoE routing mini-suite" ], "baseline": "unprojected representation steering", "metrics": [ "route agreement", "route flip rate", "steering gain" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "rd-attnres", "title": "Role-Decoupled Attention Residuals", "paper_url": "https://arxiv.org/abs/2608.01075", "detail_path": "reproductions/2608.01075-rd-attnres/README.md", "topic": [ "foundation-model", "attention", "residual-routing" ], "first_author": null, "first_author_affiliation": "Kehan Wang（论文未列机构）", "published": "2026-08-03", "code": null, "adapter": { "adapter_key": "rd-attnres", "arxiv_id": "2608.01075", "title": "Role-Decoupled Attention Residuals", "paper_url": "https://arxiv.org/abs/2608.01075", "track": "llm", "organization": "Kehan Wang（论文未列机构）", "published": "2026-08-03", "code_url": null, "topics": [ "foundation-model", "attention", "residual-routing" ], "local_code_dir": "src/auto_research/reproductions/rd_attnres", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "共享深度路由的 Block AttnRes，实验组仅解耦 QK/V route", "metrics": [ "baseline.lm_loss", "baseline.parameters", "baseline.perplexity", "method.lm_loss", "method.parameters", "method.perplexity", "method.qk_v_route_js_divergence", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=30", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "register-tokens-dllm", "title": "Register Tokens for Bounded-State Reasoning in Diffusion Language Models", "paper_url": "https://arxiv.org/abs/2609.16372", "detail_path": "foundation-models/2609.16372-register-tokens-dllm/README.md", "topic": [ "扩散语言模型", "长程推理" ], "first_author": "Albert Ge", "first_author_affiliation": "University of Wisconsin–Madison", "published": "2026-09-14", "code": null, "adapter": "register-tokens-dllm", "priority": "P1" }, { "domain": "foundation-models", "key": "repeat-aware-moe", "title": "Data Scarcity and Model Sparsity: Mixtures-of-Experts Overfit More to Repeated Data", "paper_url": "https://arxiv.org/abs/2609.11917", "detail_path": "foundation-models/2609.11917-repeat-aware-moe/README.md", "topic": [ "MoE", "数据重复与正则化" ], "first_author": "Atindra Jha", "first_author_affiliation": "Stanford University", "published": "2026-09-10", "code": null, "adapter": "repeat-aware-moe", "priority": "P1" }, { "domain": "foundation-models", "key": "retoken", "title": "ReToken: One Token to Improve Vision–Language Models for Visual Retrieval", "paper_url": "https://arxiv.org/abs/2607.28627", "detail_path": "reproductions/2607.28627-retoken/README.md", "topic": [ "vision-language-model", "visual-retrieval", "kv-cache", "sparse-attention" ], "first_author": null, "first_author_affiliation": "UIUC / Microsoft Research / Google DeepMind", "published": "2026-07-30", "code": "https://github.com/avaxiao/ReToken", "adapter": { "adapter_key": "retoken", "arxiv_id": "2607.28627", "title": "ReToken: One Token to Improve Vision–Language Models for Visual Retrieval", "paper_url": "https://arxiv.org/abs/2607.28627", "track": "llm", "organization": "UIUC / Microsoft Research / Google DeepMind", "published": "2026-07-30", "code_url": "https://github.com/avaxiao/ReToken", "topics": [ "vision-language-model", "visual-retrieval", "kv-cache", "sparse-attention" ], "local_code_dir": "src/auto_research/reproductions/retoken", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同预算 dense LLaMA，实验组为 ReToken value-cache Top-K", "metrics": [ "baseline.lm_loss", "baseline.parameters", "baseline.perplexity", "method.causal_cache_keep_rate", "method.lm_loss", "method.parameters", "method.perplexity", "method.retrieval_target_tokens", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=30", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "rope", "title": "RoFormer: Enhanced Transformer with Rotary Position Embedding", "paper_url": "https://arxiv.org/abs/2104.09864", "detail_path": "reproductions/2104.09864-rope/README.md", "topic": [ "position-encoding", "long-context", "architecture" ], "first_author": null, "first_author_affiliation": "Zhuiyi Technology", "published": "2021-04-20", "code": "https://github.com/ZhuiyiTechnology/roformer", "adapter": { "adapter_key": "rope", "arxiv_id": "2104.09864", "title": "RoFormer: Enhanced Transformer with Rotary Position Embedding", "paper_url": "https://arxiv.org/abs/2104.09864", "track": "llm", "organization": "Zhuiyi Technology", "published": "2021-04-20", "code_url": "https://github.com/ZhuiyiTechnology/roformer", "topics": [ "position-encoding", "long-context", "architecture" ], "local_code_dir": "src/auto_research/reproductions/rope", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同预算 llamamodern，实验组为 rope", "metrics": [ "baseline.composite_loss", "baseline.final_loss", "baseline.fitness", "baseline.initial_loss", "baseline.instruction_loss", "baseline.lm_loss", "baseline.muon_orthogonality_error", "baseline.parameters", "baseline.perplexity", "baseline.post_training_loss", "baseline.primary", "baseline.public_composite", "baseline.reward_bearing_group_rate", "method.architecture_stats.relative_phase", "method.composite_loss", "method.final_loss", "method.fitness", "method.initial_loss", "method.instruction_loss", "method.lm_loss", "method.muon_orthogonality_error", "method.parameters", "method.perplexity", "method.post_training_loss", "method.primary", "method.public_composite", "method.reward_bearing_group_rate", "relative.lm_loss_percent", "relative.perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=12", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "sas-attention", "title": "SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking", "paper_url": "https://arxiv.org/abs/2609.13141", "detail_path": "reproductions/2609.13141-sas-attention/README.md", "topic": [ "sparse-attention", "long-context", "post-training", "context-ranking" ], "first_author": null, "first_author_affiliation": "Tencent HY LLM Frontier / HKUST (Guangzhou) / HKUST", "published": "2026-09-11", "code": "https://github.com/Tencent-Hunyuan/Simple-Attention-Sparsification", "adapter": { "adapter_key": "sas-attention", "arxiv_id": "2609.13141", "title": "SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking", "paper_url": "https://arxiv.org/abs/2609.13141", "track": "llm", "organization": "Tencent HY LLM Frontier / HKUST (Guangzhou) / HKUST", "published": "2026-09-11", "code_url": "https://github.com/Tencent-Hunyuan/Simple-Attention-Sparsification", "topics": [ "sparse-attention", "long-context", "post-training", "context-ranking" ], "local_code_dir": "src/auto_research/reproductions/sas_attention", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "same trained tiny decoder with full causal attention", "metrics": [ "test.sas.perplexity", "test.dense.perplexity", "routing.retained_attention_fraction", "selector_training.selector_gradient_norm_mean" ], "default_seeds": [ 42, 43, 44 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "sensenova-u1-5", "title": "SenseNova-U1.5: Towards Native Unified Visual Intelligence", "paper_url": "https://arxiv.org/abs/2609.11929", "detail_path": "foundation-models/2609.11929-sensenova-u1-5/README.md", "topic": [ "统一多模态", "视觉生成与理解" ], "first_author": "SenseNova Team", "first_author_affiliation": "SenseTime Research", "published": "2026-09-10", "code": "https://github.com/OpenSenseNova/SenseNova-U1", "adapter": "sensenova-u1-5", "priority": "P1" }, { "domain": "foundation-models", "key": "siglip2", "title": "SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features", "paper_url": "https://arxiv.org/abs/2502.14786", "detail_path": "reproductions/2502.14786-siglip2/README.md", "topic": [ "multimodal", "sigmoid-contrastive", "self-distillation" ], "first_author": null, "first_author_affiliation": "Google DeepMind", "published": "2025-02-20", "code": "https://github.com/google-research/big_vision", "adapter": { "adapter_key": "siglip2", "arxiv_id": "2502.14786", "title": "SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features", "paper_url": "https://arxiv.org/abs/2502.14786", "track": "llm", "organization": "Google DeepMind", "published": "2025-02-20", "code_url": "https://github.com/google-research/big_vision", "topics": [ "multimodal", "sigmoid-contrastive", "self-distillation" ], "local_code_dir": "src/auto_research/reproductions/siglip2", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "Fashion-MNIST image-text pairs" ], "baseline": "uniform label retrieval", "metrics": [ "test_accuracy", "visual_dependency_delta" ], "default_seeds": [ 42 ], "budget": "steps=150", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "smolvlm", "title": "SmolVLM: Redefining small and efficient multimodal models", "paper_url": "https://arxiv.org/abs/2504.05299", "detail_path": "reproductions/2504.05299-smolvlm/README.md", "topic": [ "multimodal", "efficient-vlm", "visual-token-compression" ], "first_author": null, "first_author_affiliation": "Hugging Face", "published": "2025-04-07", "code": "https://github.com/huggingface/smollm", "adapter": { "adapter_key": "smolvlm", "arxiv_id": "2504.05299", "title": "SmolVLM: Redefining small and efficient multimodal models", "paper_url": "https://arxiv.org/abs/2504.05299", "track": "llm", "organization": "Hugging Face", "published": "2025-04-07", "code_url": "https://github.com/huggingface/smollm", "topics": [ "multimodal", "efficient-vlm", "visual-token-compression" ], "local_code_dir": "src/auto_research/reproductions/smolvlm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "Fashion-MNIST object QA" ], "baseline": "linear mean-pooled connector", "metrics": [ "test_accuracy", "visual_dependency_delta", "visual_tokens" ], "default_seeds": [ 42 ], "budget": "steps=120", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "speculative-decoding", "title": "Fast Inference from Transformers via Speculative Decoding", "paper_url": "https://arxiv.org/abs/2211.17192", "detail_path": "reproductions/2211.17192-speculative-decoding/README.md", "topic": [ "inference-serving", "decoding" ], "first_author": null, "first_author_affiliation": "Google Research", "published": "2022-11-30", "code": "https://github.com/google-research/google-research/tree/master/speculative_decoding", "adapter": { "adapter_key": "speculative-decoding", "arxiv_id": "2211.17192", "title": "Fast Inference from Transformers via Speculative Decoding", "paper_url": "https://arxiv.org/abs/2211.17192", "track": "llm", "organization": "Google Research", "published": "2022-11-30", "code_url": "https://github.com/google-research/google-research/tree/master/speculative_decoding", "topics": [ "inference-serving", "decoding" ], "local_code_dir": "src/auto_research/reproductions/speculative_decoding", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "target-only greedy decoding，实验组为 four-token draft and exact target verification，只改变论文核心机制", "metrics": [ "baseline.target_calls", "method.target_calls", "relative.target_call_reduction_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "switch-attention", "title": "Switch Attention: Towards Dynamic and Fine-grained Hybrid Transformers", "paper_url": "https://arxiv.org/abs/2603.26380", "detail_path": "reproductions/2603.26380-switch-attention/README.md", "topic": [ "llm-architecture", "efficient-attention", "long-context" ], "first_author": null, "first_author_affiliation": "Peking University / Huawei Technologies", "published": "2026-03-27", "code": null, "adapter": { "adapter_key": "switch-attention", "arxiv_id": "2603.26380", "title": "Switch Attention: Towards Dynamic and Fine-grained Hybrid Transformers", "paper_url": "https://arxiv.org/abs/2603.26380", "track": "llm", "organization": "Peking University / Huawei Technologies", "published": "2026-03-27", "code_url": null, "topics": [ "llm-architecture", "efficient-attention", "long-context" ], "local_code_dir": "src/auto_research/reproductions/switch_attention", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "full-attention tiny Transformer", "metrics": [ "baseline.perplexity", "method.perplexity" ], "default_seeds": [ 42 ], "budget": "steps=30", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "switch-transformer", "title": "Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity", "paper_url": "https://arxiv.org/abs/2101.03961", "detail_path": "reproductions/2101.03961-switch-transformer/README.md", "topic": [ "llm-architecture", "mixture-of-experts", "classic" ], "first_author": null, "first_author_affiliation": "Google Brain", "published": "2021-01-11", "code": "https://github.com/tensorflow/mesh", "adapter": { "adapter_key": "switch-transformer", "arxiv_id": "2101.03961", "title": "Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity", "paper_url": "https://arxiv.org/abs/2101.03961", "track": "llm", "organization": "Google Brain", "published": "2021-01-11", "code_url": "https://github.com/tensorflow/mesh", "topics": [ "llm-architecture", "mixture-of-experts", "classic" ], "local_code_dir": "src/auto_research/reproductions/switch_transformer", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "参数预算内 dense Transformer", "metrics": [ "baseline.perplexity", "method.perplexity" ], "default_seeds": [ 42 ], "budget": "steps=30", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "swrouter", "title": "SWRouter: Similarity-Contractive Window Routing for Multi-Turn Large Language Model Conversations", "paper_url": "https://arxiv.org/abs/2609.11414", "detail_path": "foundation-models/2609.11414-swrouter/README.md", "topic": [ "模型路由", "多轮上下文" ], "first_author": "Yu Wang", "first_author_affiliation": "Shanghai Jiao Tong University", "published": "2026-09-10", "code": null, "adapter": "swrouter", "priority": "P1" }, { "domain": "foundation-models", "key": "tcab", "title": "Fast A/B/n Testing: Exact Multi-Policy Comparison via Tree-Coupled Feedback Sharing", "paper_url": "https://arxiv.org/abs/2608.12831", "detail_path": "reproductions/2608.12831-tcab/README.md", "topic": [ "评测基础设施", "在线实验" ], "first_author": null, "first_author_affiliation": "New York University", "published": "2026-08-13", "code": null, "adapter": { "adapter_key": "tcab", "arxiv_id": "2608.12831", "title": "Fast A/B/n Testing: Exact Multi-Policy Comparison via Tree-Coupled Feedback Sharing", "paper_url": "https://arxiv.org/abs/2608.12831", "track": "llm", "organization": "New York University", "published": "2026-08-13", "code_url": null, "topics": [ "评测基础设施", "在线实验" ], "local_code_dir": "src/auto_research/reproductions/tcab", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "deterministic long-context public mini-suite" ], "baseline": "recent-window attention", "metrics": [ "accuracy", "relative cost" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "transmem", "title": "TransMem: Transforming Hidden States into Memory for Large Language Models", "paper_url": "https://arxiv.org/abs/2607.29032", "detail_path": "reproductions/2607.29032-transmem/README.md", "topic": [ "注意力与长上下文", "Agent 记忆" ], "first_author": null, "first_author_affiliation": "Authors did not disclose affiliation", "published": "2026-07-31", "code": null, "adapter": { "adapter_key": "transmem", "arxiv_id": "2607.29032", "title": "TransMem: Transforming Hidden States into Memory for Large Language Models", "paper_url": "https://arxiv.org/abs/2607.29032", "track": "llm", "organization": "Authors did not disclose affiliation", "published": "2026-07-31", "code_url": null, "topics": [ "注意力与长上下文", "Agent 记忆" ], "local_code_dir": "src/auto_research/reproductions/transmem", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "deterministic long-context public mini-suite" ], "baseline": "recent-window attention", "metrics": [ "accuracy", "relative cost" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "ttcd", "title": "Learning What to Remember: Test-Time Training via Context Distillation", "paper_url": "https://arxiv.org/abs/2608.01672", "detail_path": "reproductions/2608.01672-ttcd/README.md", "topic": [ "注意力与长上下文", "测试时训练" ], "first_author": null, "first_author_affiliation": "Princeton University", "published": "2026-08-03", "code": "https://github.com/dangxingyu/ttcd", "adapter": { "adapter_key": "ttcd", "arxiv_id": "2608.01672", "title": "Learning What to Remember: Test-Time Training via Context Distillation", "paper_url": "https://arxiv.org/abs/2608.01672", "track": "llm", "organization": "Princeton University", "published": "2026-08-03", "code_url": "https://github.com/dangxingyu/ttcd", "topics": [ "注意力与长上下文", "测试时训练" ], "local_code_dir": "src/auto_research/reproductions/ttcd", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "deterministic long-context public mini-suite" ], "baseline": "recent-window attention", "metrics": [ "accuracy", "relative cost" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "twinkv", "title": "TwinKV: A Composable Repair Pass for KV Cache Eviction via Pairwise Key Redundancy", "paper_url": "https://arxiv.org/abs/2608.27128", "detail_path": "reproductions/2608.27128-twinkv/README.md", "topic": [ "kv-cache", "long-context", "inference-serving", "compression" ], "first_author": null, "first_author_affiliation": "The Hong Kong University of Science and Technology (Guangzhou)", "published": "2026-08-27", "code": null, "adapter": { "adapter_key": "twinkv", "arxiv_id": "2608.27128", "title": "TwinKV: A Composable Repair Pass for KV Cache Eviction via Pairwise Key Redundancy", "paper_url": "https://arxiv.org/abs/2608.27128", "track": "llm", "organization": "The Hong Kong University of Science and Technology (Guangzhou)", "published": "2026-08-27", "code_url": null, "topics": [ "kv-cache", "long-context", "inference-serving", "compression" ], "local_code_dir": "src/auto_research/reproductions/twinkv", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2 long-context checkpoint subset" ], "baseline": "StreamingLLM sink + recent eviction at the same KV budget", "metrics": [ "attention-output cosine", "repair latency", "KV bytes", "peak GPU memory" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cuda" ], "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/twinkv-a30-20260829.json", "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "vbvr-pro", "title": "VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning", "paper_url": "https://arxiv.org/abs/2608.26105", "detail_path": "reproductions/2608.26105-vbvr-pro/README.md", "topic": [ "multimodal-foundation-model", "native-visual-reasoning", "verifiable-reward", "benchmark" ], "first_author": null, "first_author_affiliation": "Nanyang Technological University / VBVR Community", "published": "2026-08-26", "code": "https://www.video-reason.com/?v=pro", "adapter": { "adapter_key": "vbvr-pro", "arxiv_id": "2608.26105", "title": "VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning", "paper_url": "https://arxiv.org/abs/2608.26105", "track": "llm", "organization": "Nanyang Technological University / VBVR Community", "published": "2026-08-26", "code_url": "https://www.video-reason.com/?v=pro", "topics": [ "multimodal-foundation-model", "native-visual-reasoning", "verifiable-reward", "benchmark" ], "local_code_dir": "src/auto_research/reproductions/vbvr_pro", "fidelity": "core_mechanism", "evaluation_tier": "l1_mechanism", "datasets": [ "VBVR-Pro procedural task specification" ], "baseline": "scalar VLM-as-a-judge analogue on the same 300 generated tasks", "metrics": [ "reward mean", "reward standard deviation", "deterministic verifier agreement" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "wemm-embedding", "title": "WeMM-Embedding: WeChat Multi-Modal Embedding Technical Report", "paper_url": "https://arxiv.org/abs/2608.24053", "detail_path": "reproductions/2608.24053-wemm-embedding/README.md", "topic": [ "multimodal-foundation-model", "embedding", "retrieval", "matryoshka-representation" ], "first_author": "Junjie Zhou", "first_author_affiliation": "WeChat Vision, Tencent", "published": "2026-08-25", "code": "https://github.com/Tencent/WeMM-Embedding", "adapter": { "adapter_key": "wemm-embedding", "arxiv_id": "2608.24053", "title": "WeMM-Embedding: WeChat Multi-Modal Embedding Technical Report", "paper_url": "https://arxiv.org/abs/2608.24053", "track": "llm", "organization": "WeChat Vision, Tencent", "published": "2026-08-25", "code_url": "https://github.com/Tencent/WeMM-Embedding", "topics": [ "multimodal-foundation-model", "embedding", "retrieval", "matryoshka-representation" ], "local_code_dir": "src/auto_research/reproductions/wemm_embedding", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M paired content/collaborative views" ], "baseline": "stage-1 multimodal alignment without refinement", "metrics": [ "Recall@1", "Recall@10", "MRR" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "wide", "title": "WIDE: Boosting Adaptive LLM Inference via Token-level Dynamic Width Pruning", "paper_url": "https://arxiv.org/abs/2607.28418", "detail_path": "reproductions/2607.28418-wide/README.md", "topic": [ "llm-architecture", "dynamic-pruning", "inference-efficiency" ], "first_author": null, "first_author_affiliation": "EIT-NLP / LMU Munich", "published": "2026-07-30", "code": "https://github.com/EIT-NLP/LLM-Pruning/tree/main/WIDE", "adapter": { "adapter_key": "wide", "arxiv_id": "2607.28418", "title": "WIDE: Boosting Adaptive LLM Inference via Token-level Dynamic Width Pruning", "paper_url": "https://arxiv.org/abs/2607.28418", "track": "llm", "organization": "EIT-NLP / LMU Munich", "published": "2026-07-30", "code_url": "https://github.com/EIT-NLP/LLM-Pruning/tree/main/WIDE", "topics": [ "llm-architecture", "dynamic-pruning", "inference-efficiency" ], "local_code_dir": "src/auto_research/reproductions/wide", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "同预算 dense LLaMA", "metrics": [ "baseline.lm_loss", "baseline.parameters", "baseline.perplexity", "method.active_attention_head_fraction", "method.active_ffn_channel_fraction", "method.lm_loss", "method.parameters", "method.perplexity", "relative_perplexity_percent" ], "default_seeds": [ 42 ], "budget": "steps=30", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "foundation-models", "key": "windowed-mtp", "title": "Windowed-MTP: Removing the Full-Context Draft-KV Tax at Million-Token Context", "paper_url": "https://arxiv.org/abs/2607.21535", "detail_path": "reproductions/2607.21535-windowed-mtp/README.md", "topic": [ "llm-serving", "speculative-decoding", "kv-cache", "long-context" ], "first_author": null, "first_author_affiliation": "NVIDIA", "published": "2026-07-23", "code": "https://github.com/avalliappan-nvidia/windowed-mtp-b200", "adapter": { "adapter_key": "windowed-mtp", "arxiv_id": "2607.21535", "title": "Windowed-MTP: Removing the Full-Context Draft-KV Tax at Million-Token Context", "paper_url": "https://arxiv.org/abs/2607.21535", "track": "llm", "organization": "NVIDIA", "published": "2026-07-23", "code_url": "https://github.com/avalliappan-nvidia/windowed-mtp-b200", "topics": [ "llm-serving", "speculative-decoding", "kv-cache", "long-context" ], "local_code_dir": "src/auto_research/reproductions/windowed_mtp", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "WikiText-2" ], "baseline": "读取完整历史的同权重 MTP draft", "metrics": [ "context_16384.latency_reduction_percent", "target_perplexity" ], "default_seeds": [ 42 ], "budget": "steps=240", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "multimodal-models", "key": "stacktok", "title": "StackTok: Accelerating VLMs Inference with Budget-Adaptive Visual Token Selection", "paper_url": "https://arxiv.org/abs/2609.16841", "detail_path": "multimodal-models/2609.16841-stacktok/README.md", "topic": [ "视觉 Token 选择", "推理效率" ], "first_author": "Zhenbin Wang", "first_author_affiliation": "Sichuan University", "published": "2026-09-15", "code": "https://github.com/wongzbb/StackToK", "adapter": "stacktok", "priority": "P1" }, { "domain": "multimodal-models", "key": "videomm", "title": "VideoMM: Adaptive Macro-Micro Inference for Efficient Video MLLMs", "paper_url": "https://arxiv.org/abs/2609.16722", "detail_path": "multimodal-models/2609.16722-videomm/README.md", "topic": [ "视频理解", "Token 压缩" ], "first_author": "Haoyu Guo", "first_author_affiliation": "University of Science and Technology of China", "published": "2026-09-15", "code": "https://github.com/adfh917k/VideoMM", "adapter": "videomm", "priority": "P0", "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/videomm-a100-20260916.json" }, { "domain": "post-training", "key": "absolute-zero", "title": "Absolute Zero", "paper_url": "https://arxiv.org/abs/2505.03335", "detail_path": "post-training/2505.03335-absolute-zero/README.md", "topic": [ "零数据自博弈" ], "first_author": "Andrew Zhao", "first_author_affiliation": "Tsinghua University", "published": "2025-05-06", "code": "https://github.com/LeapLabTHU/Absolute-Zero-Reasoner", "adapter": "absolute-zero" }, { "domain": "post-training", "key": "adaptive-opd-gate", "title": "A Unified Per-Token Gating Family for On-Policy Distillation: FKL/RKL Mixing with Multi-Channel and Bias Coefficients", "paper_url": "https://arxiv.org/abs/2609.11768", "detail_path": "post-training/2609.11768-adaptive-opd-gate/README.md", "topic": [ "On-policy distillation", "逐 token 自适应门控" ], "first_author": "Suwan Wu", "first_author_affiliation": "Xiaohongshu", "published": "2026-09-10", "code": null, "adapter": "adaptive-opd-gate", "priority": "P1" }, { "domain": "post-training", "key": "adrs", "title": "ADRS", "paper_url": "https://arxiv.org/abs/2608.03223", "detail_path": "post-training/2608.03223-adrs/README.md", "topic": [ "回报相关奖励塑形" ], "first_author": "Ranxu Zhang", "first_author_affiliation": "University of Science and Technology of China", "published": "2026-08-04", "code": "https://github.com/gitrxh/ADRS-arxiv", "adapter": "adrs" }, { "domain": "post-training", "key": "armor", "title": "ARMOR", "paper_url": "https://arxiv.org/abs/2607.10481", "detail_path": "post-training/2607.10481-armor/README.md", "topic": [ "Reference anchor" ], "first_author": "Kexin Huang", "first_author_affiliation": "University of Science and Technology of China", "published": "2026-07-11", "code": "未发现官方代码", "adapter": "armor" }, { "domain": "post-training", "key": "beta-opsd", "title": "β-OPSD", "paper_url": "https://arxiv.org/abs/2607.28582", "detail_path": "post-training/2607.28582-beta-opsd/README.md", "topic": [ "On-policy self-distillation" ], "first_author": "Jiawei Xu", "first_author_affiliation": "University of Maryland, College Park", "published": "2026-07-30", "code": "未发现官方代码", "adapter": "beta-opsd" }, { "domain": "post-training", "key": "c2-dpo", "title": "Context Blindness in DPO: Mitigating Object Hallucination in MLLMs via Context-Calibrated Preference Optimization", "paper_url": "https://arxiv.org/abs/2608.12158", "detail_path": "post-training/2608.12158-c2-dpo/README.md", "topic": [ "多模态上下文偏好校准" ], "first_author": "Byungoh Ko", "first_author_affiliation": "Korea University", "published": "2026-08-12", "code": "https://github.com/mlvlab/C2-DPO", "adapter": "c2-dpo" }, { "domain": "post-training", "key": "calibration-aware-rl", "title": "Balancing Classification and Calibration Performance in Decision-Making LLMs via Calibration Aware Reinforcement Learning", "paper_url": "https://arxiv.org/abs/2601.13284", "detail_path": "post-training/2601.13284-calibration-aware-rl/README.md", "topic": [ "校准强化学习", "决策 token" ], "first_author": "Duygu Nur Yaldiz", "first_author_affiliation": "University of Southern California", "published": "2026-01-19", "code": null, "adapter": "system-one:rival-hybrid", "priority": "P1" }, { "domain": "post-training", "key": "causal-opd", "title": "CausalOPD: First-Wrong-Step Supervision for Distilling Causal Chain Reasoning", "paper_url": "https://arxiv.org/abs/2608.03673", "detail_path": "post-training/2608.03673-causal-opd/README.md", "topic": [ "OPD", "过程监督" ], "first_author": "Jian Zhang", "first_author_affiliation": "Authors did not disclose affiliation", "published": "2026-08-04", "code": null, "adapter": "causal-opd" }, { "domain": "post-training", "key": "chord", "title": "CHORD", "paper_url": "https://arxiv.org/abs/2508.11408", "detail_path": "post-training/2508.11408-chord/README.md", "topic": [ "SFT-RL 动态混合" ], "first_author": "Wenhao Zhang", "first_author_affiliation": "Alibaba Group", "published": "2025-08-15", "code": "https://github.com/modelscope/Trinity-RFT/tree/main/examples/mix_chord", "adapter": "chord" }, { "domain": "post-training", "key": "cispo", "title": "CISPO / MiniMax-M1", "paper_url": "https://arxiv.org/abs/2506.13585", "detail_path": "post-training/2506.13585-cispo/README.md", "topic": [ "长上下文 RL" ], "first_author": "MiniMax", "first_author_affiliation": "MiniMax", "published": "2025-06-16", "code": "https://github.com/MiniMax-AI/MiniMax-M1", "adapter": "cispo" }, { "domain": "post-training", "key": "clue-opsd", "title": "Where to Look Matters: On-Policy Self-Distillation for Long-Video Understanding", "paper_url": "https://arxiv.org/abs/2608.25356", "detail_path": "post-training/2608.25356-clue-opsd/README.md", "topic": [ "长视频特权视图 OPD" ], "first_author": "Kaishen Wang", "first_author_affiliation": "University of Maryland, College Park", "published": "2026-08-26", "code": null, "adapter": "clue-opsd" }, { "domain": "post-training", "key": "coba-rl", "title": "CoBA-RL", "paper_url": "https://arxiv.org/abs/2606.22317", "detail_path": "post-training/2606.22317-coba-rl/README.md", "topic": [ "能力边界课程" ], "first_author": "Pengxiang Cai", "first_author_affiliation": "Zhejiang University", "published": "2026-06-21", "code": "未发现官方代码", "adapter": "coba-rl" }, { "domain": "post-training", "key": "compass-opd", "title": "CompassOPD: Cross-Family On-Policy Distillation via Within-Family Likelihood Shifts", "paper_url": "https://arxiv.org/abs/2609.10154", "detail_path": "post-training/2609.10154-compass-opd/README.md", "topic": [ "跨模型族 OPD" ], "first_author": "Naibin Gu", "first_author_affiliation": "Institute of Information Engineering, Chinese Academy of Sciences", "published": "2026-09-09", "code": null, "adapter": "compass-opd" }, { "domain": "post-training", "key": "compo", "title": "A Zeroth-Order Paradigm for LLM Preference Alignment", "paper_url": "https://arxiv.org/abs/2609.19144", "detail_path": "post-training/2609.19144-compo/README.md", "topic": [ "偏好优化", "零阶优化" ], "first_author": "Peter Chen", "first_author_affiliation": "University of California, Berkeley", "published": "2026-09-16", "code": null, "adapter": "compo", "priority": "P1" }, { "domain": "post-training", "key": "conspo", "title": "ConSPO", "paper_url": "https://arxiv.org/abs/2605.12969", "detail_path": "post-training/2605.12969-conspo/README.md", "topic": [ "对比序列 RL" ], "first_author": "Feng Zhang", "first_author_affiliation": "Beijing Institute of Technology", "published": "2026-05-13", "code": "未发现官方代码", "adapter": "conspo" }, { "domain": "post-training", "key": "constitutional-ai", "title": "Constitutional AI", "paper_url": "https://arxiv.org/abs/2212.08073", "detail_path": "post-training/2212.08073-constitutional-ai/README.md", "topic": [ "AI 反馈安全对齐" ], "first_author": "Yuntao Bai", "first_author_affiliation": "Anthropic", "published": "2022-12-15", "code": "https://github.com/anthropics/ConstitutionalHarmlessnessPaper", "adapter": "constitutional-ai" }, { "domain": "post-training", "key": "cort", "title": "CoRT", "paper_url": "https://arxiv.org/abs/2607.25659", "detail_path": "post-training/2607.25659-cort/README.md", "topic": [ "Token-level credit assignment" ], "first_author": "Bo-Wen Zhang", "first_author_affiliation": "ByteDance internship", "published": "2026-07-28", "code": "未发现官方代码", "adapter": "cort" }, { "domain": "post-training", "key": "crpo", "title": "Contrastive Reinforced Policy Optimization via Privileged Self-Distillation", "paper_url": "https://arxiv.org/abs/2607.28026", "detail_path": "post-training/2607.28026-crpo/README.md", "topic": [ "OPD", "Agentic RL" ], "first_author": "Xingjian Wu", "first_author_affiliation": "Authors did not disclose affiliation", "published": "2026-07-30", "code": null, "adapter": "crpo" }, { "domain": "post-training", "key": "dapo", "title": "DAPO", "paper_url": "https://arxiv.org/abs/2503.14476", "detail_path": "post-training/2503.14476-dapo/README.md", "topic": [ "长推理 RL" ], "first_author": "Qiying Yu", "first_author_affiliation": "ByteDance Seed", "published": "2025-03-18", "code": "https://github.com/BytedTsinghua-SIA/DAPO", "adapter": "dapo" }, { "domain": "post-training", "key": "dash", "title": "DASH", "paper_url": "https://arxiv.org/abs/2608.06243", "detail_path": "post-training/2608.06243-dash/README.md", "topic": [ "自适应自蒸馏" ], "first_author": "Zhiyan Hou", "first_author_affiliation": "Nanjing University", "published": "2026-08-06", "code": "https://github.com/DBtxy/DASH-OPSD", "adapter": "dash" }, { "domain": "post-training", "key": "df-opd", "title": "Data-free On-policy Distillation", "paper_url": "https://arxiv.org/abs/2609.14193", "detail_path": "post-training/2609.14193-df-opd/README.md", "topic": [ "On-policy distillation", "数据合成" ], "first_author": "Gengsheng Li", "first_author_affiliation": "Institute of Automation, Chinese Academy of Sciences / Tencent", "published": "2026-09-12", "code": null, "adapter": "df-opd", "priority": "P0" }, { "domain": "post-training", "key": "distilled-rl", "title": "Distilled RL", "paper_url": "https://arxiv.org/abs/2607.17247", "detail_path": "post-training/2607.17247-distilled-rl/README.md", "topic": [ "教师奖励重权重" ], "first_author": "Chen Wang", "first_author_affiliation": "Nankai University", "published": "2026-07-19", "code": "https://github.com/597358816/Distilled-RL", "adapter": "distilled-rl" }, { "domain": "post-training", "key": "dpo", "title": "DPO", "paper_url": "https://arxiv.org/abs/2305.18290", "detail_path": "post-training/2305.18290-dpo/README.md", "topic": [ "直接偏好优化" ], "first_author": "Rafael Rafailov", "first_author_affiliation": "Stanford University", "published": "2023-05-29", "code": "https://github.com/eric-mitchell/direct-preference-optimization", "adapter": "dpo" }, { "domain": "post-training", "key": "dr-grpo", "title": "Dr. GRPO", "paper_url": "https://arxiv.org/abs/2503.20783", "detail_path": "post-training/2503.20783-dr-grpo/README.md", "topic": [ "GRPO 聚合偏置" ], "first_author": "Zichen Liu", "first_author_affiliation": "SAIL 研究团队", "published": "2025-03-26", "code": "https://github.com/sail-sg/understand-r1-zero", "adapter": "dr-grpo" }, { "domain": "post-training", "key": "dynamic-rubric", "title": "Co-Evolving LLM Evaluators and Policies via DynamicRubric", "paper_url": "https://arxiv.org/abs/2607.20083", "detail_path": "reproductions/2607.20083-dynamic-rubric/README.md", "topic": [ "post-training", "reward-model", "evaluator", "co-evolution" ], "first_author": "Beining Wang", "first_author_affiliation": "WeChat / Tencent", "published": "2026-07-22", "code": null, "adapter": { "adapter_key": "dynamic-rubric", "arxiv_id": "2607.20083", "title": "Co-Evolving LLM Evaluators and Policies via DynamicRubric", "paper_url": "https://arxiv.org/abs/2607.20083", "track": "llm", "organization": "WeChat / Tencent and Tsinghua University", "published": "2026-07-22", "code_url": null, "topics": [ "post-training", "reward-model", "evaluator", "co-evolution" ], "local_code_dir": "src/auto_research/reproductions/dynamic_rubric", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "Stanford Alpaca" ], "baseline": "固定 prompt-only rubric policy，实验组为 DynamicRubric evaluator-policy 共进化", "metrics": [ "baseline.preference_accuracy", "method.preference_accuracy", "relative.margin_delta", "relative.preference_accuracy_percent" ], "default_seeds": [ 42 ], "budget": "steps=1600", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": "The paper reports statistically significant online A/B gains and a full-traffic WeChat Search deployment, but withholds exact lifts.", "evolve_operators": [] } }, { "domain": "post-training", "key": "erils", "title": "Beyond On-Policy Exploration: Integrating External Policy Rollouts for Reinforcement Learning in Diffusion Language Models", "paper_url": "https://arxiv.org/abs/2608.01717", "detail_path": "post-training/2608.01717-erils/README.md", "topic": [ "外部 rollout", "Diffusion LM RL" ], "first_author": "Wonseok Lee", "first_author_affiliation": "Seoul National University", "published": "2026-08-03", "code": null, "adapter": "erils" }, { "domain": "post-training", "key": "erpo", "title": "Beyond the Stability-Exploration Dilemma: Environmental Regularization for LLM Policy Optimization", "paper_url": "https://arxiv.org/abs/2608.23311", "detail_path": "post-training/2608.23311-erpo/README.md", "topic": [ "输入侧 Query-KL" ], "first_author": "Xianlei Zhou", "first_author_affiliation": "AMAP, Alibaba Group", "published": "2026-08-24", "code": "https://github.com/alibaba/ERPO", "adapter": "erpo" }, { "domain": "post-training", "key": "flux-opd", "title": "Flux-OPD", "paper_url": "https://arxiv.org/abs/2607.28022", "detail_path": "post-training/2607.28022-flux-opd/README.md", "topic": [ "Context distillation" ], "first_author": "Yuran Wang", "first_author_affiliation": "Peking University", "published": "2026-07-30", "code": "未发现官方代码", "adapter": "flux-opd" }, { "domain": "post-training", "key": "gamma-opd", "title": "Beyond Token-Local Imitation: Reward-Compatible Temporal Credit Assignment for On-Policy Distillation", "paper_url": "https://arxiv.org/abs/2609.16937", "detail_path": "post-training/2609.16937-gamma-opd/README.md", "topic": [ "On-policy distillation", "时间信用分配" ], "first_author": "Shiqi Liu", "first_author_affiliation": "Tsinghua University / DiDi Voyager Labs", "published": "2026-09-15", "code": null, "adapter": "gamma-opd", "priority": "P0" }, { "domain": "post-training", "key": "gcpo", "title": "GCPO: Diagnosing and Constraining Subspace Geometry in Rollout RL for LLMs", "paper_url": "https://arxiv.org/abs/2608.11674", "detail_path": "post-training/2608.11674-gcpo/README.md", "topic": [ "几何约束 RL" ], "first_author": "Kai Yang", "first_author_affiliation": "Shanghai AI Laboratory", "published": "2026-08-12", "code": "https://github.com/Icarus1411/GCPO", "adapter": "gcpo" }, { "domain": "post-training", "key": "gkd", "title": "GKD", "paper_url": "https://arxiv.org/abs/2306.13649", "detail_path": "post-training/2306.13649-gkd/README.md", "topic": [ "经典 On-policy distillation" ], "first_author": "Rishabh Agarwal", "first_author_affiliation": "Google DeepMind", "published": "2023-06-23", "code": "https://github.com/huggingface/trl/blob/main/docs/source/distillation_trainer.md", "adapter": "gkd" }, { "domain": "post-training", "key": "gppo", "title": "GPPO", "paper_url": "https://arxiv.org/abs/2508.07629", "detail_path": "post-training/2508.07629-gppo/README.md", "topic": [ "梯度保留 clip" ], "first_author": "Zhenpeng Su", "first_author_affiliation": "Alibaba Group", "published": "2025-08-11", "code": "未发现官方代码", "adapter": "gppo" }, { "domain": "post-training", "key": "gprl", "title": "GPRL", "paper_url": "https://arxiv.org/abs/2605.18721", "detail_path": "post-training/2605.18721-gprl/README.md", "topic": [ "多目标 RL" ], "first_author": "Muhammad Umer", "first_author_affiliation": "Stanford University", "published": "2026-05-18", "code": "未发现官方代码", "adapter": "gprl" }, { "domain": "post-training", "key": "grin", "title": "From Memorization to Absorption: Mixed-Policy RL for Continual Knowledge Injection", "paper_url": "https://arxiv.org/abs/2608.25243", "detail_path": "post-training/2608.25243-grin/README.md", "topic": [ "混合策略知识注入 RL" ], "first_author": "Zhibo Hou", "first_author_affiliation": "University of California, Merced", "published": "2026-08-26", "code": null, "adapter": "grin" }, { "domain": "post-training", "key": "grip", "title": "GRIP: Granular Reward-Guided Parameter Interpolation for Efficient Reasoning", "paper_url": "https://arxiv.org/abs/2608.25583", "detail_path": "post-training/2608.25583-grip/README.md", "topic": [ "奖励引导参数插值" ], "first_author": "Lam So", "first_author_affiliation": "Peking University", "published": "2026-08-26", "code": null, "adapter": "grip" }, { "domain": "post-training", "key": "growmtp", "title": "GrowMTP: Can RL Grow Its Own Draft Head?", "paper_url": "https://arxiv.org/abs/2609.16648", "detail_path": "post-training/2609.16648-growmtp/README.md", "topic": [ "RL 训练加速", "推测解码" ], "first_author": "Minghua He", "first_author_affiliation": "WeChat AI, Tencent / Peking University", "published": "2026-09-15", "code": null, "adapter": "growmtp", "priority": "P1" }, { "domain": "post-training", "key": "grpo", "title": "DeepSeekMath / GRPO", "paper_url": "https://arxiv.org/abs/2402.03300", "detail_path": "post-training/2402.03300-grpo/README.md", "topic": [ "在线推理 RL" ], "first_author": "Zhihong Shao", "first_author_affiliation": "DeepSeek-AI", "published": "2024-02-05", "code": "https://github.com/deepseek-ai/DeepSeek-Math", "adapter": "grpo" }, { "domain": "post-training", "key": "gspo", "title": "GSPO", "paper_url": "https://arxiv.org/abs/2507.18071", "detail_path": "post-training/2507.18071-gspo/README.md", "topic": [ "稳定序列 RL" ], "first_author": "Chujie Zheng", "first_author_affiliation": "Alibaba Qwen Team", "published": "2025-07-24", "code": "https://alibaba.github.io/ROLL/docs/User%20Guides/Algorithms/GSPO/", "adapter": "gspo" }, { "domain": "post-training", "key": "icepop", "title": "IcePop", "paper_url": "https://arxiv.org/abs/2510.18855", "detail_path": "post-training/2510.18855-icepop/README.md", "topic": [ "MoE 训推失配" ], "first_author": "Ling Team", "first_author_affiliation": "Ant Group", "published": "2025-10-21", "code": "未发现官方代码", "adapter": "icepop" }, { "domain": "post-training", "key": "intuitor", "title": "INTUITOR", "paper_url": "https://arxiv.org/abs/2505.19590", "detail_path": "post-training/2505.19590-intuitor/README.md", "topic": [ "自置信奖励" ], "first_author": "Xuandong Zhao", "first_author_affiliation": "University of California, Berkeley", "published": "2025-05-26", "code": "https://github.com/sunblaze-ucb/Intuitor", "adapter": "intuitor" }, { "domain": "post-training", "key": "ipo", "title": "IPO", "paper_url": "https://arxiv.org/abs/2310.12036", "detail_path": "post-training/2310.12036-ipo/README.md", "topic": [ "偏好正则" ], "first_author": "Mohammad Gheshlaghi Azar", "first_author_affiliation": "Google DeepMind", "published": "2023-10-18", "code": "未发现官方代码", "adapter": "ipo" }, { "domain": "post-training", "key": "iso-rlvr", "title": "ISO: An RLVR-Native Optimization Stack", "paper_url": "https://arxiv.org/abs/2607.19331", "detail_path": "post-training/2607.19331-iso-rlvr/README.md", "topic": [ "RLVR", "优化器" ], "first_author": "Hanqing Zhu", "first_author_affiliation": "The University of Texas at Austin", "published": "2026-07-21", "code": "https://github.com/zhuhanqing/ISO", "adapter": "iso-rlvr" }, { "domain": "post-training", "key": "kpop", "title": "KPop", "paper_url": "https://arxiv.org/abs/2606.15079", "detail_path": "post-training/2606.15079-kpop/README.md", "topic": [ "异步训推失配" ], "first_author": "Ang Li", "first_author_affiliation": "Ling / Ring Team", "published": "2026-06-13", "code": "未发现官方代码", "adapter": "kpop" }, { "domain": "post-training", "key": "kto", "title": "KTO", "paper_url": "https://arxiv.org/abs/2402.01306", "detail_path": "post-training/2402.01306-kto/README.md", "topic": [ "二元反馈对齐" ], "first_author": "Kawin Ethayarajh", "first_author_affiliation": "Contextual AI", "published": "2024-02-02", "code": "https://github.com/ContextualAI/HALOs", "adapter": "kto" }, { "domain": "post-training", "key": "lightning-opd", "title": "Lightning OPD", "paper_url": "https://arxiv.org/abs/2604.13010", "detail_path": "post-training/2604.13010-lightning-opd/README.md", "topic": [ "On-policy distillation" ], "first_author": "Yecheng Wu", "first_author_affiliation": "MIT HAN Lab", "published": "2026-04-14", "code": "https://github.com/jet-ai-projects/Lightning-OPD", "adapter": "lightning-opd" }, { "domain": "post-training", "key": "locus", "title": "LOCUS: Task-Aware Low-Rank Post-Training for Token-Efficient Language Generation", "paper_url": "https://arxiv.org/abs/2609.11739", "detail_path": "post-training/2609.11739-locus/README.md", "topic": [ "低秩后训练", "Token 效率" ], "first_author": "Dongfang Zhao", "first_author_affiliation": "University of Washington", "published": "2026-09-10", "code": null, "adapter": "locus", "priority": "P1" }, { "domain": "post-training", "key": "luffy", "title": "LUFFY", "paper_url": "https://arxiv.org/abs/2504.14945", "detail_path": "post-training/2504.14945-luffy/README.md", "topic": [ "离策略推理 RL" ], "first_author": "Jianhao Yan", "first_author_affiliation": "University of Washington", "published": "2025-04-21", "code": "https://github.com/Simplified-Reasoning/LUFFY", "adapter": "luffy" }, { "domain": "post-training", "key": "luspo", "title": "LUSPO", "paper_url": "https://arxiv.org/abs/2602.05261", "detail_path": "post-training/2602.05261-luspo/README.md", "topic": [ "长度无偏 RL" ], "first_author": "Fanfan Liu", "first_author_affiliation": "Meituan", "published": "2026-02-05", "code": "未发现官方代码", "adapter": "luspo" }, { "domain": "post-training", "key": "math-shepherd", "title": "Math-Shepherd", "paper_url": "https://arxiv.org/abs/2312.08935", "detail_path": "post-training/2312.08935-math-shepherd/README.md", "topic": [ "自动过程奖励" ], "first_author": "Peiyi Wang", "first_author_affiliation": "Peking University", "published": "2023-12-14", "code": "未发现官方代码", "adapter": "math-shepherd" }, { "domain": "post-training", "key": "minillm", "title": "MiniLLM", "paper_url": "https://arxiv.org/abs/2306.08543", "detail_path": "post-training/2306.08543-minillm/README.md", "topic": [ "Reverse-KL distillation" ], "first_author": "Yuxian Gu", "first_author_affiliation": "Tsinghua University", "published": "2023-06-14", "code": "https://github.com/microsoft/LMOps/tree/main/minillm", "adapter": "minillm" }, { "domain": "post-training", "key": "minirl", "title": "Stabilizing RL with LLMs", "paper_url": "https://arxiv.org/abs/2512.01374", "detail_path": "post-training/2512.01374-minirl/README.md", "topic": [ "稳定 MoE RL" ], "first_author": "Chujie Zheng", "first_author_affiliation": "Alibaba Qwen Team", "published": "2025-12-01", "code": "未发现官方代码", "adapter": "minirl" }, { "domain": "post-training", "key": "missing-old-logits", "title": "Missing Old Logits", "paper_url": "https://arxiv.org/abs/2605.12070", "detail_path": "post-training/2605.12070-missing-old-logits/README.md", "topic": [ "异步 off-policy" ], "first_author": "Zhong Guan", "first_author_affiliation": "Tianjin University", "published": "2026-05-12", "code": "未发现官方代码", "adapter": "missing-old-logits" }, { "domain": "post-training", "key": "mopd", "title": "MOPD", "paper_url": "https://arxiv.org/abs/2606.30406", "detail_path": "post-training/2606.30406-mopd/README.md", "topic": [ "多教师能力整合" ], "first_author": "Wenhan Ma", "first_author_affiliation": "Xiaomi", "published": "2026-06-29", "code": "未发现官方代码", "adapter": "mopd" }, { "domain": "post-training", "key": "normalized-dpo", "title": "Disentangling Optimization Scale from Preference Scale in DPO", "paper_url": "https://arxiv.org/abs/2608.27032", "detail_path": "post-training/2608.27032-normalized-dpo/README.md", "topic": [ "DPO 偏好/优化尺度解耦" ], "first_author": "Ivan Kruzhilov", "first_author_affiliation": "Affiliation not listed in the paper", "published": "2026-08-27", "code": "https://github.com/ivankru/bayesian_dpo", "adapter": "normalized-dpo" }, { "domain": "post-training", "key": "nsd", "title": "Negative Self-Distillation: Learning to Reason by Avoiding Flaws", "paper_url": "https://arxiv.org/abs/2609.11699", "detail_path": "post-training/2609.11699-nsd/README.md", "topic": [ "负向自蒸馏", "推理后训练" ], "first_author": "Rongcan Pei", "first_author_affiliation": "University of Virginia", "published": "2026-09-10", "code": "https://github.com/Prongcan/NSD", "adapter": "nsd", "priority": "P0" }, { "domain": "post-training", "key": "off-context-grpo", "title": "Off-Context GRPO: Learning to Reason on Hard Problems using Privileged Information", "paper_url": "https://arxiv.org/abs/2607.19313", "detail_path": "reproductions/2607.19313-off-context-grpo/README.md", "topic": [ "rlvr", "grpo", "reasoning", "privileged-information" ], "first_author": "Priyank Agrawal", "first_author_affiliation": "Meta AI", "published": "2026-07-21", "code": "https://github.com/AgPriyank/OC-GRPO", "adapter": { "adapter_key": "off-context-grpo", "arxiv_id": "2607.19313", "title": "Off-Context GRPO: Learning to Reason on Hard Problems using Privileged Information", "paper_url": "https://arxiv.org/abs/2607.19313", "track": "llm", "organization": "Meta AI / Columbia University", "published": "2026-07-21", "code_url": "https://github.com/AgPriyank/OC-GRPO", "topics": [ "rlvr", "grpo", "reasoning", "privileged-information" ], "local_code_dir": "src/auto_research/reproductions/off_context_grpo", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "GSM8K official" ], "baseline": "同 2,400 step、group size 8 的 vanilla GRPO，实验组为 Off-Context GRPO", "metrics": [ "baseline.pass_at_1", "baseline.pass_at_8", "method.pass_at_1", "method.pass_at_8", "relative.pass_at_1_percent" ], "default_seeds": [ 42 ], "budget": "steps=2400", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "post-training", "key": "online-icepop", "title": "Online IcePop", "paper_url": "web-2025-online-icepop/README.md", "detail_path": "post-training/web-2025-online-icepop/README.md", "topic": [ "纯在线训推校正" ], "first_author": "Jian Hu", "first_author_affiliation": "Ant Group", "published": "2025-12-16", "code": "未发现官方代码", "adapter": "online-icepop" }, { "domain": "post-training", "key": "opcd", "title": "OPCD", "paper_url": "https://arxiv.org/abs/2602.12275", "detail_path": "post-training/2602.12275-opcd/README.md", "topic": [ "Context distillation" ], "first_author": "Tianzhu Ye", "first_author_affiliation": "Microsoft Research", "published": "2026-02-12", "code": "https://github.com/microsoft/LMOps/tree/main/opcd", "adapter": "opcd" }, { "domain": "post-training", "key": "opd-aha", "title": "OPD-Aha: From Linguistic Momentum to Visual Reflection in Multimodal On-Policy Distillation", "paper_url": "https://arxiv.org/abs/2609.16459", "detail_path": "post-training/2609.16459-opd-aha/README.md", "topic": [ "多模态后训练", "On-policy distillation" ], "first_author": "Chenhao Qiu", "first_author_affiliation": "Arizona State University / University of Virginia / Stevens Institute of Technology", "published": "2026-09-15", "code": "https://github.com/Echochef/OPD-Aha", "adapter": "opd-aha", "priority": "P0" }, { "domain": "post-training", "key": "opd-lm", "title": "OPDLM", "paper_url": "https://arxiv.org/abs/2606.06712", "detail_path": "post-training/2606.06712-opd-lm/README.md", "topic": [ "AR-to-Diffusion 蒸馏" ], "first_author": "Xingyu Su", "first_author_affiliation": "Texas A&M University", "published": "2026-06-04", "code": "未发现官方代码", "adapter": "opd-lm" }, { "domain": "post-training", "key": "opd-search-plus", "title": "OPDSearch+: Search-Enhanced On-Policy Distillation with Reinforcement Learning", "paper_url": "https://arxiv.org/abs/2608.24310", "detail_path": "post-training/2608.24310-opd-search-plus/README.md", "topic": [ "搜索增强 OPD + RL" ], "first_author": "Qinglin Ye", "first_author_affiliation": "University of Chinese Academy of Sciences", "published": "2026-08-25", "code": null, "adapter": "opd-search-plus" }, { "domain": "post-training", "key": "opd2", "title": "On-Policy Delta Distillation for Multilingual Math Reasoning", "paper_url": "https://arxiv.org/abs/2608.05802", "detail_path": "post-training/2608.05802-opd2/README.md", "topic": [ "OPD", "多语言推理" ], "first_author": "Byeongho Heo", "first_author_affiliation": "NAVER AI Lab", "published": "2026-08-06", "code": "https://github.com/naver-ai/opd2", "adapter": "opd2" }, { "domain": "post-training", "key": "opdvr", "title": "OPDVR: On-Policy Distillation with Verifiable Rewards", "paper_url": "https://arxiv.org/abs/2608.24696", "detail_path": "post-training/2608.24696-opdvr/README.md", "topic": [ "可验证奖励 OPD" ], "first_author": "Wenze Lin", "first_author_affiliation": "LeapLab, Tsinghua University", "published": "2026-08-25", "code": "https://github.com/LeapLabTHU/OPDVR", "adapter": "opdvr" }, { "domain": "post-training", "key": "oprd", "title": "Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation", "paper_url": "https://arxiv.org/abs/2609.08798", "detail_path": "post-training/2609.08798-oprd/README.md", "topic": [ "弱到强反向蒸馏" ], "first_author": "Youngrok Park", "first_author_affiliation": "KAIST AI", "published": "2026-09-08", "code": null, "adapter": "oprd" }, { "domain": "post-training", "key": "opsd", "title": "OPSD", "paper_url": "https://arxiv.org/abs/2601.18734", "detail_path": "post-training/2601.18734-opsd/README.md", "topic": [ "On-policy self-distillation" ], "first_author": "Siyan Zhao", "first_author_affiliation": "University of California, Los Angeles", "published": "2026-01-26", "code": "https://github.com/siyan-zhao/OPSD", "adapter": "opsd" }, { "domain": "post-training", "key": "orpo", "title": "ORPO", "paper_url": "https://arxiv.org/abs/2403.07691", "detail_path": "post-training/2403.07691-orpo/README.md", "topic": [ "单阶段偏好" ], "first_author": "Jiwoo Hong", "first_author_affiliation": "KAIST", "published": "2024-03-12", "code": "https://github.com/xfactlab/orpo", "adapter": "orpo" }, { "domain": "post-training", "key": "pcsd", "title": "PCSD", "paper_url": "https://arxiv.org/abs/2608.01837", "detail_path": "post-training/2608.01837-pcsd/README.md", "topic": [ "持续一致性蒸馏" ], "first_author": "Chunji Lv", "first_author_affiliation": "论文未列机构", "published": "2026-08-03", "code": "未发现官方代码", "adapter": "pcsd" }, { "domain": "post-training", "key": "ppo-rlhf", "title": "InstructGPT / PPO-RLHF", "paper_url": "https://arxiv.org/abs/2203.02155", "detail_path": "post-training/2203.02155-ppo-rlhf/README.md", "topic": [ "经典 RLHF" ], "first_author": "Long Ouyang", "first_author_affiliation": "OpenAI", "published": "2022-03-04", "code": "https://github.com/openai/following-instructions-human-feedback", "adapter": "ppo-rlhf" }, { "domain": "post-training", "key": "probe-erpo", "title": "Entropy-Regularized Rank-Masked Policy Optimization for Test-Time Reinforcement Learning in Code Generation", "paper_url": "https://arxiv.org/abs/2609.09135", "detail_path": "post-training/2609.09135-probe-erpo/README.md", "topic": [ "代码测试时 RL" ], "first_author": "Jiacheng Xu", "first_author_affiliation": "Nanyang Technological University", "published": "2026-09-08", "code": null, "adapter": "probe-erpo" }, { "domain": "post-training", "key": "process-supervision", "title": "Let's Verify Step by Step", "paper_url": "https://arxiv.org/abs/2305.20050", "detail_path": "post-training/2305.20050-process-supervision/README.md", "topic": [ "过程监督" ], "first_author": "Hunter Lightman", "first_author_affiliation": "OpenAI", "published": "2023-05-31", "code": "https://github.com/openai/prm800k", "adapter": "process-supervision" }, { "domain": "post-training", "key": "pto", "title": "Preference Tree Optimization: Enhancing Goal-Oriented Dialogue with Look-Ahead Simulations", "paper_url": "https://arxiv.org/abs/2608.12062", "detail_path": "post-training/2608.12062-pto/README.md", "topic": [ "前瞻偏好树" ], "first_author": "Lior Baruch", "first_author_affiliation": "Reichman University", "published": "2026-08-12", "code": null, "adapter": "pto" }, { "domain": "post-training", "key": "r2-opd", "title": "Beyond Imitation: Filtering On-Policy Distillation by Reasoning Progress", "paper_url": "https://arxiv.org/abs/2608.19408", "detail_path": "post-training/2608.19408-r2-opd/README.md", "topic": [ "OPD", "过程蒸馏" ], "first_author": "Chen Yang", "first_author_affiliation": "Authors did not disclose affiliation", "published": "2026-08-19", "code": null, "adapter": "r2-opd" }, { "domain": "post-training", "key": "raft", "title": "RAFT", "paper_url": "https://arxiv.org/abs/2304.06767", "detail_path": "post-training/2304.06767-raft/README.md", "topic": [ "Reward 选优微调" ], "first_author": "Hanze Dong", "first_author_affiliation": "HKUST", "published": "2023-04-13", "code": "https://github.com/OptimalScale/LMFlow", "adapter": "raft" }, { "domain": "post-training", "key": "rail", "title": "Optimizing What Policies Learn From: Recoverability-Aware Rollout Intervention Learning", "paper_url": "https://arxiv.org/abs/2608.05080", "detail_path": "post-training/2608.05080-rail/README.md", "topic": [ "rollout 预算分配" ], "first_author": "Zheyuan Zhang", "first_author_affiliation": "University of Notre Dame / Amazon", "published": "2026-08-05", "code": null, "adapter": "rail" }, { "domain": "post-training", "key": "reco-grpo", "title": "ReCo", "paper_url": "https://arxiv.org/abs/2607.26862", "detail_path": "post-training/2607.26862-reco/README.md", "topic": [ "分布保持 RL" ], "first_author": "Junoh Park", "first_author_affiliation": "Seoul National University", "published": "2026-07-29", "code": "未发现官方代码", "adapter": "reco-grpo" }, { "domain": "post-training", "key": "reinforce-plus", "title": "REINFORCE++", "paper_url": "https://arxiv.org/abs/2501.03262", "detail_path": "post-training/2501.03262-reinforce-plus/README.md", "topic": [ "全局优势估计" ], "first_author": "Jian Hu", "first_author_affiliation": "Independent researchers", "published": "2025-01-04", "code": "未发现官方代码", "adapter": "reinforce-plus" }, { "domain": "post-training", "key": "relay-opd", "title": "Relay-OPD", "paper_url": "https://arxiv.org/abs/2607.26057", "detail_path": "post-training/2607.26057-relay-opd/README.md", "topic": [ "On-policy distillation" ], "first_author": "Haolei Xu", "first_author_affiliation": "Zhejiang University", "published": "2026-07-28", "code": "https://github.com/ZJU-REAL/Relay-OPD", "adapter": "relay-opd" }, { "domain": "post-training", "key": "remax", "title": "ReMax", "paper_url": "https://arxiv.org/abs/2310.10505", "detail_path": "post-training/2310.10505-remax/README.md", "topic": [ "经典 RLHF" ], "first_author": "Ziniu Li", "first_author_affiliation": "香港中文大学（深圳）/ 深圳市大数据研究院", "published": "2023-10-16", "code": "https://github.com/liziniu/ReMax", "adapter": "remax" }, { "domain": "post-training", "key": "retire-opd", "title": "RetireOPD: Self-Retiring On-Policy Distillation for Agentic Reinforcement Learning", "paper_url": "https://arxiv.org/abs/2609.20784", "detail_path": "post-training/2609.20784-retire-opd/README.md", "topic": [ "On-policy distillation", "Agent RL" ], "first_author": "ZJU-REAL team", "first_author_affiliation": "Zhejiang University", "published": "2026-09-17", "code": "https://github.com/ZJU-REAL/SDAR", "adapter": "retire-opd", "priority": "P0" }, { "domain": "post-training", "key": "ripo", "title": "RIPO", "paper_url": "https://arxiv.org/abs/2607.10169", "detail_path": "post-training/2607.10169-ripo/README.md", "topic": [ "几何信任域" ], "first_author": "Zhicheng Cai", "first_author_affiliation": "Tsinghua University", "published": "2026-07-11", "code": "未发现官方代码", "adapter": "ripo" }, { "domain": "post-training", "key": "rlaif", "title": "RLAIF", "paper_url": "https://arxiv.org/abs/2309.00267", "detail_path": "post-training/2309.00267-rlaif/README.md", "topic": [ "AI 反馈" ], "first_author": "Harrison Lee", "first_author_affiliation": "Google Research", "published": "2023-09-01", "code": "未发现官方代码", "adapter": "rlaif" }, { "domain": "post-training", "key": "rlcr", "title": "Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty", "paper_url": "https://arxiv.org/abs/2507.16806", "detail_path": "post-training/2507.16806-rlcr/README.md", "topic": [ "校准奖励", "Proper scoring rule" ], "first_author": "Mehul Damani", "first_author_affiliation": "Massachusetts Institute of Technology", "published": "2025-07-22", "code": "https://github.com/damanimehul/RLCR", "adapter": "system-one:rival-brier", "priority": "P1" }, { "domain": "post-training", "key": "rloo", "title": "RLOO", "paper_url": "https://arxiv.org/abs/2402.14740", "detail_path": "post-training/2402.14740-rloo/README.md", "topic": [ "经典 RLHF" ], "first_author": "Arash Ahmadian", "first_author_affiliation": "Cohere For AI", "published": "2024-02-22", "code": "未发现官方代码", "adapter": "rloo" }, { "domain": "post-training", "key": "rlvr-fusion", "title": "Consolidating RLVR Capabilities Across Domains: A Deep Dive into Fusion Paradigms", "paper_url": "https://arxiv.org/abs/2608.27409", "detail_path": "post-training/2608.27409-rlvr-fusion/README.md", "topic": [ "跨领域 RLVR 能力融合" ], "first_author": "Siye Wu", "first_author_affiliation": "Fudan University", "published": "2026-08-27", "code": "https://github.com/Di-viner/LLM-Fusion", "adapter": "rlvr-fusion" }, { "domain": "post-training", "key": "route-opd", "title": "Distillation as Probability Transport: Routed On-Policy Distillation", "paper_url": "https://arxiv.org/abs/2609.08337", "detail_path": "post-training/2609.08337-route-opd/README.md", "topic": [ "概率传输 OPD" ], "first_author": "Tianle Xia", "first_author_affiliation": "Tencent", "published": "2026-09-08", "code": null, "adapter": "route-opd" }, { "domain": "post-training", "key": "rp-opsd", "title": "RP-OPSD", "paper_url": "https://arxiv.org/abs/2608.06347", "detail_path": "post-training/2608.06347-rp-opsd/README.md", "topic": [ "推理枢纽蒸馏" ], "first_author": "Xinye Wang", "first_author_affiliation": "Nanjing University", "published": "2026-08-06", "code": "https://github.com/NJUNLP/RP-OPSD", "adapter": "rp-opsd" }, { "domain": "post-training", "key": "rrc", "title": "RRC: Unlocking Generative Reward Models in LLM Reinforcement Learning via Ranking-Based Reward Construction", "paper_url": "https://arxiv.org/abs/2608.06310", "detail_path": "post-training/2608.06310-rrc/README.md", "topic": [ "生成式奖励模型" ], "first_author": "Chenglong Wang", "first_author_affiliation": "Northeastern University", "published": "2026-08-06", "code": "https://github.com/wangclnlp/RRC", "adapter": "rrc" }, { "domain": "post-training", "key": "rrhf", "title": "RRHF", "paper_url": "https://arxiv.org/abs/2304.05302", "detail_path": "post-training/2304.05302-rrhf/README.md", "topic": [ "全排序偏好" ], "first_author": "Zheng Yuan", "first_author_affiliation": "Alibaba DAMO Academy", "published": "2023-04-11", "code": "https://github.com/GanjinZero/RRHF", "adapter": "rrhf" }, { "domain": "post-training", "key": "rstg", "title": "Distill Where You Fail: Recovering Learning Signals of Negative RL-Groups from Adaptive Teacher Guidance", "paper_url": "https://arxiv.org/abs/2608.00782", "detail_path": "post-training/2608.00782-rstg/README.md", "topic": [ "OPD", "RLVR" ], "first_author": "Zhuowen Han", "first_author_affiliation": "Tianjin University", "published": "2026-08-01", "code": null, "adapter": "rstg" }, { "domain": "post-training", "key": "rubric-dropout", "title": "Rubric Dropout: A Simple Way to Mitigate Reward Hacking in Rubric-as-Reward RL", "paper_url": "https://arxiv.org/abs/2608.11669", "detail_path": "post-training/2608.11669-rubric-dropout/README.md", "topic": [ "Rubric RL", "reward hacking" ], "first_author": "Minglai Yang", "first_author_affiliation": "Scale AI", "published": "2026-08-12", "code": null, "adapter": "rubric-dropout" }, { "domain": "post-training", "key": "sa-mrpo", "title": "Learn What's Left, Not What's Mastered: Saturation Aware Advantage Reweighting for Multi-Reward Policy Optimization", "paper_url": "https://arxiv.org/abs/2608.16072", "detail_path": "post-training/2608.16072-sa-mrpo/README.md", "topic": [ "多奖励 RL", "稳定性" ], "first_author": "Yixuan Wang", "first_author_affiliation": "University of Florida", "published": "2026-08-17", "code": null, "adapter": "sa-mrpo" }, { "domain": "post-training", "key": "sd-dpo", "title": "Style-Debiased DPO: Updating LLM Knowledge with Factuality-Aware Synthetic Preference Data", "paper_url": "https://arxiv.org/abs/2609.16532", "detail_path": "post-training/2609.16532-sd-dpo/README.md", "topic": [ "偏好优化", "知识更新" ], "first_author": "Takayuki Yamamoto", "first_author_affiliation": "Waseda University", "published": "2026-09-15", "code": null, "adapter": "sd-dpo", "priority": "P1" }, { "domain": "post-training", "key": "self-rewarding", "title": "Self-Rewarding LM", "paper_url": "https://arxiv.org/abs/2401.10020", "detail_path": "post-training/2401.10020-self-rewarding/README.md", "topic": [ "自奖励" ], "first_author": "Weizhe Yuan", "first_author_affiliation": "Meta AI", "published": "2024-01-18", "code": "未发现官方代码", "adapter": "self-rewarding" }, { "domain": "post-training", "key": "serpo", "title": "SERPO: Self-Evolving Rubric Policy Optimization for Open-Ended Test-Time Reinforcement Learning", "paper_url": "https://arxiv.org/abs/2607.26873", "detail_path": "post-training/2607.26873-serpo/README.md", "topic": [ "Rubric RL", "测试时自进化" ], "first_author": "Jianze Wang", "first_author_affiliation": "Authors did not disclose affiliation", "published": "2026-07-29", "code": null, "adapter": "serpo" }, { "domain": "post-training", "key": "simpo", "title": "SimPO", "paper_url": "https://arxiv.org/abs/2405.14734", "detail_path": "post-training/2405.14734-simpo/README.md", "topic": [ "Reference-free 偏好" ], "first_author": "Yu Meng", "first_author_affiliation": "Princeton University", "published": "2024-05-23", "code": "https://github.com/princeton-nlp/SimPO", "adapter": "simpo" }, { "domain": "post-training", "key": "sis", "title": "Turning Off-Policy Tokens On-Policy: A Plug-in Approach for Improving LLM Alignment", "paper_url": "https://arxiv.org/abs/2607.04728", "detail_path": "reproductions/2607.04728-sis/README.md", "topic": [ "post-training", "off-policy-correction", "importance-sampling" ], "first_author": "Yu Li", "first_author_affiliation": "Renmin University of China", "published": "2026-07-06", "code": null, "adapter": { "adapter_key": "sis", "arxiv_id": "2607.04728", "title": "Turning Off-Policy Tokens On-Policy: A Plug-in Approach for Improving LLM Alignment", "paper_url": "https://arxiv.org/abs/2607.04728", "track": "llm", "organization": "Paper author team", "published": "2026-07-06", "code_url": null, "topics": [ "post-training", "off-policy-correction", "importance-sampling" ], "local_code_dir": "src/auto_research/reproductions/sis", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "Tiny Shakespeare" ], "baseline": "标准 token-level Importance Sampling", "metrics": [ "sis.reported_objective" ], "default_seeds": [ 41, 42, 43 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "post-training", "key": "slic-hf", "title": "SLiC-HF", "paper_url": "https://arxiv.org/abs/2305.10425", "detail_path": "post-training/2305.10425-slic-hf/README.md", "topic": [ "序列概率校准" ], "first_author": "Yao Zhao", "first_author_affiliation": "Google DeepMind", "published": "2023-05-17", "code": "未发现官方代码", "adapter": "slic-hf" }, { "domain": "post-training", "key": "smopd", "title": "SMOPD: Multi-Reward Reinforcement Learning via Specialize-and-Merge Online Policy Distillation", "paper_url": "https://arxiv.org/abs/2608.03092", "detail_path": "post-training/2608.03092-smopd/README.md", "topic": [ "OPD", "多奖励 RL" ], "first_author": "Wen Wang", "first_author_affiliation": "Alibaba / Qwen", "published": "2026-08-04", "code": null, "adapter": "smopd" }, { "domain": "post-training", "key": "sparse-opd", "title": "Extremely Sparse Supervision Incentivizes Reasoning Ability", "paper_url": "https://arxiv.org/abs/2609.04565", "detail_path": "post-training/2609.04565-sparse-opd/README.md", "topic": [ "极稀疏 On-policy distillation" ], "first_author": "Zhishuai Liu", "first_author_affiliation": "Amazon / Duke University", "published": "2026-09-03", "code": null, "adapter": "sparse-opd" }, { "domain": "post-training", "key": "spear", "title": "SPEAR: Distilling Domain-Adaptive Reasoning Skeletons via Sequential Symbolic Alignment in Reinforcement Learning", "paper_url": "https://arxiv.org/abs/2608.26550", "detail_path": "post-training/2608.26550-spear/README.md", "topic": [ "过程奖励" ], "first_author": "Zhuochun Li", "first_author_affiliation": "University of Pittsburgh", "published": "2026-08-27", "code": "https://github.com/zhuochunli/SPEAR", "adapter": "spear" }, { "domain": "post-training", "key": "specroll", "title": "SpecRoll: Fast-Slow Verifier-Feedback Adaptation for Speculative Reinforcement Learning Rollouts", "paper_url": "https://arxiv.org/abs/2608.04962", "detail_path": "post-training/2608.04962-specroll/README.md", "topic": [ "RL rollout 加速" ], "first_author": "Nhat Minh Pham", "first_author_affiliation": "VNU University of Engineering and Technology / Viettel AI", "published": "2026-08-05", "code": "https://anonymous.4open.science/r/SpecRoll-26062006", "adapter": "specroll" }, { "domain": "post-training", "key": "spin", "title": "SPIN", "paper_url": "https://arxiv.org/abs/2401.01335", "detail_path": "post-training/2401.01335-spin/README.md", "topic": [ "自博弈微调" ], "first_author": "Zixiang Chen", "first_author_affiliation": "University of California, Los Angeles", "published": "2024-01-02", "code": "https://github.com/uclaml/SPIN", "adapter": "spin" }, { "domain": "post-training", "key": "spiral", "title": "SPIRAL", "paper_url": "https://arxiv.org/abs/2506.24119", "detail_path": "post-training/2506.24119-spiral/README.md", "topic": [ "自博弈课程" ], "first_author": "Bo Liu", "first_author_affiliation": "Apple", "published": "2025-06-30", "code": "https://github.com/spiral-rl/spiral", "adapter": "spiral" }, { "domain": "post-training", "key": "sr-opsd", "title": "SR-OPSD: Self-Referenced On-Policy Self-Distillation", "paper_url": "https://arxiv.org/abs/2608.09745", "detail_path": "post-training/2608.09745-sr-opsd/README.md", "topic": [ "OPD", "自蒸馏" ], "first_author": "Zhuo Sun", "first_author_affiliation": "Independent Researcher", "published": "2026-08-10", "code": null, "adapter": "sr-opsd" }, { "domain": "post-training", "key": "srpo", "title": "SRPO: Self-Reflective Policy Optimization for Long-Horizon Reasoning", "paper_url": "https://arxiv.org/abs/2608.23493", "detail_path": "post-training/2608.23493-srpo/README.md", "topic": [ "反思式 token 信用" ], "first_author": "Jialong Liu", "first_author_affiliation": "Wuhan University", "published": "2026-08-24", "code": "https://github.com/Galleons2029/SRPO", "adapter": "srpo" }, { "domain": "post-training", "key": "stare", "title": "STARE", "paper_url": "https://arxiv.org/abs/2606.19236", "detail_path": "post-training/2606.19236-stare/README.md", "topic": [ "Entropy 稳定" ], "first_author": "Haipeng Luo", "first_author_affiliation": "Tsinghua University", "published": "2026-06-17", "code": "https://github.com/hp-luo/STARE", "adapter": "stare" }, { "domain": "post-training", "key": "steerlm", "title": "SteerLM", "paper_url": "https://arxiv.org/abs/2310.05344", "detail_path": "post-training/2310.05344-steerlm/README.md", "topic": [ "多属性可控 SFT" ], "first_author": "Yi Dong", "first_author_affiliation": "NVIDIA", "published": "2023-10-09", "code": "https://github.com/NVIDIA/NeMo-Aligner", "adapter": "steerlm" }, { "domain": "post-training", "key": "stride-opd", "title": "Know When to Stop, Where to Restart: Accelerating Multi-Turn Agentic On-Policy Distillation", "paper_url": "https://arxiv.org/abs/2609.14636", "detail_path": "post-training/2609.14636-stride-opd/README.md", "topic": [ "On-policy distillation", "多轮加速" ], "first_author": "Zhiyu Gui", "first_author_affiliation": "University of Science and Technology of China / Ant Group", "published": "2026-09-13", "code": null, "adapter": "stride-opd", "priority": "P0" }, { "domain": "post-training", "key": "taco", "title": "TACO", "paper_url": "https://arxiv.org/abs/2607.07976", "detail_path": "post-training/2607.07976-taco/README.md", "topic": [ "Token 信用校准" ], "first_author": "Xiuyi Lou", "first_author_affiliation": "Johns Hopkins University", "published": "2026-07-08", "code": "https://github.com/xiuyilou/TACO", "adapter": "taco" }, { "domain": "post-training", "key": "tasco", "title": "Beyond Confidence: Stability-Aware Test-Time Adaptation for LLM Reasoning", "paper_url": "https://arxiv.org/abs/2609.11393", "detail_path": "post-training/2609.11393-tasco/README.md", "topic": [ "测试时适配", "稳定性优化" ], "first_author": "Bincheng Gu", "first_author_affiliation": "Chongqing University", "published": "2026-09-10", "code": null, "adapter": "tasco", "priority": "P1" }, { "domain": "post-training", "key": "tcr", "title": "TCR", "paper_url": "https://arxiv.org/abs/2607.19824", "detail_path": "post-training/2607.19824-tcr/README.md", "topic": [ "过程奖励" ], "first_author": "Xubo Liu", "first_author_affiliation": "论文未列机构", "published": "2026-07-22", "code": "未发现官方代码", "adapter": "tcr" }, { "domain": "post-training", "key": "tiao", "title": "TIAO: Token Importance-Aware Policy Optimization for Text Summarization", "paper_url": "https://arxiv.org/abs/2609.16748", "detail_path": "post-training/2609.16748-tiao/README.md", "topic": [ "强化学习", "Token 信用分配" ], "first_author": "Qixiu Li", "first_author_affiliation": "National University of Defense Technology", "published": "2026-09-15", "code": "https://github.com/TechCloud-x/TIAO", "adapter": "tiao", "priority": "P0", "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/tiao-a100-20260916.json" }, { "domain": "post-training", "key": "tis", "title": "TIS", "paper_url": "web-2025-tis/README.md", "detail_path": "post-training/web-2025-tis/README.md", "topic": [ "训推失配校正" ], "first_author": "Feng Yao", "first_author_affiliation": "University of California, San Diego", "published": "2025-08-05", "code": "未发现官方代码", "adapter": "tis" }, { "domain": "post-training", "key": "tlm-dre", "title": "Turn-level Multiscale Density Ratio Estimation for LLM Agents", "paper_url": "https://arxiv.org/abs/2609.16760", "detail_path": "post-training/2609.16760-tlm-dre/README.md", "topic": [ "多轮 Agent 对齐", "密度比估计" ], "first_author": "Zishuo Zhao", "first_author_affiliation": "Alibaba Group", "published": "2026-09-15", "code": null, "adapter": "tlm-dre", "priority": "P0" }, { "domain": "post-training", "key": "trajectory-learnability", "title": "Trajectory Learnability for Offline On-Policy Distillation", "paper_url": "https://arxiv.org/abs/2609.18321", "detail_path": "post-training/2609.18321-trajectory-learnability/README.md", "topic": [ "On-policy distillation", "轨迹加权" ], "first_author": "Yihao Ai", "first_author_affiliation": "National University of Singapore", "published": "2026-09-16", "code": null, "adapter": "trajectory-learnability", "priority": "P1" }, { "domain": "post-training", "key": "ttpo", "title": "TTPO: Test-Time Policy Optimization", "paper_url": "https://arxiv.org/abs/2608.27448", "detail_path": "post-training/2608.27448-ttpo/README.md", "topic": [ "测试时强化学习" ], "first_author": "Aozhe Wang", "first_author_affiliation": "Zhejiang University", "published": "2026-08-27", "code": "https://github.com/ZJU-REAL/TTPO", "adapter": "ttpo" }, { "domain": "post-training", "key": "ttrl", "title": "TTRL", "paper_url": "https://arxiv.org/abs/2504.16084", "detail_path": "post-training/2504.16084-ttrl/README.md", "topic": [ "测试时强化学习" ], "first_author": "Yuxin Zuo", "first_author_affiliation": "PRIME-RL author team", "published": "2025-04-22", "code": "https://github.com/PRIME-RL/TTRL", "adapter": "ttrl" }, { "domain": "post-training", "key": "u-opsd", "title": "U-OPSD", "paper_url": "https://arxiv.org/abs/2608.06296", "detail_path": "post-training/2608.06296-u-opsd/README.md", "topic": [ "无监督自蒸馏" ], "first_author": "Yijiang Li", "first_author_affiliation": "University of California, San Diego", "published": "2026-08-06", "code": "未发现官方代码", "adapter": "u-opsd" }, { "domain": "post-training", "key": "uc-mopd", "title": "Preserving General Capabilities during Domain Specialization with Uncertainty-Calibrated MOPD", "paper_url": "https://arxiv.org/abs/2608.26735", "detail_path": "post-training/2608.26735-uc-mopd/README.md", "topic": [ "多教师能力整合" ], "first_author": "Ziyuan Liu", "first_author_affiliation": "Peking University", "published": "2026-08-27", "code": null, "adapter": "uc-mopd" }, { "domain": "post-training", "key": "v-rubrics", "title": "V-Rubrics: Visual Faithfulness via Rubric-Based Reinforcement Learning", "paper_url": "https://arxiv.org/abs/2608.25580", "detail_path": "post-training/2608.25580-v-rubrics/README.md", "topic": [ "多模态 rubric RL" ], "first_author": "Shulin Tian", "first_author_affiliation": "S-Lab, Nanyang Technological University", "published": "2026-08-26", "code": "https://shulin16.github.io/v-rubrics/", "adapter": "v-rubrics" }, { "domain": "post-training", "key": "vad", "title": "VAD", "paper_url": "https://arxiv.org/abs/2607.28590", "detail_path": "post-training/2607.28590-vad/README.md", "topic": [ "多模态证据归因蒸馏" ], "first_author": "Kangning Zhang", "first_author_affiliation": "Shanghai Jiao Tong University", "published": "2026-07-30", "code": "https://github.com/DeepExperience/VAD_Multimodal_OPD", "adapter": "vad" }, { "domain": "post-training", "key": "vapo", "title": "VAPO", "paper_url": "https://arxiv.org/abs/2504.05118", "detail_path": "post-training/2504.05118-vapo/README.md", "topic": [ "Critic PPO" ], "first_author": "Yu Yue", "first_author_affiliation": "ByteDance Seed", "published": "2025-04-07", "code": "未发现官方代码", "adapter": "vapo" }, { "domain": "post-training", "key": "video-opsd", "title": "Video-OPSD: Exploiting Privileged Visual Evidence for On-Policy Self-Distillation in Video Large Language Models", "paper_url": "https://arxiv.org/abs/2608.27065", "detail_path": "post-training/2608.27065-video-opsd/README.md", "topic": [ "证据帧特权自蒸馏" ], "first_author": "Ziyue Wang", "first_author_affiliation": "Nanyang Technological University", "published": "2026-08-27", "code": null, "adapter": "video-opsd" }, { "domain": "post-training", "key": "weak-guide-rlvr", "title": "Boosting LLM Exploration via Weak-Model Guidance in RLVR", "paper_url": "https://arxiv.org/abs/2608.27420", "detail_path": "post-training/2608.27420-weak-guide-rlvr/README.md", "topic": [ "弱模型前缀探索" ], "first_author": "Xingyu Shen", "first_author_affiliation": "Peking University", "published": "2026-08-27", "code": null, "adapter": "weak-guide-rlvr" }, { "domain": "recommendation", "key": "adaf2m2", "title": "AdaF²M²: Comprehensive Learning and Responsive Leveraging Features in Recommendation System", "paper_url": "https://arxiv.org/abs/2501.15816", "detail_path": "reproductions/2501.15816-adaf2m2/README.md", "topic": [ "ranking", "feature-masking", "state-aware-adapter", "multi-forward" ], "first_author": null, "first_author_affiliation": "ByteDance / Douyin", "published": "2025-01-27", "code": null, "adapter": { "adapter_key": "adaf2m2", "arxiv_id": "2501.15816", "title": "AdaF²M²: Comprehensive Learning and Responsive Leveraging Features in Recommendation System", "paper_url": "https://arxiv.org/abs/2501.15816", "track": "recommendation", "organization": "ByteDance / Douyin", "published": "2025-01-27", "code_url": null, "topics": [ "ranking", "feature-masking", "state-aware-adapter", "multi-forward" ], "local_code_dir": "src/auto_research/reproductions/adaf2m2", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "single-forward full-feature ranker", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.mrr_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.mrr_at_10", "method.ndcg_at_10", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.mrr_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Douyin", "metric": "cumulative active days", "lift_percent": 1.37, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2501.15816", "source_location": "original paper online-result disclosure: product=Douyin; metric=cumulative active days" }, { "product": "Douyin", "metric": "app duration", "lift_percent": 1.89, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2501.15816", "source_location": "original paper online-result disclosure: product=Douyin; metric=app duration" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "adaptive-ad-load", "title": "Adaptive Ad Load Design for Sponsored Search Markets: Evidence, Theory, and Deployment", "paper_url": "https://arxiv.org/abs/2607.14418", "detail_path": "reproductions/2607.14418-adaptive-ad-load/README.md", "topic": [ "advertising", "constrained-optimization", "marketplace" ], "first_author": null, "first_author_affiliation": "University of Washington", "published": "2026-07-15", "code": null, "adapter": { "adapter_key": "adaptive-ad-load", "arxiv_id": "2607.14418", "title": "Adaptive Ad Load Design for Sponsored Search Markets: Evidence, Theory, and Deployment", "paper_url": "https://arxiv.org/abs/2607.14418", "track": "recommendation", "organization": "University of Washington", "published": "2026-07-15", "code_url": null, "topics": [ "advertising", "constrained-optimization", "marketplace" ], "local_code_dir": "src/auto_research/reproductions/adaptive_ad_load", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Android app-store sponsored search", "metric": "revenue", "lift_percent": 36.8, "traffic": "66-day randomized field experiment", "source_url": "https://arxiv.org/html/2607.14418v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "adasid", "title": "Beyond Static Collision Handling: Adaptive Semantic ID Learning for Multimodal Recommendation at Industrial Scale", "paper_url": "https://arxiv.org/abs/2604.23522", "detail_path": "reproductions/2604.23522-adasid/README.md", "topic": [ "semantic-id", "multimodal-recommendation", "retrieval" ], "first_author": null, "first_author_affiliation": "University of Electronic Science and Technology of China / Kuaishou", "published": "2026-04-26", "code": null, "adapter": { "adapter_key": "adasid", "arxiv_id": "2604.23522", "title": "Beyond Static Collision Handling: Adaptive Semantic ID Learning for Multimodal Recommendation at Industrial Scale", "paper_url": "https://arxiv.org/abs/2604.23522", "track": "recommendation", "organization": "University of Electronic Science and Technology of China / Kuaishou", "published": "2026-04-26", "code_url": null, "topics": [ "semantic-id", "multimodal-recommendation", "retrieval" ], "local_code_dir": "src/auto_research/reproductions/adasid", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou e-commerce", "metric": "GPM", "lift_percent": 1.16, "traffic": "tens of millions of users, four days", "source_url": "https://arxiv.org/html/2604.23522v1", "source_location": "Section 5.3 / Table 3", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "agentic-rec-tune", "title": "AgenticRecTune: Multi-Agent with Self-Evolving Skillhub for Recommendation System Optimization", "paper_url": "https://arxiv.org/abs/2604.26969", "detail_path": "reproductions/2604.26969-agentic-rec-tune/README.md", "topic": [ "agent", "auto-research", "retrieval", "ranking", "reranking" ], "first_author": null, "first_author_affiliation": "Google / Discover", "published": "2026-04-21", "code": null, "adapter": { "adapter_key": "agentic-rec-tune", "arxiv_id": "2604.26969", "title": "AgenticRecTune: Multi-Agent with Self-Evolving Skillhub for Recommendation System Optimization", "paper_url": "https://arxiv.org/abs/2604.26969", "track": "recommendation", "organization": "Google / Discover", "published": "2026-04-21", "code_url": null, "topics": [ "agent", "auto-research", "retrieval", "ranking", "reranking" ], "local_code_dir": "src/auto_research/reproductions/agentic_rec_tune", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (220 users / 360 items)" ], "baseline": "固定手调排序配方，实验组运行三轮 Actor–Critic–SkillHub", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Google Discover value retrieval", "metric": "Engagement 2", "lift_percent": 0.9, "traffic": "live-traffic A/B", "source_url": "https://arxiv.org/html/2604.26969v2#S5", "source_location": "Section 5 Table 1", "significance": "p<0.05", "retrieved_at": "2026-08-09" }, { "product": "Google Discover diversity reranking", "metric": "Diversity", "lift_percent": 3.43, "traffic": "live-traffic A/B", "source_url": "https://arxiv.org/html/2604.26969v2#S5", "source_location": "Section 5 Table 1", "significance": "p<0.05", "retrieved_at": "2026-08-09" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "aigq", "title": "AIGQ: An End-to-End Hybrid Generative Architecture for E-commerce Query Recommendation", "paper_url": "https://arxiv.org/abs/2603.19710", "detail_path": "reproductions/2603.19710-aigq/README.md", "topic": [ "query-recommendation", "generative-recommendation", "reinforcement-learning" ], "first_author": null, "first_author_affiliation": "Taobao & Tmall Group / Alibaba", "published": "2026-03-20", "code": null, "adapter": { "adapter_key": "aigq", "arxiv_id": "2603.19710", "title": "AIGQ: An End-to-End Hybrid Generative Architecture for E-commerce Query Recommendation", "paper_url": "https://arxiv.org/abs/2603.19710", "track": "recommendation", "organization": "Taobao & Tmall Group / Alibaba", "published": "2026-03-20", "code_url": null, "topics": [ "query-recommendation", "generative-recommendation", "reinforcement-learning" ], "local_code_dir": "src/auto_research/reproductions/aigq", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao HintQ", "metric": "GMV", "lift_percent": 10.68, "traffic": "thirty-day A/B", "source_url": "https://arxiv.org/html/2603.19710v1", "source_location": "Section 5.4", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "airbnb-ebr", "title": "Applying Embedding-Based Retrieval to Airbnb Search", "paper_url": "https://arxiv.org/abs/2601.06873", "detail_path": "reproductions/2601.06873-airbnb-ebr/README.md", "topic": [ "search-retrieval", "embedding-retrieval", "hard-negative-mining" ], "first_author": null, "first_author_affiliation": "Airbnb", "published": "2026-01-11", "code": null, "adapter": { "adapter_key": "airbnb-ebr", "arxiv_id": "2601.06873", "title": "Applying Embedding-Based Retrieval to Airbnb Search", "paper_url": "https://arxiv.org/abs/2601.06873", "track": "recommendation", "organization": "Airbnb", "published": "2026-01-11", "code_url": null, "topics": [ "search-retrieval", "embedding-retrieval", "hard-negative-mining" ], "local_code_dir": "src/auto_research/reproductions/airbnb_ebr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Airbnb search", "metric": "Booking conversion", "lift_percent": 0.31, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2601.06873v1", "source_location": "Section 6.3", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "context:journey-retrieval" ] } }, { "domain": "recommendation", "key": "akt-rec", "title": "From Head to Tail: Asymmetric Knowledge Transfer in Long-tail Recommendation with Generative Semantic IDs", "paper_url": "https://arxiv.org/abs/2605.23310", "detail_path": "reproductions/2605.23310-akt-rec/README.md", "topic": [ "llm-recommendation", "semantic-id", "long-tail", "ctr-ranking" ], "first_author": null, "first_author_affiliation": "Alibaba Group / Peking University", "published": "2026-05-22", "code": null, "adapter": { "adapter_key": "akt-rec", "arxiv_id": "2605.23310", "title": "From Head to Tail: Asymmetric Knowledge Transfer in Long-tail Recommendation with Generative Semantic IDs", "paper_url": "https://arxiv.org/abs/2605.23310", "track": "recommendation", "organization": "Alibaba Group / Peking University", "published": "2026-05-22", "code_url": null, "topics": [ "llm-recommendation", "semantic-id", "long-tail", "ctr-ranking" ], "local_code_dir": "src/auto_research/reproductions/akt_rec", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K sampled CTR" ], "baseline": "只使用 individual embedding 和实例级 target attention 的 onlinebase", "metrics": [ "relative_akt_rec_vs_online_base_percent.auc", "relative_akt_rec_vs_online_base_percent.gauc", "relative_akt_rec_vs_online_base_percent.tail_auc", "setup.ctr_steps_per_variant_seed", "test.akt_rec.auc", "test.akt_rec.auc_std", "test.akt_rec.gauc", "test.akt_rec.gauc_std", "test.akt_rec.tail_auc", "test.akt_rec.tail_auc_std", "test.online_base.auc", "test.online_base.auc_std", "test.online_base.gauc", "test.online_base.gauc_std", "test.online_base.tail_auc", "test.online_base.tail_auc_std", "validation.akt_rec.auc", "validation.akt_rec.gauc", "validation.akt_rec.tail_auc", "validation.online_base.auc", "validation.online_base.gauc", "validation.online_base.tail_auc" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=400", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Tmall", "metric": "CTR", "lift_percent": 2.76, "traffic": "10% control / 10% treatment, two weeks", "source_url": "https://arxiv.org/abs/2605.23310", "source_location": "original paper online-result disclosure: product=Tmall; metric=CTR" }, { "product": "Tmall", "metric": "GMV", "lift_percent": 3.47, "traffic": "10% control / 10% treatment, two weeks", "source_url": "https://arxiv.org/abs/2605.23310", "source_location": "original paper online-result disclosure: product=Tmall; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "allecompanion", "title": "Beyond Co-purchase Relation: Evolution of Complementary Recommendations at Allegro", "paper_url": "https://arxiv.org/abs/2609.05063", "detail_path": "reproductions/2609.05063-allecompanion/README.md", "topic": [ "complementary-recommendation", "two-tower", "category-adapter", "e-commerce" ], "first_author": null, "first_author_affiliation": "Allegro.com", "published": "2026-09-04", "code": null, "adapter": { "adapter_key": "allecompanion", "arxiv_id": "2609.05063", "title": "Beyond Co-purchase Relation: Evolution of Complementary Recommendations at Allegro", "paper_url": "https://arxiv.org/abs/2609.05063", "track": "recommendation", "organization": "Allegro.com", "published": "2026-09-04", "code_url": null, "topics": [ "complementary-recommendation", "two-tower", "category-adapter", "e-commerce" ], "local_code_dir": "src/auto_research/reproductions/allecompanion", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "shared two-tower without category adapter", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; 80 CPU updates; validation-only candidate selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Allegro product and cart placements", "metric": "attributed GMV", "lift_percent": 21.25, "traffic": "two-week tests routed over 100% platform traffic", "source_url": "https://arxiv.org/html/2609.05063", "source_location": "Section 4.4, Table 5", "experiment_duration": "two weeks", "retrieved_at": "2026-09-07" } ], "selection_exception": null, "evolve_operators": [ "context:complementary-category-adapter" ] } }, { "domain": "recommendation", "key": "angle", "title": "One-Step Retrieval Framework for Real-Time Sponsored Search Ads Using Hierarchical Text Representations", "paper_url": "https://arxiv.org/abs/2609.18296", "detail_path": "reproductions/2609.18296-angle/README.md", "topic": [ "sponsored-search", "generative-retrieval", "hierarchical-identifiers" ], "first_author": "Jinrui Zhang", "first_author_affiliation": "Tencent", "published": "2026-09-16", "code": null, "adapter": { "adapter_key": "angle", "arxiv_id": "2609.18296", "title": "One-Step Retrieval Framework for Real-Time Sponsored Search Ads Using Hierarchical Text Representations", "paper_url": "https://arxiv.org/abs/2609.18296", "track": "recommendation", "organization": "Tencent", "published": "2026-09-16", "code_url": null, "topics": [ "sponsored-search", "generative-retrieval", "hierarchical-identifiers" ], "local_code_dir": "src/auto_research/reproductions/angle", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Weixin Top Stories sponsored search", "metric": "consumption", "lift_percent": 1.81, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2609.18296v1", "source_location": "online experiment table", "retrieved_at": "2026-09-19" }, { "product": "Weixin Top Stories sponsored search", "metric": "GMV", "lift_percent": 2.16, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2609.18296v1", "source_location": "online experiment table", "retrieved_at": "2026-09-19" }, { "product": "Weixin Top Stories sponsored search", "metric": "clicks", "lift_percent": 1.5, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2609.18296v1", "source_location": "online experiment table", "retrieved_at": "2026-09-19" }, { "product": "Weixin Top Stories sponsored search", "metric": "conversions", "lift_percent": 1.44, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2609.18296v1", "source_location": "online experiment table", "retrieved_at": "2026-09-19" }, { "product": "Weixin Top Stories sponsored search", "metric": "exposure", "lift_percent": 2.49, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2609.18296v1", "source_location": "online experiment table", "retrieved_at": "2026-09-19" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "argus", "title": "Scaling Recommender Transformers to One Billion Parameters", "paper_url": "https://arxiv.org/abs/2507.15994", "detail_path": "reproductions/2507.15994-argus/README.md", "topic": [ "long-sequence", "transformer", "multi-task" ], "first_author": null, "first_author_affiliation": "Yandex", "published": "2025-07", "code": null, "adapter": { "adapter_key": "argus", "arxiv_id": "2507.15994", "title": "Scaling Recommender Transformers to One Billion Parameters", "paper_url": "https://arxiv.org/abs/2507.15994", "track": "recommendation", "organization": "Yandex", "published": "2025-07", "code_url": null, "topics": [ "long-sequence", "transformer", "multi-task" ], "local_code_dir": "src/auto_research/reproductions/argus", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "DIN，ARGUS NDCG@10 相对 DIN -4.12%", "metrics": [ "results.din.hit_at_10", "results.din.ndcg_at_10", "results.feedback_then_item.hit_at_10", "results.feedback_then_item.ndcg_at_10", "results.ndcg_ablation_percent", "results.ndcg_vs_din_percent", "results.next_item_only.hit_at_10", "results.next_item_only.ndcg_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=100", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Yandex Music", "metric": "total listening time", "lift_percent": 2.26, "traffic": "large-scale online experiment", "source_url": "https://arxiv.org/abs/2507.15994", "source_location": "original paper online-result disclosure: product=Yandex Music; metric=total listening time" }, { "product": "Yandex Music", "metric": "likes", "lift_percent": 6.37, "traffic": "large-scale online experiment", "source_url": "https://arxiv.org/abs/2507.15994", "source_location": "original paper online-result disclosure: product=Yandex Music; metric=likes" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "asarl", "title": "ASARL: Autonomous Social-Aware Relevance Learning for QQ Search", "paper_url": "https://arxiv.org/abs/2607.26593", "detail_path": "reproductions/2607.26593-asarl/README.md", "topic": [ "search-relevance", "multi-agent-data-curation", "knowledge-distillation" ], "first_author": null, "first_author_affiliation": "Tencent PCG", "published": "2026-07-29", "code": null, "adapter": { "adapter_key": "asarl", "arxiv_id": "2607.26593", "title": "ASARL: Autonomous Social-Aware Relevance Learning for QQ Search", "paper_url": "https://arxiv.org/abs/2607.26593", "track": "recommendation", "organization": "Tencent PCG", "published": "2026-07-29", "code_url": null, "topics": [ "search-relevance", "multi-agent-data-curation", "knowledge-distillation" ], "local_code_dir": "src/auto_research/reproductions/asarl", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "matched control without the asarl mechanism", "metrics": [ "paper_results.channel_ctr_lift_percent", "paper_results.group_ctr_lift_percent", "relative.curated_examples_percent", "relative.final_loss_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.initial_loss_percent", "relative.ndcg_at_10_percent", "relative.parameters_percent", "relative.reason_critic_rounds_percent", "relative.social_distillation_percent", "variants.ASARL distilled.final_loss", "variants.ASARL distilled.hit_at_10", "variants.ASARL distilled.initial_loss", "variants.ASARL distilled.ndcg_at_10", "variants.online student.final_loss", "variants.online student.hit_at_10", "variants.online student.initial_loss", "variants.online student.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "steps=60", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "QQ channel search", "metric": "CTR", "lift_percent": 2.69, "traffic": "20% treatment / 20% control, >=7 days", "source_url": "https://arxiv.org/abs/2607.26593", "source_location": "original paper online-result disclosure: product=QQ channel search; metric=CTR" }, { "product": "QQ group search", "metric": "GSB", "lift_percent": 16.66, "traffic": "20% treatment / 20% control, >=7 days", "source_url": "https://arxiv.org/abs/2607.26593", "source_location": "original paper online-result disclosure: product=QQ group search; metric=GSB" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "atomic-intent", "title": "Atomic Intent Reasoning: Bringing LLM Semantics to Industrial Cross-Domain Recommendations", "paper_url": "https://arxiv.org/abs/2606.10357", "detail_path": "reproductions/2606.10357-atomic-intent/README.md", "topic": [ "cross-domain", "llm-recommendation", "intent-retrieval" ], "first_author": null, "first_author_affiliation": "The Hong Kong Polytechnic University", "published": "2026-06-09", "code": null, "adapter": { "adapter_key": "atomic-intent", "arxiv_id": "2606.10357", "title": "Atomic Intent Reasoning: Bringing LLM Semantics to Industrial Cross-Domain Recommendations", "paper_url": "https://arxiv.org/abs/2606.10357", "track": "recommendation", "organization": "The Hong Kong Polytechnic University", "published": "2026-06-09", "code_url": null, "topics": [ "cross-domain", "llm-recommendation", "intent-retrieval" ], "local_code_dir": "src/auto_research/reproductions/atomic_intent", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou E-commerce", "metric": "GMV", "lift_percent": 3.446, "traffic": "5.08% production traffic", "source_url": "https://arxiv.org/html/2606.10357v1", "source_location": "Section 5.6", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "context:atomic-intent-tree" ] } }, { "domain": "recommendation", "key": "autolr", "title": "AutoLR: Automating the Path from Research to Launch Review in Industrial Recommender Systems", "paper_url": "https://arxiv.org/abs/2609.04871", "detail_path": "reproductions/2609.04871-autolr/README.md", "topic": [ "auto-research", "experiment-controller", "launch-review", "industrial-recommendation" ], "first_author": null, "first_author_affiliation": "NetEase, Inc.", "published": "2026-09-04", "code": null, "adapter": { "adapter_key": "autolr", "arxiv_id": "2609.04871", "title": "AutoLR: Automating the Path from Research to Launch Review in Industrial Recommender Systems", "paper_url": "https://arxiv.org/abs/2609.04871", "track": "recommendation", "organization": "NetEase, Inc.", "published": "2026-09-04", "code_url": null, "topics": [ "auto-research", "experiment-controller", "launch-review", "industrial-recommendation" ], "local_code_dir": "src/auto_research/reproductions/autolr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; 80 CPU updates; validation-only candidate selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "NetEase DASHEN feed and immersive-video feed", "metric": "content-consumption time", "lift_percent": 10.83, "traffic": "nine Launch Review records with online A/B and production decisions", "source_url": "https://arxiv.org/html/2609.04871", "source_location": "Section 5.2, Table 1", "retrieved_at": "2026-09-07" } ], "selection_exception": null, "evolve_operators": [ "controller:autolr-evidence-council" ] } }, { "domain": "recommendation", "key": "baff", "title": "BAFF: Bid-Aware Filter Family for Mitigating Training Data Interference in RTB A/B Tests", "paper_url": "https://arxiv.org/abs/2609.08725", "detail_path": "reproductions/2609.08725-baff/README.md", "topic": [ "advertising", "rtb", "ab-testing", "training-data-interference" ], "first_author": null, "first_author_affiliation": "Dable", "published": "2026-09-08", "code": null, "adapter": { "adapter_key": "baff", "arxiv_id": "2609.08725", "title": "BAFF: Bid-Aware Filter Family for Mitigating Training Data Interference in RTB A/B Tests", "paper_url": "https://arxiv.org/abs/2609.08725", "track": "recommendation", "organization": "Dable", "published": "2026-09-08", "code_url": null, "topics": [ "advertising", "rtb", "ab-testing", "training-data-interference" ], "local_code_dir": "src/auto_research/reproductions/baff", "fidelity": "core_mechanism", "evaluation_tier": "l1_mechanism", "datasets": [ "controlled RTB simulation" ], "baseline": "naive pooled training", "metrics": [ "parameter_deviation", "retained_control_fraction" ], "default_seeds": [ 42, 43, 44 ], "budget": "2,400 simulated auctions", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "production DSP real-time bidding", "metric": "absolute CPC treatment-effect gap reduction versus naive shared log", "lift_percent": 95.53, "traffic": "single advertiser/SSP; two- and three-day measurement phases", "source_url": "https://arxiv.org/html/2609.08725", "source_location": "Section 5.4, Tables 6-7", "retrieved_at": "2026-09-12" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "bahe", "title": "Breaking the Length Barrier: LLM-Enhanced CTR Prediction in Long Textual User Behaviors", "paper_url": "https://arxiv.org/abs/2403.19347", "detail_path": "reproductions/2403.19347-bahe/README.md", "topic": [ "llm-recommendation", "ctr-ranking", "long-sequence", "efficiency" ], "first_author": null, "first_author_affiliation": "Ant Group", "published": "2024-03", "code": null, "adapter": { "adapter_key": "bahe", "arxiv_id": "2403.19347", "title": "Breaking the Length Barrier: LLM-Enhanced CTR Prediction in Long Textual User Behaviors", "paper_url": "https://arxiv.org/abs/2403.19347", "track": "recommendation", "organization": "Ant Group", "published": "2024-03", "code_url": null, "topics": [ "llm-recommendation", "ctr-ranking", "long-sequence", "efficiency" ], "local_code_dir": "src/auto_research/reproductions/bahe", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens 100K title text" ], "baseline": "逐样本运行完整文本上层微调", "metrics": [ "auc.bahe_atomic_hierarchical.mean", "auc.bahe_atomic_hierarchical.std", "auc.full_text_upper_tuning.mean", "auc.full_text_upper_tuning.std", "auc.mean_absolute_gain", "auc.mean_relative_gain_percent", "latency_milliseconds_per_example.bahe_atomic_hierarchical.mean", "latency_milliseconds_per_example.bahe_atomic_hierarchical.std", "latency_milliseconds_per_example.full_text_upper_tuning.mean", "latency_milliseconds_per_example.full_text_upper_tuning.std", "latency_milliseconds_per_example.mean_relative_change_percent" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=80", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "e-commerce ads", "metric": "CTR", "lift_percent": 9.65, "traffic": "two-week A/B test", "source_url": "https://arxiv.org/abs/2403.19347", "source_location": "original paper online-result disclosure: product=e-commerce ads; metric=CTR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "barge", "title": "Bridging the Structural Gap: Adapting Autoregressive Generation for Recommendation", "paper_url": "https://arxiv.org/abs/2607.21028", "detail_path": "reproductions/2607.21028-barge/README.md", "topic": [ "generative-recommendation", "semantic-id", "reranking" ], "first_author": null, "first_author_affiliation": "Tencent", "published": "2026-07-23", "code": null, "adapter": { "adapter_key": "barge", "arxiv_id": "2607.21028", "title": "Bridging the Structural Gap: Adapting Autoregressive Generation for Recommendation", "paper_url": "https://arxiv.org/abs/2607.21028", "track": "recommendation", "organization": "Tencent", "published": "2026-07-23", "code_url": null, "topics": [ "generative-recommendation", "semantic-id", "reranking" ], "local_code_dir": "src/auto_research/reproductions/barge", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "同 width、层数、训练行、optimizer 和 140 steps 的 single-path RQ-VAE/TIGER", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "osq.final_loss", "osq.initial_loss", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Tencent commercial media platform", "metric": "CTR", "lift_percent": 0.6, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2607.21028", "source_location": "original paper online-result disclosure: product=Tencent commercial media platform; metric=CTR" }, { "product": "Tencent commercial media platform", "metric": "click UV", "lift_percent": 1.34, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2607.21028", "source_location": "original paper online-result disclosure: product=Tencent commercial media platform; metric=click UV" }, { "product": "Tencent commercial media platform", "metric": "total reading time", "lift_percent": 1.7, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2607.21028", "source_location": "original paper online-result disclosure: product=Tencent commercial media platform; metric=total reading time" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "beque", "title": "Large Language Model based Long-tail Query Rewriting in Taobao Search", "paper_url": "https://arxiv.org/abs/2311.03758", "detail_path": "reproductions/2311.03758-beque/README.md", "topic": [ "llm-recommendation", "query-rewriting", "preference-optimization", "search" ], "first_author": null, "first_author_affiliation": "Alibaba", "published": "2023-11", "code": null, "adapter": { "adapter_key": "beque", "arxiv_id": "2311.03758", "title": "Large Language Model based Long-tail Query Rewriting in Taobao Search", "paper_url": "https://arxiv.org/abs/2311.03758", "track": "recommendation", "organization": "Alibaba", "published": "2023-11", "code_url": null, "topics": [ "llm-recommendation", "query-rewriting", "preference-optimization", "search" ], "local_code_dir": "src/auto_research/reproductions/beque", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens 100K catalog text (public query-rewrite proxy)" ], "baseline": "T5 SFT", "metrics": [ "hit_at_1.pro_mean", "hit_at_1.pro_std", "hit_at_1.relative_gain_percent", "hit_at_1.sft_mean", "hit_at_1.sft_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=40", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Mobile Taobao Search", "metric": "GMV", "lift_percent": 0.4, "traffic": "14-day online test", "source_url": "https://arxiv.org/abs/2311.03758", "source_location": "original paper online-result disclosure: product=Mobile Taobao Search; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "bst", "title": "Behavior Sequence Transformer for E-commerce Recommendation in Alibaba", "paper_url": "https://arxiv.org/abs/1905.06874", "detail_path": "reproductions/1905.06874-bst/README.md", "topic": [ "ranking", "classic", "sequence-modeling" ], "first_author": null, "first_author_affiliation": "Alibaba", "published": "2019-05-15", "code": null, "adapter": { "adapter_key": "bst", "arxiv_id": "1905.06874", "title": "Behavior Sequence Transformer for E-commerce Recommendation in Alibaba", "paper_url": "https://arxiv.org/abs/1905.06874", "track": "recommendation", "organization": "Alibaba", "published": "2019-05-15", "code_url": null, "topics": [ "ranking", "classic", "sequence-modeling" ], "local_code_dir": "src/auto_research/reproductions/bst", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K compact" ], "baseline": "DIN", "metrics": [ "baseline.ndcg_at_10", "method.ndcg_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao recommendation", "metric": "CTR", "lift_percent": 7.57, "traffic": "online A/B vs WDL control", "source_url": "https://arxiv.org/abs/1905.06874", "source_location": "original paper online-result disclosure: product=Taobao recommendation; metric=CTR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "cadet", "title": "CADET: Context-Conditioned Ads CTR Prediction With a Decoder-Only Transformer", "paper_url": "https://arxiv.org/abs/2602.11410", "detail_path": "reproductions/2602.11410-cadet/README.md", "topic": [ "advertising", "ctr-prediction", "decoder-only-transformer" ], "first_author": null, "first_author_affiliation": "LinkedIn", "published": "2026-02-11", "code": null, "adapter": { "adapter_key": "cadet", "arxiv_id": "2602.11410", "title": "CADET: Context-Conditioned Ads CTR Prediction With a Decoder-Only Transformer", "paper_url": "https://arxiv.org/abs/2602.11410", "track": "recommendation", "organization": "LinkedIn", "published": "2026-02-11", "code_url": null, "topics": [ "advertising", "ctr-prediction", "decoder-only-transformer" ], "local_code_dir": "src/auto_research/reproductions/cadet", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "LinkedIn sponsored feed", "metric": "CTR", "lift_percent": 11.04, "traffic": "large-scale online A/B", "source_url": "https://arxiv.org/html/2602.11410v1", "source_location": "Section 5.3 / Table 3", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "camie", "title": "CAMIE: Co-Engagement-Aware Multimodal Item Embeddings for Snap Dynamic Product Ads Retrieval", "paper_url": "https://arxiv.org/abs/2608.30255", "detail_path": "reproductions/2608.30255-camie/README.md", "topic": [ "multimodal-retrieval", "co-engagement", "advertising" ], "first_author": null, "first_author_affiliation": "Snap Inc.", "published": "2026-08-31", "code": null, "adapter": { "adapter_key": "camie", "arxiv_id": "2608.30255", "title": "CAMIE: Co-Engagement-Aware Multimodal Item Embeddings for Snap Dynamic Product Ads Retrieval", "paper_url": "https://arxiv.org/abs/2608.30255", "track": "recommendation", "organization": "Snap Inc.", "published": "2026-08-31", "code_url": null, "topics": [ "multimodal-retrieval", "co-engagement", "advertising" ], "local_code_dir": "src/auto_research/reproductions/camie", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Snap Dynamic Product Ads", "metric": "overall CVR", "lift_percent": 1.911, "traffic": "production deployment on overall DPA traffic", "source_url": "https://arxiv.org/html/2608.30255v1", "source_location": "Abstract", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "context:coengagement-embedding" ] } }, { "domain": "recommendation", "key": "capts", "title": "CAPTS: Channel-Aware, Preference-Aligned Trigger Selection for Multi-Channel Item-to-Item Retrieval", "paper_url": "https://arxiv.org/abs/2602.12564", "detail_path": "reproductions/2602.12564-capts/README.md", "topic": [ "retrieval", "trigger-selection", "multi-channel-routing" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-02-13", "code": null, "adapter": { "adapter_key": "capts", "arxiv_id": "2602.12564", "title": "CAPTS: Channel-Aware, Preference-Aligned Trigger Selection for Multi-Channel Item-to-Item Retrieval", "paper_url": "https://arxiv.org/abs/2602.12564", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-02-13", "code_url": null, "topics": [ "retrieval", "trigger-selection", "multi-channel-routing" ], "local_code_dir": "src/auto_research/reproductions/capts", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kwai multi-channel retrieval", "metric": "Total app time spent", "lift_percent": 0.713, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2602.12564v1", "source_location": "Section 4.4.1", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "context:channel-trigger-routing" ] } }, { "domain": "recommendation", "key": "causal-representation", "title": "Causal Representation Learning for Generalisable Recommendation", "paper_url": "https://arxiv.org/abs/2605.27043", "detail_path": "reproductions/2605.27043-causal-representation/README.md", "topic": [ "causal-recommendation", "distribution-shift", "representation-learning" ], "first_author": null, "first_author_affiliation": "University of Warwick", "published": "2026-05-26", "code": null, "adapter": { "adapter_key": "causal-representation", "arxiv_id": "2605.27043", "title": "Causal Representation Learning for Generalisable Recommendation", "paper_url": "https://arxiv.org/abs/2605.27043", "track": "recommendation", "organization": "University of Warwick", "published": "2026-05-26", "code_url": null, "topics": [ "causal-recommendation", "distribution-shift", "representation-learning" ], "local_code_dir": "src/auto_research/reproductions/causal_representation", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Spotify playlist generation", "metric": "Track streams", "lift_percent": 0.75, "traffic": "2 weeks; millions of users; 95% CI [+0.54%, +0.96%]", "source_url": "https://arxiv.org/html/2605.27043v1", "source_location": "Section 4.3, Table 2", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "head:causal-bottleneck" ] } }, { "domain": "recommendation", "key": "causal-retrieval", "title": "Deep-learning Causal Retrieval Optimization for Efficient e-commerce Distribution in Pinterest", "paper_url": "https://arxiv.org/abs/2607.14161", "detail_path": "reproductions/2607.14161-causal-retrieval/README.md", "topic": [ "causal-retrieval", "uplift", "candidate-generation" ], "first_author": null, "first_author_affiliation": "Pinterest", "published": "2026-07-14", "code": null, "adapter": { "adapter_key": "causal-retrieval", "arxiv_id": "2607.14161", "title": "Deep-learning Causal Retrieval Optimization for Efficient e-commerce Distribution in Pinterest", "paper_url": "https://arxiv.org/abs/2607.14161", "track": "recommendation", "organization": "Pinterest", "published": "2026-07-14", "code_url": null, "topics": [ "causal-retrieval", "uplift", "candidate-generation" ], "local_code_dir": "src/auto_research/reproductions/causal_retrieval", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K with synthetic randomized treatment" ], "baseline": "固定的非因果 transition/content/popularity retrieval ensemble", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Pinterest shopping retrieval", "metric": "total sessions", "lift_percent": 0.26, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2607.14161", "source_location": "original paper online-result disclosure: product=Pinterest shopping retrieval; metric=total sessions" }, { "product": "Pinterest shopping retrieval", "metric": "Pin saves", "lift_percent": 1.1, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2607.14161", "source_location": "original paper online-result disclosure: product=Pinterest shopping retrieval; metric=Pin saves" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "ccformer", "title": "CCFormer: Efficient Cross-Field Interaction and Hierarchical Sequence Compression for Industrial Recommendation at Tencent", "paper_url": "https://arxiv.org/abs/2607.28070", "detail_path": "reproductions/2607.28070-ccformer/README.md", "topic": [ "ranking", "long-sequence", "sequence-compression", "token-mixing" ], "first_author": null, "first_author_affiliation": "Tencent Platform and Content Group", "published": "2026-07-30", "code": null, "adapter": { "adapter_key": "ccformer", "arxiv_id": "2607.28070", "title": "CCFormer: Efficient Cross-Field Interaction and Hierarchical Sequence Compression for Industrial Recommendation at Tencent", "paper_url": "https://arxiv.org/abs/2607.28070", "track": "recommendation", "organization": "Tencent Platform and Content Group", "published": "2026-07-30", "code_url": null, "topics": [ "ranking", "long-sequence", "sequence-compression", "token-mixing" ], "local_code_dir": "src/auto_research/reproductions/ccformer", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M (240 users / 400 items)" ], "baseline": "同预算 GRU", "metrics": [ "baseline.encoded_sequence_tokens", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.encoded_sequence_tokens", "method.hit_at_10", "method.ndcg_at_10", "relative_ndcg_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Tencent video recommendation", "metric": "CTR", "lift_percent": 3.57, "traffic": "over one million exposed users/day, two weeks; fully deployed", "source_url": "https://arxiv.org/abs/2607.28070", "source_location": "original paper online-result disclosure: product=Tencent video recommendation; metric=CTR" }, { "product": "Tencent advertising ranking", "metric": "advertising revenue", "lift_percent": 1.71, "traffic": "over one million exposed users/day, two weeks; fully deployed", "source_url": "https://arxiv.org/abs/2607.28070", "source_location": "original paper online-result disclosure: product=Tencent advertising ranking; metric=advertising revenue" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "cdm", "title": "Contextual Distillation Model for Diversified Recommendation", "paper_url": "https://arxiv.org/abs/2406.09021", "detail_path": "reproductions/2406.09021-cdm/README.md", "topic": [ "reranking", "diversity", "knowledge-distillation" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2024-06-13", "code": null, "adapter": { "adapter_key": "cdm", "arxiv_id": "2406.09021", "title": "Contextual Distillation Model for Diversified Recommendation", "paper_url": "https://arxiv.org/abs/2406.09021", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2024-06-13", "code_url": null, "topics": [ "reranking", "diversity", "knowledge-distillation" ], "local_code_dir": "src/auto_research/reproductions/cdm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享 transition + content scorer，实验组只加入 cdm 核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou", "metric": "Watch Time", "lift_percent": 0.406, "traffic": "main-app A/B", "source_url": "https://arxiv.org/abs/2406.09021", "source_location": "original paper online-result disclosure: product=Kuaishou; metric=Watch Time" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "cgr", "title": "Constraint-Aware Generative Re-ranking for Multi-Objective Optimization in Advertising Feeds", "paper_url": "https://arxiv.org/abs/2603.04227", "detail_path": "reproductions/2603.04227-cgr/README.md", "topic": [ "advertising", "generative-reranking", "constraint-optimization" ], "first_author": null, "first_author_affiliation": "Bilibili", "published": "2026-03-04", "code": null, "adapter": { "adapter_key": "cgr", "arxiv_id": "2603.04227", "title": "Constraint-Aware Generative Re-ranking for Multi-Objective Optimization in Advertising Feeds", "paper_url": "https://arxiv.org/abs/2603.04227", "track": "recommendation", "organization": "Bilibili", "published": "2026-03-04", "code_url": null, "topics": [ "advertising", "generative-reranking", "constraint-optimization" ], "local_code_dir": "src/auto_research/reproductions/cgr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Bilibili advertising feeds", "metric": "Inference latency reduction", "lift_percent": 85.0, "traffic": "millions of daily requests; production online A/B", "source_url": "https://arxiv.org/html/2603.04227v1", "source_location": "Conclusion", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "reward:constraint-aware" ] } }, { "domain": "recommendation", "key": "chronicle-rec", "title": "ChronicleRec: Pre-training Temporally Anchored Tokens for Lifelong User Modeling", "paper_url": "https://arxiv.org/abs/2609.12375", "detail_path": "reproductions/2609.12375-chronicle-rec/README.md", "topic": [ "ranking", "long-sequence", "sequence-compression", "pretraining" ], "first_author": null, "first_author_affiliation": "Tencent", "published": "2026-09-11", "code": null, "adapter": { "adapter_key": "chronicle-rec", "arxiv_id": "2609.12375", "title": "ChronicleRec: Pre-training Temporally Anchored Tokens for Lifelong User Modeling", "paper_url": "https://arxiv.org/abs/2609.12375", "track": "recommendation", "organization": "Tencent", "published": "2026-09-11", "code_url": null, "topics": [ "ranking", "long-sequence", "sequence-compression", "pretraining" ], "local_code_dir": "src/auto_research/reproductions/chronicle_rec", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Weixin Moments Ads pCVR", "metric": "GMV", "lift_percent": 1.61, "traffic": "seven-day production A/B", "source_url": "https://arxiv.org/html/2609.12375v1", "source_location": "Section 4.9", "experiment_duration": "seven days", "significance": "95% CI excludes zero", "retrieved_at": "2026-09-15" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "click-a-buy-b", "title": "Click A, Buy B: Rethinking Conversion Attribution in E-Commerce Recommendations", "paper_url": "https://arxiv.org/abs/2507.15113", "detail_path": "reproductions/2507.15113-click-a-buy-b/README.md", "topic": [ "conversion-attribution", "multi-task-learning", "taxonomy", "ads-ranking" ], "first_author": null, "first_author_affiliation": "Pinterest", "published": "2025-07-20", "code": null, "adapter": { "adapter_key": "click-a-buy-b", "arxiv_id": "2507.15113", "title": "Click A, Buy B: Rethinking Conversion Attribution in E-Commerce Recommendations", "paper_url": "https://arxiv.org/abs/2507.15113", "track": "recommendation", "organization": "Pinterest", "published": "2025-07-20", "code_url": null, "topics": [ "conversion-attribution", "multi-task-learning", "taxonomy", "ads-ranking" ], "local_code_dir": "src/auto_research/reproductions/click_a_buy_b", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "last-click single-item conversion attribution", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.mrr_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.mrr_at_10", "method.ndcg_at_10", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.mrr_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Pinterest shopping", "metric": "primary business metric", "lift_percent": 0.25, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2507.15113", "source_location": "original paper online-result disclosure: product=Pinterest shopping; metric=primary business metric" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "climber-pilot", "title": "Climber-Pilot: A Non-Myopic Generative Recommendation Model Towards Better Instruction-Following", "paper_url": "https://arxiv.org/abs/2602.13581", "detail_path": "reproductions/2602.13581-climber-pilot/README.md", "topic": [ "generative-retrieval", "long-horizon", "instruction-following" ], "first_author": null, "first_author_affiliation": "NetEase Cloud Music", "published": "2026-02-14", "code": null, "adapter": { "adapter_key": "climber-pilot", "arxiv_id": "2602.13581", "title": "Climber-Pilot: A Non-Myopic Generative Recommendation Model Towards Better Instruction-Following", "paper_url": "https://arxiv.org/abs/2602.13581", "track": "recommendation", "organization": "NetEase Cloud Music", "published": "2026-02-14", "code_url": null, "topics": [ "generative-retrieval", "long-horizon", "instruction-following" ], "local_code_dir": "src/auto_research/reproductions/climber_pilot", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "NetEase Cloud Music", "metric": "Core business metric", "lift_percent": 4.24, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2602.13581v1", "source_location": "Abstract", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "context:instruction-foresight" ] } }, { "domain": "recommendation", "key": "clockrope", "title": "ClockRoPE: Random Fourier Rotations for Temporal Routine Modeling", "paper_url": "https://arxiv.org/abs/2607.26369", "detail_path": "reproductions/2607.26369-clockrope/README.md", "topic": [ "generative-retrieval", "temporal-modeling", "rope", "long-sequence" ], "first_author": null, "first_author_affiliation": "YouTube / Google DeepMind", "published": "2026-07-29", "code": null, "adapter": { "adapter_key": "clockrope", "arxiv_id": "2607.26369", "title": "ClockRoPE: Random Fourier Rotations for Temporal Routine Modeling", "paper_url": "https://arxiv.org/abs/2607.26369", "track": "recommendation", "organization": "YouTube / Google DeepMind", "published": "2026-07-29", "code_url": null, "topics": [ "generative-retrieval", "temporal-modeling", "rope", "long-sequence" ], "local_code_dir": "src/auto_research/reproductions/clockrope", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "RoPE-style recency ranker", "metrics": [ "Hit@10", "NDCG@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "major video platform", "metric": "engagement", "lift_percent": 0.08, "traffic": "14-day A/B; 1% traffic per variant", "source_url": "https://arxiv.org/html/2607.26369v2#S4.SS2", "source_location": "Section 4.2 Table 6", "experiment_duration": "14 days", "retrieved_at": "2026-08-24" }, { "product": "major video platform", "metric": "valued engagement", "lift_percent": 0.08, "traffic": "14-day A/B; 1% traffic per variant", "source_url": "https://arxiv.org/html/2607.26369v2#S4.SS2", "source_location": "Section 4.2 Table 6", "experiment_duration": "14 days", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "cluster-goobs", "title": "Real-Time Hard Negative Sampling via LLM-based Clustering for Large-Scale Two-Tower Retrieval", "paper_url": "https://arxiv.org/abs/2607.00448", "detail_path": "reproductions/2607.00448-cluster-goobs/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "cluster-goobs", "arxiv_id": "2607.00448", "title": "Real-Time Hard Negative Sampling via LLM-based Clustering for Large-Scale Two-Tower Retrieval", "paper_url": "https://arxiv.org/abs/2607.00448", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/cluster_goobs", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "Random OOB negative sampler", "metrics": [ "relative_ndcg_change_percent", "results.cluster_goobs.hit_at_10", "results.cluster_goobs.hit_at_10_std", "results.cluster_goobs.ndcg_at_10", "results.cluster_goobs.ndcg_at_10_std", "results.random_oob.hit_at_10", "results.random_oob.hit_at_10_std", "results.random_oob.ndcg_at_10", "results.random_oob.ndcg_at_10_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "cmsl", "title": "CMSL: Constructive Multi-Sequence Learning for Recommendation Systems", "paper_url": "https://arxiv.org/abs/2606.28533", "detail_path": "reproductions/2606.28533-cmsl/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "cmsl", "arxiv_id": "2606.28533", "title": "CMSL: Constructive Multi-Sequence Learning for Recommendation Systems", "paper_url": "https://arxiv.org/abs/2606.28533", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/cmsl", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "single sequence hstu", "metrics": [ "relative_ndcg_change_percent", "results.cmsl.hit_at_10", "results.cmsl.hit_at_10_std", "results.cmsl.ndcg_at_10", "results.cmsl.ndcg_at_10_std", "results.single_sequence_hstu.hit_at_10", "results.single_sequence_hstu.hit_at_10_std", "results.single_sequence_hstu.ndcg_at_10", "results.single_sequence_hstu.ndcg_at_10_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=90", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "cobra", "title": "Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations", "paper_url": "https://arxiv.org/abs/2503.02453", "detail_path": "reproductions/2503.02453-cobra/README.md", "topic": [ "generative-retrieval", "semantic-id", "dense-retrieval" ], "first_author": null, "first_author_affiliation": "Baidu", "published": "2025-03", "code": null, "adapter": { "adapter_key": "cobra", "arxiv_id": "2503.02453", "title": "Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations", "paper_url": "https://arxiv.org/abs/2503.02453", "track": "recommendation", "organization": "Baidu", "published": "2025-03", "code_url": null, "topics": [ "generative-retrieval", "semantic-id", "dense-retrieval" ], "local_code_dir": "src/auto_research/reproductions/cobra", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens-100K" ], "baseline": "DIN，COBRA NDCG@10 相对 DIN +25.75%", "metrics": [ "results.cobra_cascade.hit_at_10", "results.cobra_cascade.ndcg_at_10", "results.din.hit_at_10", "results.din.ndcg_at_10", "results.ndcg_ablation_percent", "results.ndcg_vs_din_percent", "results.sparse_only.hit_at_10", "results.sparse_only.ndcg_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=120", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Baidu advertising", "metric": "conversion", "lift_percent": 3.6, "traffic": "production online A/B", "source_url": "https://arxiv.org/abs/2503.02453", "source_location": "original paper online-result disclosure: product=Baidu advertising; metric=conversion" }, { "product": "Baidu advertising", "metric": "ARPU", "lift_percent": 4.15, "traffic": "production online A/B", "source_url": "https://arxiv.org/abs/2503.02453", "source_location": "original paper online-result disclosure: product=Baidu advertising; metric=ARPU" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "connectionmind", "title": "ConnectionMind: A General Social-Personalized Recommendation System with LLM Reasoning", "paper_url": "https://arxiv.org/abs/2608.10187", "detail_path": "reproductions/2608.10187-connectionmind/README.md", "topic": [ "social-recommendation", "llm-recommendation", "graph-reasoning", "reinforcement-learning", "distillation" ], "first_author": "Haoyu Han", "first_author_affiliation": "Michigan State University", "published": "2026-08-10", "code": null, "adapter": { "adapter_key": "connectionmind", "arxiv_id": "2608.10187", "title": "ConnectionMind: A General Social-Personalized Recommendation System with LLM Reasoning", "paper_url": "https://arxiv.org/abs/2608.10187", "track": "recommendation", "organization": "Michigan State University / Meta Platforms, Inc.", "published": "2026-08-10", "code_url": null, "topics": [ "social-recommendation", "llm-recommendation", "graph-reasoning", "reinforcement-learning", "distillation" ], "local_code_dir": "src/auto_research/reproductions/connectionmind", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "HetRec 2011 Delicious-2K" ], "baseline": "fixed heterogeneous graph aggregation", "metrics": [ "Recall@10", "Precision@10", "NDCG@10" ], "default_seeds": [ 42 ], "budget": "3 SFT epochs + 180 GRPO groups", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta short-video recommendation", "metric": "exposure", "lift_percent": 0.33, "traffic": "multi-week test over tens of millions of users", "source_url": "https://arxiv.org/pdf/2608.10187", "source_location": "Section 5.4 Table 2", "experiment_duration": "multiple weeks", "significance": "±0.08%", "retrieved_at": "2026-08-20" }, { "product": "Meta short-video recommendation", "metric": "watch time", "lift_percent": 0.43, "traffic": "multi-week test over tens of millions of users", "source_url": "https://arxiv.org/pdf/2608.10187", "source_location": "Section 5.4 Table 2", "experiment_duration": "multiple weeks", "significance": "±0.14%", "retrieved_at": "2026-08-20" }, { "product": "Meta short-video recommendation", "metric": "video sessions", "lift_percent": 0.22, "traffic": "multi-week test over tens of millions of users", "source_url": "https://arxiv.org/pdf/2608.10187", "source_location": "Section 5.4 Table 2", "experiment_duration": "multiple weeks", "significance": "±0.13%", "retrieved_at": "2026-08-20" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "coral", "title": "CORAL: An LLM-Native Harness for Production Recommender Systems", "paper_url": "https://arxiv.org/abs/2609.02730", "detail_path": "reproductions/2609.02730-coral/README.md", "topic": [ "auto-research", "agent", "retrieval-budget", "constrained-optimization" ], "first_author": null, "first_author_affiliation": "Meta AI", "published": "2026-09-02", "code": null, "adapter": { "adapter_key": "coral", "arxiv_id": "2609.02730", "title": "CORAL: An LLM-Native Harness for Production Recommender Systems", "paper_url": "https://arxiv.org/abs/2609.02730", "track": "recommendation", "organization": "Meta AI", "published": "2026-09-02", "code_url": null, "topics": [ "auto-research", "agent", "retrieval-budget", "constrained-optimization" ], "local_code_dir": "src/auto_research/reproductions/coral", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "fixed equal retrieval allocation", "metrics": [ "hit_at_10", "ndcg_at_10", "budget feasibility", "cycle effect" ], "default_seeds": [ 42, 43, 44 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta large-scale video service", "metric": "video-viewing sessions", "lift_percent": 0.16, "traffic": "millions-user online A/B; deployed configuration", "source_url": "https://arxiv.org/pdf/2609.02730v1", "source_location": "Section 5.1, Table 1", "retrieved_at": "2026-09-06" }, { "product": "Meta serving-capacity service", "metric": "annualized capacity savings", "lift_percent": 44.0, "traffic": "second A/B round increased first-round savings by 44%; engagement neutral", "source_url": "https://arxiv.org/pdf/2609.02730v1", "source_location": "Section 5.2", "retrieved_at": "2026-09-06" } ], "selection_exception": null, "evolve_operators": [ "controller:coral-constrained-feedback-loop" ] } }, { "domain": "recommendation", "key": "core-relevance", "title": "CORE: A Unified Cascaded Ordinal Relevance Estimation Framework for E-commerce Search", "paper_url": "https://arxiv.org/abs/2607.24417", "detail_path": "reproductions/2607.24417-core-relevance/README.md", "topic": [ "search-ranking", "ordinal-relevance", "step-grpo", "llm-distillation" ], "first_author": null, "first_author_affiliation": "Meituan / Beijing Institute of Technology", "published": "2026-07-27", "code": null, "adapter": { "adapter_key": "core-relevance", "arxiv_id": "2607.24417", "title": "CORE: A Unified Cascaded Ordinal Relevance Estimation Framework for E-commerce Search", "paper_url": "https://arxiv.org/abs/2607.24417", "track": "recommendation", "organization": "Meituan / Beijing Institute of Technology", "published": "2026-07-27", "code_url": null, "topics": [ "search-ranking", "ordinal-relevance", "step-grpo", "llm-distillation" ], "local_code_dir": "src/auto_research/reproductions/core_relevance", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M ordinal relevance" ], "baseline": "MovieLens-1M 构造的同一 Low/Mid/High query-item 集上的 flat 三分类", "metrics": [ "paper_results.cascaded_bert_accuracy", "paper_results.direct_bert_accuracy", "paper_results.distilled_accuracy", "paper_results.online_ndcg_at_5_percent", "paper_results.postcot_core_accuracy", "paper_results.step_grpo_accuracy", "relative.accuracy_points", "relative.badcase_reduction_percent", "relative.ndcg_at_5_percent", "training.cascade_sft.final_loss", "training.cascade_sft.initial_loss", "training.flat.final_loss", "training.flat.initial_loss", "training.postcot_distillation.final_loss", "training.postcot_distillation.initial_loss", "training.step_grpo.final_policy_loss", "training.step_grpo.mean_step_reward", "variants.Cascaded + step-GRPO.accuracy", "variants.Cascaded + step-GRPO.ndcg_at_5", "variants.Cascaded classifier.accuracy", "variants.Cascaded classifier.ndcg_at_5", "variants.Flat classifier.accuracy", "variants.Flat classifier.ndcg_at_5", "variants.PostCoT distilled cascade.accuracy", "variants.PostCoT distilled cascade.ndcg_at_5" ], "default_seeds": [ 42 ], "budget": "steps=100", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meituan e-commerce search", "metric": "NDCG@5", "lift_percent": 0.2, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2607.24417", "source_location": "original paper online-result disclosure: product=Meituan e-commerce search; metric=NDCG@5" }, { "product": "Meituan e-commerce search", "metric": "Badcase@5 reduction", "lift_percent": 15.9, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2607.24417", "source_location": "original paper online-result disclosure: product=Meituan e-commerce search; metric=Badcase@5 reduction" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "cq-sid", "title": "Efficient Generative Retrieval for E-commerce Search with Semantic Cluster IDs and Expert-Guided RL", "paper_url": "https://arxiv.org/abs/2605.14434", "detail_path": "reproductions/2605.14434-cq-sid/README.md", "topic": [ "generative-retrieval", "semantic-id", "reinforcement-learning" ], "first_author": null, "first_author_affiliation": "Alibaba Taobao & Tmall Group", "published": "2026-05-14", "code": null, "adapter": { "adapter_key": "cq-sid", "arxiv_id": "2605.14434", "title": "Efficient Generative Retrieval for E-commerce Search with Semantic Cluster IDs and Expert-Guided RL", "paper_url": "https://arxiv.org/abs/2605.14434", "track": "recommendation", "organization": "Alibaba Taobao & Tmall Group", "published": "2026-05-14", "code_url": null, "topics": [ "generative-retrieval", "semantic-id", "reinforcement-learning" ], "local_code_dir": "src/auto_research/reproductions/cq_sid", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K compact" ], "baseline": "协同 item-ID 检索", "metrics": [ "baseline.ndcg_at_10", "method.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "TmallAPP Search", "metric": "GMV", "lift_percent": 1.15, "traffic": "two-week online A/B", "source_url": "https://arxiv.org/abs/2605.14434", "source_location": "original paper online-result disclosure: product=TmallAPP Search; metric=GMV" }, { "product": "TmallAPP Search", "metric": "UCTCVR", "lift_percent": 0.4, "traffic": "two-week online A/B", "source_url": "https://arxiv.org/abs/2605.14434", "source_location": "original paper online-result disclosure: product=TmallAPP Search; metric=UCTCVR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "cross-domain-kd", "title": "Zero-shot Cross-domain Knowledge Distillation: A Case study on YouTube Music", "paper_url": "https://arxiv.org/abs/2603.28994", "detail_path": "reproductions/2603.28994-cross-domain-kd/README.md", "topic": [ "distillation", "cross-domain", "ranking" ], "first_author": null, "first_author_affiliation": "Google / YouTube", "published": "2026-03", "code": null, "adapter": { "adapter_key": "cross-domain-kd", "arxiv_id": "2603.28994", "title": "Zero-shot Cross-domain Knowledge Distillation: A Case study on YouTube Music", "paper_url": "https://arxiv.org/abs/2603.28994", "track": "recommendation", "organization": "Google / YouTube", "published": "2026-03", "code_url": null, "topics": [ "distillation", "cross-domain", "ranking" ], "local_code_dir": "src/auto_research/reproductions/cross_domain_kd", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K source/target domains" ], "baseline": "在相同 target split 上训练的 DIN，实验组 KD student 的 NDCG@10 相对 DIN -68.46%", "metrics": [ "results.din.hit_at_10", "results.din.ndcg_at_10", "results.ndcg_ablation_percent", "results.ndcg_vs_din_percent", "results.target_only.hit_at_10", "results.target_only.ndcg_at_10", "results.zero_shot_kd.hit_at_10", "results.zero_shot_kd.ndcg_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=100", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "YouTube Music", "metric": "discovery", "lift_percent": 1.12, "traffic": "two-week live experiments", "source_url": "https://arxiv.org/abs/2603.28994", "source_location": "original paper online-result disclosure: product=YouTube Music; metric=discovery" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "crsd", "title": "From Reasoning LLMs to BERT: A Two-Stage Distillation Framework for Search Relevance", "paper_url": "https://arxiv.org/abs/2510.11056", "detail_path": "reproductions/2510.11056-crsd/README.md", "topic": [ "search", "content-understanding", "llm-distillation" ], "first_author": null, "first_author_affiliation": "Meituan", "published": "2025-10-13", "code": null, "adapter": { "adapter_key": "crsd", "arxiv_id": "2510.11056", "title": "From Reasoning LLMs to BERT: A Two-Stage Distillation Framework for Search Relevance", "paper_url": "https://arxiv.org/abs/2510.11056", "track": "recommendation", "organization": "Meituan", "published": "2025-10-13", "code_url": null, "topics": [ "search", "content-understanding", "llm-distillation" ], "local_code_dir": "src/auto_research/reproductions/crsd", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "shared transition + content baseline，实验组为 CRSD，只改变论文核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "paper_results.adctr_percent", "paper_results.adcvr_percent", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meituan search advertising", "metric": "AdCTR", "lift_percent": 0.91, "traffic": "30% traffic", "source_url": "https://arxiv.org/abs/2510.11056", "source_location": "original paper online-result disclosure: product=Meituan search advertising; metric=AdCTR" }, { "product": "Meituan search advertising", "metric": "AdCVR", "lift_percent": 1.06, "traffic": "30% traffic", "source_url": "https://arxiv.org/abs/2510.11056", "source_location": "original paper online-result disclosure: product=Meituan search advertising; metric=AdCVR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "cs3", "title": "CS3: Efficient Online Capability Synergy for Two-Tower Recommendation", "paper_url": "https://arxiv.org/abs/2604.19269", "detail_path": "reproductions/2604.19269-cs3/README.md", "topic": [ "retrieval", "two-tower", "online-learning" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-04-21", "code": "https://github.com/lixiangwang/CS3Rec", "adapter": { "adapter_key": "cs3", "arxiv_id": "2604.19269", "title": "CS3: Efficient Online Capability Synergy for Two-Tower Recommendation", "paper_url": "https://arxiv.org/abs/2604.19269", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-04-21", "code_url": "https://github.com/lixiangwang/CS3Rec", "topics": [ "retrieval", "two-tower", "online-learning" ], "local_code_dir": "src/auto_research/reproductions/cs3", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K compact" ], "baseline": "matched two-tower", "metrics": [ "baseline.ndcg_at_10", "method.ndcg_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou advertising", "metric": "Revenue", "lift_percent": 8.356, "traffic": "Scenario A; three production scenarios", "source_url": "https://arxiv.org/abs/2604.19269", "source_location": "original paper online-result disclosure: product=Kuaishou advertising; metric=Revenue" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "cwm", "title": "Counteracting Duration Bias in Video Recommendation via Counterfactual Watch Time", "paper_url": "https://arxiv.org/abs/2406.07932", "detail_path": "reproductions/2406.07932-cwm/README.md", "topic": [ "ranking", "counterfactual-learning", "watch-time", "bias" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology / Renmin University of China", "published": "2024-06-12", "code": "https://github.com/hyz20/CWM", "adapter": { "adapter_key": "cwm", "arxiv_id": "2406.07932", "title": "Counteracting Duration Bias in Video Recommendation via Counterfactual Watch Time", "paper_url": "https://arxiv.org/abs/2406.07932", "track": "recommendation", "organization": "Kuaishou Technology / Renmin University of China", "published": "2024-06-12", "code_url": "https://github.com/hyz20/CWM", "topics": [ "ranking", "counterfactual-learning", "watch-time", "bias" ], "local_code_dir": "src/auto_research/reproductions/cwm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享 transition + content scorer，实验组只加入 cwm 核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "paper_results.ctr_percent", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.cost_transform_to_interest", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou", "metric": "Mean Watch Time", "lift_percent": 2.9, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2406.07932", "source_location": "original paper online-result disclosure: product=Kuaishou; metric=Mean Watch Time" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "dadf", "title": "DADF: A Distribution-Aware Debiasing Framework for Watch-Time Regression in Recommender Systems", "paper_url": "https://arxiv.org/abs/2605.17863", "detail_path": "reproductions/2605.17863-dadf/README.md", "topic": [ "ranking", "watch-time", "debiasing" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-05-18", "code": "https://github.com/liuzhao09/DADF", "adapter": { "adapter_key": "dadf", "arxiv_id": "2605.17863", "title": "DADF: A Distribution-Aware Debiasing Framework for Watch-Time Regression in Recommender Systems", "paper_url": "https://arxiv.org/abs/2605.17863", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-05-18", "code_url": "https://github.com/liuzhao09/DADF", "topics": [ "ranking", "watch-time", "debiasing" ], "local_code_dir": "src/auto_research/reproductions/dadf", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou full ranking", "metric": "average time spent per device", "lift_percent": 0.649, "traffic": "7-day A/B, then 100% traffic", "source_url": "https://arxiv.org/html/2605.17863v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "danet", "title": "Cheaper is Better: A Discount-Aware Network for Conversion Rate Prediction in E-commerce Recommendation System", "paper_url": "https://arxiv.org/abs/2607.12578", "detail_path": "reproductions/2607.12578-danet/README.md", "topic": [ "ranking", "cvr", "discount" ], "first_author": null, "first_author_affiliation": "Alibaba Group / Tmall", "published": "2026-07-14", "code": "https://github.com/tangrc/DANet", "adapter": { "adapter_key": "danet", "arxiv_id": "2607.12578", "title": "Cheaper is Better: A Discount-Aware Network for Conversion Rate Prediction in E-commerce Recommendation System", "paper_url": "https://arxiv.org/abs/2607.12578", "track": "recommendation", "organization": "Alibaba Group / Tmall", "published": "2026-07-14", "code_url": "https://github.com/tangrc/DANet", "topics": [ "ranking", "cvr", "discount" ], "local_code_dir": "src/auto_research/reproductions/danet", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "movielens 100k" ], "baseline": "IPN interest score", "metrics": [ "baseline.fresh_hit_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Tmall", "metric": "pCVR", "lift_percent": 3.63, "traffic": "online A/B test", "source_url": "https://arxiv.org/abs/2607.12578", "source_location": "original paper online-result disclosure: product=Tmall; metric=pCVR" }, { "product": "Tmall", "metric": "GMV", "lift_percent": 2.23, "traffic": "online A/B test", "source_url": "https://arxiv.org/abs/2607.12578", "source_location": "original paper online-result disclosure: product=Tmall; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "dceo", "title": "DCEO: Direct Causal Effect Optimization for Long-Term User Value Modeling in E-commerce Search", "paper_url": "https://arxiv.org/abs/2608.25635", "detail_path": "reproductions/2608.25635-dceo/README.md", "topic": [ "search-ranking", "long-term-value", "causal-optimization", "multi-objective-ranking" ], "first_author": null, "first_author_affiliation": "Taobao & Tmall Group, Alibaba", "published": "2026-08-26", "code": null, "adapter": { "adapter_key": "dceo", "arxiv_id": "2608.25635", "title": "DCEO: Direct Causal Effect Optimization for Long-Term User Value Modeling in E-commerce Search", "paper_url": "https://arxiv.org/abs/2608.25635", "track": "recommendation", "organization": "Taobao & Tmall Group, Alibaba", "published": "2026-08-26", "code_url": null, "topics": [ "search-ranking", "long-term-value", "causal-optimization", "multi-objective-ranking" ], "local_code_dir": "src/auto_research/reproductions/dceo", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "fixed multi-objective log fusion over the same four proxy scores", "metrics": [ "Hit@10", "NDCG@10", "head share@10", "estimated relative causal effect" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Alibaba e-commerce search", "metric": "GMV", "lift_percent": 0.36, "traffic": "41-day production A/B test", "source_url": "https://arxiv.org/html/2608.25635v1#S5.SS7", "source_location": "Section 5.7", "experiment_duration": "41 days", "retrieved_at": "2026-08-28" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "dcn-v2", "title": "DCN V2: Improved Deep & Cross Network and Practical Lessons for Web-scale Learning to Rank Systems", "paper_url": "https://arxiv.org/abs/2008.13535", "detail_path": "reproductions/2008.13535-dcn-v2/README.md", "topic": [ "ranking", "classic", "feature-interaction" ], "first_author": null, "first_author_affiliation": "Google", "published": "2020-08-19", "code": null, "adapter": { "adapter_key": "dcn-v2", "arxiv_id": "2008.13535", "title": "DCN V2: Improved Deep & Cross Network and Practical Lessons for Web-scale Learning to Rank Systems", "paper_url": "https://arxiv.org/abs/2008.13535", "track": "recommendation", "organization": "Google", "published": "2020-08-19", "code_url": null, "topics": [ "ranking", "classic", "feature-interaction" ], "local_code_dir": "src/auto_research/reproductions/dcn_v2", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K compact" ], "baseline": "deep-only", "metrics": [ "baseline.ndcg_at_10", "method.ndcg_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": "用户明确要求补齐经典搜广推论文；原文确认 Google 线上业务指标显著提升但未披露具体 lift。", "evolve_operators": [] } }, { "domain": "recommendation", "key": "deepfm", "title": "DeepFM: A Factorization-Machine based Neural Network for CTR Prediction", "paper_url": "https://arxiv.org/abs/1703.04247", "detail_path": "reproductions/1703.04247-deepfm/README.md", "topic": [ "ranking", "classic", "feature-interaction" ], "first_author": null, "first_author_affiliation": "Huawei Noah's Ark Lab", "published": "2017-03-13", "code": null, "adapter": { "adapter_key": "deepfm", "arxiv_id": "1703.04247", "title": "DeepFM: A Factorization-Machine based Neural Network for CTR Prediction", "paper_url": "https://arxiv.org/abs/1703.04247", "track": "recommendation", "organization": "Huawei Noah's Ark Lab", "published": "2017-03-13", "code_url": null, "topics": [ "ranking", "classic", "feature-interaction" ], "local_code_dir": "src/auto_research/reproductions/deepfm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K compact" ], "baseline": "embedding MLP，实验组为共享 embedding 的 FM + deep 分支", "metrics": [ "baseline.head_share_at_10", "baseline.head_share_at_10_std", "baseline.hit_at_10", "baseline.hit_at_10_std", "baseline.ndcg_at_10", "baseline.ndcg_at_10_std", "method.head_share_at_10", "method.head_share_at_10_std", "method.hit_at_10", "method.hit_at_10_std", "method.ndcg_at_10", "method.ndcg_at_10_std", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=80", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": "用户明确批准 DeepFM 为经典例外；不以该例外放宽新工业论文的量化线上证据门槛。", "evolve_operators": [] } }, { "domain": "recommendation", "key": "degr", "title": "DEGR: Dual Exploration-Driven Generative Re-Ranking for Adaptive Cross-Request Context Bridging", "paper_url": "https://arxiv.org/abs/2608.04809", "detail_path": "reproductions/2608.04809-degr/README.md", "topic": [ "reranking", "slate-optimization", "exploration", "preference-optimization" ], "first_author": null, "first_author_affiliation": "JD.com", "published": "2026-08-05", "code": null, "adapter": { "adapter_key": "degr", "arxiv_id": "2608.04809", "title": "DEGR: Dual Exploration-Driven Generative Re-Ranking for Adaptive Cross-Request Context Bridging", "paper_url": "https://arxiv.org/abs/2608.04809", "track": "recommendation", "organization": "JD.com", "published": "2026-08-05", "code_url": null, "topics": [ "reranking", "slate-optimization", "exploration", "preference-optimization" ], "local_code_dir": "src/auto_research/reproductions/degr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "同预算 CE generator，实验组加入 diversity 与 adaptive reward ORPO", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "degr.adaptive_reward_orpo_updates", "degr.hit_at_10", "degr.ndcg_at_10", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "JD homepage recommendation", "metric": "UCTR", "lift_percent": 1.22, "traffic": "online A/B test", "source_url": "https://arxiv.org/abs/2608.04809", "source_location": "original paper online-result disclosure: product=JD homepage recommendation; metric=UCTR" }, { "product": "JD homepage recommendation", "metric": "PV", "lift_percent": 0.2, "traffic": "online A/B test", "source_url": "https://arxiv.org/abs/2608.04809", "source_location": "original paper online-result disclosure: product=JD homepage recommendation; metric=PV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "degre", "title": "DeGRe: Listwise Generative Reranking with Offline Lookahead Distillation", "paper_url": "https://arxiv.org/abs/2605.25749", "detail_path": "reproductions/2605.25749-degre/README.md", "topic": [ "generative-recommendation", "reranking", "distillation" ], "first_author": null, "first_author_affiliation": "Alibaba Group / Zhejiang University", "published": "2026-05-25", "code": null, "adapter": { "adapter_key": "degre", "arxiv_id": "2605.25749", "title": "DeGRe: Listwise Generative Reranking with Offline Lookahead Distillation", "paper_url": "https://arxiv.org/abs/2605.25749", "track": "recommendation", "organization": "Alibaba Group / Zhejiang University", "published": "2026-05-25", "code_url": null, "topics": [ "generative-recommendation", "reranking", "distillation" ], "local_code_dir": "src/auto_research/reproductions/degre", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "movielens 100k" ], "baseline": "相同候选与特征的 pointwise generator", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao Flash Shopping", "metric": "CTR", "lift_percent": 2.85, "traffic": "8 days, 2% traffic", "source_url": "https://arxiv.org/abs/2605.25749", "source_location": "original paper online-result disclosure: product=Taobao Flash Shopping; metric=CTR" }, { "product": "Taobao Flash Shopping", "metric": "GMV", "lift_percent": 3.75, "traffic": "8 days, 2% traffic", "source_url": "https://arxiv.org/abs/2605.25749", "source_location": "original paper online-result disclosure: product=Taobao Flash Shopping; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "dien", "title": "Deep Interest Evolution Network for Click-Through Rate Prediction", "paper_url": "https://arxiv.org/abs/1809.03672", "detail_path": "reproductions/1809.03672-dien/README.md", "topic": [ "ranking", "classic", "sequence-modeling" ], "first_author": null, "first_author_affiliation": "Alibaba", "published": "2018-09-11", "code": "https://github.com/mouna99/dien", "adapter": { "adapter_key": "dien", "arxiv_id": "1809.03672", "title": "Deep Interest Evolution Network for Click-Through Rate Prediction", "paper_url": "https://arxiv.org/abs/1809.03672", "track": "recommendation", "organization": "Alibaba", "published": "2018-09-11", "code_url": "https://github.com/mouna99/dien", "topics": [ "ranking", "classic", "sequence-modeling" ], "local_code_dir": "src/auto_research/reproductions/dien", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K compact" ], "baseline": "同一训练协议的 DIN", "metrics": [ "baseline.ndcg_at_10", "method.ndcg_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao display advertising", "metric": "CTR", "lift_percent": 20.7, "traffic": "2018-06-07 to 2018-07-12", "source_url": "https://arxiv.org/abs/1809.03672", "source_location": "original paper online-result disclosure: product=Taobao display advertising; metric=CTR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "diffureason", "title": "DiffuReason: Bridging Latent Reasoning and Generative Refinement for Sequential Recommendation", "paper_url": "https://arxiv.org/abs/2602.09744", "detail_path": "reproductions/2602.09744-diffureason/README.md", "topic": [ "generative-recommendation", "diffusion", "reinforcement-learning" ], "first_author": null, "first_author_affiliation": "Tencent", "published": "2026-02-10", "code": null, "adapter": { "adapter_key": "diffureason", "arxiv_id": "2602.09744", "title": "DiffuReason: Bridging Latent Reasoning and Generative Refinement for Sequential Recommendation", "paper_url": "https://arxiv.org/abs/2602.09744", "track": "recommendation", "organization": "Tencent", "published": "2026-02-10", "code_url": null, "topics": [ "generative-recommendation", "diffusion", "reinforcement-learning" ], "local_code_dir": "src/auto_research/reproductions/diffureason", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Tencent Ads / WeChat Channels", "metric": "ad consumption", "lift_percent": 1.1462, "traffic": "20% traffic, five days", "source_url": "https://arxiv.org/html/2602.09744v1", "source_location": "Section 4.7", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "din", "title": "Deep Interest Network for Click-Through Rate Prediction", "paper_url": "https://arxiv.org/abs/1706.06978", "detail_path": "reproductions/1706.06978-din/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": "https://github.com/zhougr1993/DeepInterestNetwork", "adapter": { "adapter_key": "din", "arxiv_id": "1706.06978", "title": "Deep Interest Network for Click-Through Rate Prediction", "paper_url": "https://arxiv.org/abs/1706.06978", "track": "recommendation", "organization": null, "published": null, "code_url": "https://github.com/zhougr1993/DeepInterestNetwork", "topics": [], "local_code_dir": "src/auto_research/reproductions/din", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "移除 candidate-aware attention、改用 mean pooling 的同参数模型", "metrics": [ "din.hit_at_10", "din.hit_at_10_std", "din.ndcg_at_10", "din.ndcg_at_10_std", "mean_pool.hit_at_10", "mean_pool.hit_at_10_std", "mean_pool.ndcg_at_10", "mean_pool.ndcg_at_10_std", "ndcg_gain_percent" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=320", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "director", "title": "DIRECTOR: Dynamic Index-based Recommendation with Transport-Optimized Retrieval", "paper_url": "https://arxiv.org/abs/2607.26418", "detail_path": "reproductions/2607.26418-director/README.md", "topic": [ "reranking", "dynamic-index", "optimal-transport", "non-autoregressive" ], "first_author": null, "first_author_affiliation": "University of Science and Technology of China", "published": "2026-07-29", "code": null, "adapter": { "adapter_key": "director", "arxiv_id": "2607.26418", "title": "DIRECTOR: Dynamic Index-based Recommendation with Transport-Optimized Retrieval", "paper_url": "https://arxiv.org/abs/2607.26418", "track": "recommendation", "organization": "University of Science and Technology of China", "published": "2026-07-29", "code_url": null, "topics": [ "reranking", "dynamic-index", "optimal-transport", "non-autoregressive" ], "local_code_dir": "src/auto_research/reproductions/director", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "independent position-wise dynamic-index retrieval", "metrics": [ "hit_at_10", "ndcg_at_10", "matched_duplicate_count", "transport_row_error" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; six parallel positions", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou main application", "metric": "Valid View", "lift_percent": 0.519, "traffic": "7-day test; two mutually exclusive 10% buckets", "source_url": "https://arxiv.org/html/2607.26418v1", "source_location": "Section 6.2 and Appendix online protocol, Table 4", "significance": "p < 0.05; 95% CI [0.45%, 0.59%]", "retrieved_at": "2026-09-01" } ], "selection_exception": null, "evolve_operators": [ "head:transport-index" ] } }, { "domain": "recommendation", "key": "dme", "title": "Douyin Multimodal Embedding Model Technical Report", "paper_url": "https://arxiv.org/abs/2608.02148", "detail_path": "reproductions/2608.02148-dme/README.md", "topic": [ "content-understanding", "multimodal-retrieval", "embedding", "llm-recommendation" ], "first_author": null, "first_author_affiliation": "ByteDance / Douyin", "published": "2026-08-03", "code": null, "adapter": { "adapter_key": "dme", "arxiv_id": "2608.02148", "title": "Douyin Multimodal Embedding Model Technical Report", "paper_url": "https://arxiv.org/abs/2608.02148", "track": "recommendation", "organization": "ByteDance / Douyin", "published": "2026-08-03", "code_url": null, "topics": [ "content-understanding", "multimodal-retrieval", "embedding", "llm-recommendation" ], "local_code_dir": "src/auto_research/reproductions/dme", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享 transition + content scorer，实验组只加入 DME 核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Douyin Search", "metric": "Lifetime", "lift_percent": 0.1, "traffic": "online A/B test", "source_url": "https://arxiv.org/abs/2608.02148", "source_location": "original paper online-result disclosure: product=Douyin Search; metric=Lifetime" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "dos", "title": "DOS: Dual-Flow Orthogonal Semantic IDs for Recommendation in Meituan", "paper_url": "https://arxiv.org/abs/2602.04460", "detail_path": "reproductions/2602.04460-dos/README.md", "topic": [ "generative-recommendation", "semantic-id", "orthogonal-quantization" ], "first_author": null, "first_author_affiliation": "Meituan", "published": "2026-02-04", "code": null, "adapter": { "adapter_key": "dos", "arxiv_id": "2602.04460", "title": "DOS: Dual-Flow Orthogonal Semantic IDs for Recommendation in Meituan", "paper_url": "https://arxiv.org/abs/2602.04460", "track": "recommendation", "organization": "Meituan", "published": "2026-02-04", "code_url": null, "topics": [ "generative-recommendation", "semantic-id", "orthogonal-quantization" ], "local_code_dir": "src/auto_research/reproductions/dos", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "未旋转的 content/collaborative ensemble", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meituan app", "metric": "revenue", "lift_percent": 1.15, "traffic": "30% traffic, one week", "source_url": "https://arxiv.org/abs/2602.04460", "source_location": "original paper online-result disclosure: product=Meituan app; metric=revenue" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "downstream-rewards", "title": "Long-term User Engagement Optimization through Model-agnostic Downstream Rewards Learning", "paper_url": "https://arxiv.org/abs/2607.14192", "detail_path": "reproductions/2607.14192-downstream-rewards/README.md", "topic": [ "long-term-value", "reward-design", "multi-surface", "retention" ], "first_author": null, "first_author_affiliation": "Pinterest", "published": "2026-07-15", "code": null, "adapter": { "adapter_key": "downstream-rewards", "arxiv_id": "2607.14192", "title": "Long-term User Engagement Optimization through Model-agnostic Downstream Rewards Learning", "paper_url": "https://arxiv.org/abs/2607.14192", "track": "recommendation", "organization": "Pinterest", "published": "2026-07-15", "code_url": null, "topics": [ "long-term-value", "reward-design", "multi-surface", "retention" ], "local_code_dir": "src/auto_research/reproductions/downstream_rewards", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "即时 next-item engagement，实验组加入筛选后的 downstream reward", "metrics": [ "baseline.fresh_hit_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Pinterest Homefeed", "metric": "Successful Sessions", "lift_percent": 0.36, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2607.14192", "source_location": "original paper online-result disclosure: product=Pinterest Homefeed; metric=Successful Sessions" }, { "product": "Pinterest Search", "metric": "Search Fulfillment Rate", "lift_percent": 0.25, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2607.14192", "source_location": "original paper online-result disclosure: product=Pinterest Search; metric=Search Fulfillment Rate" }, { "product": "Pinterest Related Pins", "metric": "Successful Sessions", "lift_percent": 0.15, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2607.14192", "source_location": "original paper online-result disclosure: product=Pinterest Related Pins; metric=Successful Sessions" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "dream", "title": "DREAM: A Dual-Loop Recommendation Evolution Framework Powered by Large Language Models", "paper_url": "https://arxiv.org/abs/2608.09408", "detail_path": "reproductions/2608.09408-dream/README.md", "topic": [ "llm-recommendation", "ranking", "reranking", "intent-understanding", "online-policy", "serving" ], "first_author": "DREAM author team (alphabetical order)", "first_author_affiliation": "Taobao & Tmall Group / Alibaba", "published": "2026-08-10", "code": null, "adapter": { "adapter_key": "dream", "arxiv_id": "2608.09408", "title": "DREAM: A Dual-Loop Recommendation Evolution Framework Powered by Large Language Models", "paper_url": "https://arxiv.org/abs/2608.09408", "track": "recommendation", "organization": "Taobao & Tmall Group / Alibaba", "published": "2026-08-10", "code_url": null, "topics": [ "llm-recommendation", "ranking", "reranking", "intent-understanding", "online-policy", "serving" ], "local_code_dir": "src/auto_research/reproductions/dream", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "retrieval/ranking backbone without DREAM overlay", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "3 offline replay generations", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao homepage recommendation", "metric": "IPV", "lift_percent": 2.71, "traffic": "platform-scale online deployment of fine-rank + rerank", "source_url": "https://arxiv.org/pdf/2608.09408", "source_location": "Section 5.3.1 Table 7", "retrieved_at": "2026-08-20" }, { "product": "Taobao homepage recommendation", "metric": "Core IPV", "lift_percent": 3.06, "traffic": "platform-scale online deployment of fine-rank + rerank", "source_url": "https://arxiv.org/pdf/2608.09408", "source_location": "Section 5.3.1 Table 7", "retrieved_at": "2026-08-20" }, { "product": "Taobao homepage recommendation", "metric": "GMV", "lift_percent": 1.31, "traffic": "platform-scale online deployment of fine-rank + rerank", "source_url": "https://arxiv.org/pdf/2608.09408", "source_location": "Section 5.3.1 Table 7", "retrieved_at": "2026-08-20" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "drem", "title": "DrEM: Dual-Side Robust Ensemble Ranking from Noisy User Preference Predictions in Video Recommendation", "paper_url": "https://arxiv.org/abs/2608.12778", "detail_path": "reproductions/2608.12778-drem/README.md", "topic": [ "ranking", "multi-objective", "robust-learning", "video-recommendation" ], "first_author": null, "first_author_affiliation": "Shenzhen University", "published": "2026-08-13", "code": null, "adapter": { "adapter_key": "drem", "arxiv_id": "2608.12778", "title": "DrEM: Dual-Side Robust Ensemble Ranking from Noisy User Preference Predictions in Video Recommendation", "paper_url": "https://arxiv.org/abs/2608.12778", "track": "recommendation", "organization": "Shenzhen University", "published": "2026-08-13", "code_url": null, "topics": [ "ranking", "multi-objective", "robust-learning", "video-recommendation" ], "local_code_dir": "src/auto_research/reproductions/drem", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (proxy multi-objective channels)" ], "baseline": "naive weighted pxtr fusion under the same perturbations", "metrics": [ "hit_at_10", "ndcg_at_10", "naive_mean_absolute_drift", "robust_mean_absolute_drift" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; 24 fixed logit perturbations", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "industrial video recommendation", "metric": "Comment", "lift_percent": 1.388, "traffic": "7-day experiment; 5.1% main traffic per group", "source_url": "https://arxiv.org/html/2608.12778v1", "source_location": "Section 5.4, Table 2", "significance": "p < 0.005", "retrieved_at": "2026-09-01" } ], "selection_exception": null, "evolve_operators": [ "reward:robust-preference" ] } }, { "domain": "recommendation", "key": "drl-put", "title": "Deep Reinforcement Learning for Ranking Utility Tuning in the Ad Recommender System at Pinterest", "paper_url": "https://arxiv.org/abs/2509.05292", "detail_path": "reproductions/2509.05292-drl-put/README.md", "topic": [ "ads-ranking", "reinforcement-learning", "utility-tuning", "off-policy" ], "first_author": null, "first_author_affiliation": "Pinterest", "published": "2025-09-05", "code": null, "adapter": { "adapter_key": "drl-put", "arxiv_id": "2509.05292", "title": "Deep Reinforcement Learning for Ranking Utility Tuning in the Ad Recommender System at Pinterest", "paper_url": "https://arxiv.org/abs/2509.05292", "track": "recommendation", "organization": "Pinterest", "published": "2025-09-05", "code_url": null, "topics": [ "ads-ranking", "reinforcement-learning", "utility-tuning", "off-policy" ], "local_code_dir": "src/auto_research/reproductions/drl_put", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "fixed ranking utility weights", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.mrr_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.mrr_at_10", "method.ndcg_at_10", "paper_results.ctr_percent", "paper_results.cvr_percent", "paper_results.platform_revenue_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.mrr_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Pinterest ads", "metric": "platform revenue", "lift_percent": 0.27, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2509.05292", "source_location": "original paper online-result disclosure: product=Pinterest ads; metric=platform revenue" }, { "product": "Pinterest ads", "metric": "CTR", "lift_percent": 1.62, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2509.05292", "source_location": "original paper online-result disclosure: product=Pinterest ads; metric=CTR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "dual-rerank", "title": "Dual-Rerank: Fusing Sequential Dependencies and Utility for Generative Reranking", "paper_url": "https://arxiv.org/abs/2604.07420", "detail_path": "reproductions/2604.07420-dual-rerank/README.md", "topic": [ "reranking", "knowledge-distillation", "listwise-rl", "serving" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-04-08", "code": null, "adapter": { "adapter_key": "dual-rerank", "arxiv_id": "2604.07420", "title": "Dual-Rerank: Fusing Sequential Dependencies and Utility for Generative Reranking", "paper_url": "https://arxiv.org/abs/2604.07420", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-04-08", "code_url": null, "topics": [ "reranking", "knowledge-distillation", "listwise-rl", "serving" ], "local_code_dir": "src/auto_research/reproductions/dual_rerank", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享 transition + content scorer，实验组只加入 dual-rerank 核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "paper_results.latency_ms", "paper_results.whole_page_ctr_percent", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.ldro_detached_list_reward", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou Search", "metric": "Long View", "lift_percent": 1.107, "traffic": "5% traffic, one month", "source_url": "https://arxiv.org/abs/2604.07420", "source_location": "original paper online-result disclosure: product=Kuaishou Search; metric=Long View" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "dual-sid", "title": "Tokens are All You Need: Dual-purpose Semantic IDs for Achieving LLM-Level I/O Efficiency in Recommendation Systems", "paper_url": "https://arxiv.org/abs/2607.24865", "detail_path": "reproductions/2607.24865-dual-sid/README.md", "topic": [ "semantic-id", "retrieval", "ranking", "content-understanding", "serving" ], "first_author": null, "first_author_affiliation": "Google DeepMind / YouTube", "published": "2026-07-26", "code": null, "adapter": { "adapter_key": "dual-sid", "arxiv_id": "2607.24865", "title": "Tokens are All You Need: Dual-purpose Semantic IDs for Achieving LLM-Level I/O Efficiency in Recommendation Systems", "paper_url": "https://arxiv.org/abs/2607.24865", "track": "recommendation", "organization": "Google DeepMind / YouTube", "published": "2026-07-26", "code_url": null, "topics": [ "semantic-id", "retrieval", "ranking", "content-understanding", "serving" ], "local_code_dir": "src/auto_research/reproductions/dual_sid", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (220 users / 360 items)" ], "baseline": "同切分的 dense transition-content retrieval，实验组加入 Dual-purpose SID 与 SiDec", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "YouTube Watchpage ranking", "metric": "sitewide objective", "lift_percent": 0.09, "traffic": "production online experiment", "source_url": "https://arxiv.org/html/2607.24865v1#S4.SS2", "source_location": "Section 4.2 Table 1", "retrieved_at": "2026-08-09" }, { "product": "YouTube retrieval", "metric": "homepage objective", "lift_percent": 0.13, "traffic": "production online experiment", "source_url": "https://arxiv.org/html/2607.24865v1#S4.SS2", "source_location": "Section 4.2 Table 1", "retrieved_at": "2026-08-09" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "dualgr", "title": "DualGR: Generative Retrieval with Long and Short-Term Interests Modeling", "paper_url": "https://arxiv.org/abs/2511.12518", "detail_path": "reproductions/2511.12518-dualgr/README.md", "topic": [ "retrieval", "generative-recommendation", "semantic-id", "long-short-interest" ], "first_author": null, "first_author_affiliation": "USTC / Kuaishou Technology", "published": "2025-11-16", "code": null, "adapter": { "adapter_key": "dualgr", "arxiv_id": "2511.12518", "title": "DualGR: Generative Retrieval with Long and Short-Term Interests Modeling", "paper_url": "https://arxiv.org/abs/2511.12518", "track": "recommendation", "organization": "USTC / Kuaishou Technology", "published": "2025-11-16", "code_url": null, "topics": [ "retrieval", "generative-recommendation", "semantic-id", "long-short-interest" ], "local_code_dir": "src/auto_research/reproductions/dualgr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享 transition + content scorer，实验组只加入 dualgr 核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.exposure_aware_next_token_loss", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou", "metric": "Video Views", "lift_percent": 0.527, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2511.12518", "source_location": "original paper online-result disclosure: product=Kuaishou; metric=Video Views" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "dynamic-codebook", "title": "From a Static Multi-Level Small Semantic Codebook to a Dynamic Single-Level Large Semantic Codebook for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2608.21012", "detail_path": "reproductions/2608.21012-dynamic-codebook/README.md", "topic": [ "generative-recommendation", "semantic-id", "serving-efficiency" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-08-21", "code": null, "adapter": { "adapter_key": "dynamic-codebook", "arxiv_id": "2608.21012", "title": "From a Static Multi-Level Small Semantic Codebook to a Dynamic Single-Level Large Semantic Codebook for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2608.21012", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-08-21", "code_url": null, "topics": [ "generative-recommendation", "semantic-id", "serving-efficiency" ], "local_code_dir": "src/auto_research/reproductions/dynamic_codebook", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou", "metric": "primary consumption", "lift_percent": 0.792, "traffic": "2.5% traffic, 5 days", "source_url": "https://arxiv.org/html/2608.21012v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "easq", "title": "Towards End-to-End Alignment of User Satisfaction via Questionnaire in Video Recommendation", "paper_url": "https://arxiv.org/abs/2601.20215", "detail_path": "reproductions/2601.20215-easq/README.md", "topic": [ "satisfaction-alignment", "preference-learning", "online-learning" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-01-28", "code": null, "adapter": { "adapter_key": "easq", "arxiv_id": "2601.20215", "title": "Towards End-to-End Alignment of User Satisfaction via Questionnaire in Video Recommendation", "paper_url": "https://arxiv.org/abs/2601.20215", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-01-28", "code_url": null, "topics": [ "satisfaction-alignment", "preference-learning", "online-learning" ], "local_code_dir": "src/auto_research/reproductions/easq", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou video recommendation", "metric": "User satisfaction metric", "lift_percent": 5.1, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2601.20215v1", "source_location": "Section 5.3", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "context:questionnaire-alignment" ] } }, { "domain": "recommendation", "key": "egr", "title": "EGR: Embedding-Native Generative Retrieval with a Shared LLM", "paper_url": "https://arxiv.org/abs/2607.23038", "detail_path": "reproductions/2607.23038-egr/README.md", "topic": [ "retrieval", "generative-retrieval", "shared-encoder" ], "first_author": null, "first_author_affiliation": "Snap Inc.", "published": "2026-07-25", "code": null, "adapter": { "adapter_key": "egr", "arxiv_id": "2607.23038", "title": "EGR: Embedding-Native Generative Retrieval with a Shared LLM", "paper_url": "https://arxiv.org/abs/2607.23038", "track": "recommendation", "organization": "Snap Inc.", "published": "2026-07-25", "code_url": null, "topics": [ "retrieval", "generative-retrieval", "shared-encoder" ], "local_code_dir": "src/auto_research/reproductions/egr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "separate transition/content retrieval", "metrics": [ "hit_at_10", "ndcg_at_10", "shared_encoder_passes", "indexed_items" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Snap DPA", "metric": "CVR", "lift_percent": 2.91, "traffic": "two-week test; 10%/10% traffic split", "source_url": "https://arxiv.org/html/2607.23038v1", "source_location": "Section 4.5, Table 6", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "head:embedding-native" ] } }, { "domain": "recommendation", "key": "ektm", "title": "Effective Knowledge Transfer for Multi-Task Recommendation Models", "paper_url": "https://arxiv.org/abs/2605.05730", "detail_path": "reproductions/2605.05730-ektm/README.md", "topic": [ "multi-task-learning", "knowledge-transfer", "conversion" ], "first_author": null, "first_author_affiliation": "Huawei Technologies", "published": "2026-05-07", "code": null, "adapter": { "adapter_key": "ektm", "arxiv_id": "2605.05730", "title": "Effective Knowledge Transfer for Multi-Task Recommendation Models", "paper_url": "https://arxiv.org/abs/2605.05730", "track": "recommendation", "organization": "Huawei Technologies", "published": "2026-05-07", "code_url": null, "topics": [ "multi-task-learning", "knowledge-transfer", "conversion" ], "local_code_dir": "src/auto_research/reproductions/ektm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "commercial recommendation platform", "metric": "eCPM", "lift_percent": 3.93, "traffic": "5%→20%→50%→100% traffic", "source_url": "https://arxiv.org/html/2605.05730v1", "source_location": "Section 4.4.2 / Figure 3", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "elise", "title": "Multilingual Semantic Retrieval for Apple Music Search", "paper_url": "https://arxiv.org/abs/2607.10239", "detail_path": "reproductions/2607.10239-elise/README.md", "topic": [ "search", "multilingual-retrieval", "score-calibration" ], "first_author": null, "first_author_affiliation": "Apple", "published": "2026-07-11", "code": null, "adapter": { "adapter_key": "elise", "arxiv_id": "2607.10239", "title": "Multilingual Semantic Retrieval for Apple Music Search", "paper_url": "https://arxiv.org/abs/2607.10239", "track": "recommendation", "organization": "Apple", "published": "2026-07-11", "code_url": null, "topics": [ "search", "multilingual-retrieval", "score-calibration" ], "local_code_dir": "src/auto_research/reproductions/elise", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "lexical/token-index retrieval", "metrics": [ "hit_at_10", "ndcg_at_10", "retrieval_sources", "quantile_matched" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Apple Music worldwide search", "metric": "Conversion Rate", "lift_percent": 2.28, "traffic": "18-day test; 10% treatment and 10% control", "source_url": "https://arxiv.org/html/2607.10239v1", "source_location": "Section 7.2, Table 8", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "context:quantile-fusion" ] } }, { "domain": "recommendation", "key": "esmm", "title": "Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate", "paper_url": "https://arxiv.org/abs/1804.07931", "detail_path": "reproductions/1804.07931-esmm/README.md", "topic": [ "ranking", "classic", "multi-task", "conversion" ], "first_author": null, "first_author_affiliation": "Alibaba", "published": "2018-04-21", "code": null, "adapter": { "adapter_key": "esmm", "arxiv_id": "1804.07931", "title": "Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate", "paper_url": "https://arxiv.org/abs/1804.07931", "track": "recommendation", "organization": "Alibaba", "published": "2018-04-21", "code_url": null, "topics": [ "ranking", "classic", "multi-task", "conversion" ], "local_code_dir": "src/auto_research/reproductions/esmm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K entire-space multitask construction" ], "baseline": "只在 clicked space 训练 CVR tower", "metrics": [ "baseline.click_auc", "baseline.click_auc_std", "baseline.click_logloss", "baseline.click_logloss_std", "baseline.conversion_auc", "baseline.conversion_auc_std", "baseline.conversion_logloss", "baseline.conversion_logloss_std", "baseline.mean_auc", "baseline.mean_auc_std", "method.click_auc", "method.click_auc_std", "method.click_logloss", "method.click_logloss_std", "method.conversion_auc", "method.conversion_auc_std", "method.conversion_logloss", "method.conversion_logloss_std", "method.mean_auc", "method.mean_auc_std", "relative.conversion_auc_percent", "relative.mean_auc_percent" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=200", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": "用户明确批准 ESMM 为经典例外；原文工业部署不满足当前新论文的量化线上证据格式。", "evolve_operators": [] } }, { "domain": "recommendation", "key": "evorec", "title": "EvoRec: Self-Evolving Agentic Recommender Systems", "paper_url": "https://arxiv.org/abs/2606.28368", "detail_path": "reproductions/2606.28368-evorec/README.md", "topic": [ "auto-research", "agent", "skill-evolution" ], "first_author": null, "first_author_affiliation": "Alibaba International Digital Commerce Group", "published": "2026-06-15", "code": null, "adapter": { "adapter_key": "evorec", "arxiv_id": "2606.28368", "title": "EvoRec: Self-Evolving Agentic Recommender Systems", "paper_url": "https://arxiv.org/abs/2606.28368", "track": "recommendation", "organization": "Alibaba International Digital Commerce Group", "published": "2026-06-15", "code_url": null, "topics": [ "auto-research", "agent", "skill-evolution" ], "local_code_dir": "src/auto_research/reproductions/evorec", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (260 users / 420 items)" ], "baseline": "固定 ensemble，实验组为三代 skill-memory 进化结果", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Alibaba international recommendation", "metric": "Revenue", "lift_percent": 1.85, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2606.28368", "source_location": "original paper online-result disclosure: product=Alibaba international recommendation; metric=Revenue" }, { "product": "Alibaba international recommendation", "metric": "CTR", "lift_percent": 1.02, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2606.28368", "source_location": "original paper online-result disclosure: product=Alibaba international recommendation; metric=CTR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "feedback-policy", "title": "From Understanding to Action: Feedback-Grounded Policy Discovery for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2607.27789", "detail_path": "reproductions/2607.27789-feedback-policy/README.md", "topic": [ "generative-recommendation", "policy-optimization", "distillation" ], "first_author": null, "first_author_affiliation": "Huazhong Agricultural University (Kuaishou internship)", "published": "2026-07-30", "code": null, "adapter": { "adapter_key": "feedback-policy", "arxiv_id": "2607.27789", "title": "From Understanding to Action: Feedback-Grounded Policy Discovery for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2607.27789", "track": "recommendation", "organization": "Huazhong Agricultural University (Kuaishou internship)", "published": "2026-07-30", "code_url": null, "topics": [ "generative-recommendation", "policy-optimization", "distillation" ], "local_code_dir": "src/auto_research/reproductions/feedback_policy", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou advertising", "metric": "Revenue", "lift_percent": 4.506, "traffic": "large-scale online A/B", "source_url": "https://arxiv.org/html/2607.27789v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "filterllm", "title": "FilterLLM: Text-To-Distribution LLM for Billion-Scale Cold-Start Recommendation", "paper_url": "https://arxiv.org/abs/2502.16924", "detail_path": "reproductions/2502.16924-filterllm/README.md", "topic": [ "llm-recommendation", "cold-start", "retrieval", "text-to-distribution" ], "first_author": null, "first_author_affiliation": "Alibaba", "published": "2025-02-24", "code": null, "adapter": { "adapter_key": "filterllm", "arxiv_id": "2502.16924", "title": "FilterLLM: Text-To-Distribution LLM for Billion-Scale Cold-Start Recommendation", "paper_url": "https://arxiv.org/abs/2502.16924", "track": "recommendation", "organization": "Alibaba", "published": "2025-02-24", "code_url": null, "topics": [ "llm-recommendation", "cold-start", "retrieval", "text-to-distribution" ], "local_code_dir": "src/auto_research/reproductions/filterllm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "content-to-item judgment", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.mrr_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.mrr_at_10", "method.ndcg_at_10", "paper_results.gmv_percent", "paper_results.latency_reduction_percent", "paper_results.pctr_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.mrr_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Alibaba cold-start", "metric": "Cold-PV", "lift_percent": 5.13, "traffic": "two-month A/B", "source_url": "https://arxiv.org/abs/2502.16924", "source_location": "original paper online-result disclosure: product=Alibaba cold-start; metric=Cold-PV" }, { "product": "Alibaba cold-start", "metric": "GMV", "lift_percent": 10.86, "traffic": "two-month A/B", "source_url": "https://arxiv.org/abs/2502.16924", "source_location": "original paper online-result disclosure: product=Alibaba cold-start; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "fluid", "title": "FLUID: From Ephemeral IDs to Multimodal Semantic Codes for Industrial-Scale Livestreaming Recommendation", "paper_url": "https://arxiv.org/abs/2605.21832", "detail_path": "reproductions/2605.21832-fluid/README.md", "topic": [ "llm-recommendation", "multimodal", "semantic-id", "cold-start", "ranking" ], "first_author": null, "first_author_affiliation": "TikTok / ByteDance", "published": "2026-05-20", "code": null, "adapter": { "adapter_key": "fluid", "arxiv_id": "2605.21832", "title": "FLUID: From Ephemeral IDs to Multimodal Semantic Codes for Industrial-Scale Livestreaming Recommendation", "paper_url": "https://arxiv.org/abs/2605.21832", "track": "recommendation", "organization": "TikTok / ByteDance", "published": "2026-05-20", "code_url": null, "topics": [ "llm-recommendation", "multimodal", "semantic-id", "cold-start", "ranking" ], "local_code_dir": "src/auto_research/reproductions/fluid", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "含 item transition/popularity 的 ID ranker", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "TikTok / ByteDance livestreaming", "metric": "Quality Watch Duration", "lift_percent": 0.55, "traffic": "production online A/B test", "source_url": "https://arxiv.org/abs/2605.21832", "source_location": "original paper online-result disclosure: product=TikTok / ByteDance livestreaming; metric=Quality Watch Duration" }, { "product": "TikTok / ByteDance livestreaming", "metric": "Cold-Start Room Views", "lift_percent": 2.05, "traffic": "production online A/B test", "source_url": "https://arxiv.org/abs/2605.21832", "source_location": "original paper online-result disclosure: product=TikTok / ByteDance livestreaming; metric=Cold-Start Room Views" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "friend-gnn", "title": "Scaling Graph Neural Networks for Friend Recommendation: Multi-Hash User Embeddings and Temporal Neighbor Sampling", "paper_url": "https://arxiv.org/abs/2608.27413", "detail_path": "reproductions/2608.27413-friend-gnn/README.md", "topic": [ "friend-recommendation", "graph-ranking", "multi-hash-embedding", "temporal-sampling" ], "first_author": null, "first_author_affiliation": "AI VK", "published": "2026-08-27", "code": "https://github.com/makut/VK-GNN", "adapter": { "adapter_key": "friend-gnn", "arxiv_id": "2608.27413", "title": "Scaling Graph Neural Networks for Friend Recommendation: Multi-Hash User Embeddings and Temporal Neighbor Sampling", "paper_url": "https://arxiv.org/abs/2608.27413", "track": "recommendation", "organization": "AI VK", "published": "2026-08-27", "code_url": "https://github.com/makut/VK-GNN", "topics": [ "friend-recommendation", "graph-ranking", "multi-hash-embedding", "temporal-sampling" ], "local_code_dir": "src/auto_research/reproductions/friend_gnn", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "popularity ranker on the identical full candidate catalog", "metrics": [ "Hit@10", "NDCG@10", "hash table compression", "temporal sampling complexity" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "VK friend recommendation", "metric": "friend additions", "lift_percent": 16.0, "traffic": "production A/B test", "source_url": "https://arxiv.org/html/2608.27413v1#S6.SS6", "source_location": "Section 6.6", "retrieved_at": "2026-08-29" }, { "product": "VK friend recommendation", "metric": "unique friend adders", "lift_percent": 11.5, "traffic": "production A/B test", "source_url": "https://arxiv.org/html/2608.27413v1#S6.SS6", "source_location": "Section 6.6", "retrieved_at": "2026-08-29" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "fuxi-alpha", "title": "FuXi-α: Scaling Recommendation Model with Feature Interaction Enhanced Transformer", "paper_url": "https://arxiv.org/abs/2502.03036", "detail_path": "reproductions/2502.03036-fuxi-alpha/README.md", "topic": [ "ranking", "transformer", "feature-interaction", "scaling" ], "first_author": null, "first_author_affiliation": "Huawei / USTC", "published": "2025-02-05", "code": "https://github.com/USTC-StarTeam/FuXi-alpha", "adapter": { "adapter_key": "fuxi-alpha", "arxiv_id": "2502.03036", "title": "FuXi-α: Scaling Recommendation Model with Feature Interaction Enhanced Transformer", "paper_url": "https://arxiv.org/abs/2502.03036", "track": "recommendation", "organization": "Huawei / USTC", "published": "2025-02-05", "code_url": "https://github.com/USTC-StarTeam/FuXi-alpha", "topics": [ "ranking", "transformer", "feature-interaction", "scaling" ], "local_code_dir": "src/auto_research/reproductions/fuxi_alpha", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "single-channel sequential attention", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.mrr_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.mrr_at_10", "method.ndcg_at_10", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.mrr_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Huawei Music", "metric": "songs played", "lift_percent": 4.67, "traffic": "30% traffic, 7 days", "source_url": "https://arxiv.org/abs/2502.03036", "source_location": "original paper online-result disclosure: product=Huawei Music; metric=songs played" }, { "product": "Huawei Music", "metric": "listening duration", "lift_percent": 5.1, "traffic": "30% traffic, 7 days", "source_url": "https://arxiv.org/abs/2502.03036", "source_location": "original paper online-result disclosure: product=Huawei Music; metric=listening duration" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "g2rec", "title": "Structuring and Tokenizing Distributed User Interest Context for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2606.20554", "detail_path": "reproductions/2606.20554-g2rec/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "g2rec", "arxiv_id": "2606.20554", "title": "Structuring and Tokenizing Distributed User Interest Context for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2606.20554", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/g2rec", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "Amazon Beauty 5-core" ], "baseline": "参数级别匹配的 Item-only Autoregressive Decoder", "metrics": [ "configuration.profile_loss_weight", "ndcg_gain_percent", "results.g2rec_item_plus_interest_tokens.hit_at_10", "results.g2rec_item_plus_interest_tokens.ndcg_at_10", "results.item_tokens_only.hit_at_10", "results.item_tokens_only.ndcg_at_10", "training.g2rec_final_loss", "training.item_only_final_loss" ], "default_seeds": [ 42 ], "budget": "steps=240", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "gala", "title": "GALA: Generative Aligned Learning for Adaptive Multimodal Representation in the Taobao Shangou Recommender System", "paper_url": "https://arxiv.org/abs/2607.29213", "detail_path": "reproductions/2607.29213-gala/README.md", "topic": [ "multimodal-recommendation", "representation-alignment", "reinforcement-learning" ], "first_author": null, "first_author_affiliation": "Rajax Network Technology / Taobao Shangou / Alibaba", "published": "2026-07-31", "code": null, "adapter": { "adapter_key": "gala", "arxiv_id": "2607.29213", "title": "GALA: Generative Aligned Learning for Adaptive Multimodal Representation in the Taobao Shangou Recommender System", "paper_url": "https://arxiv.org/abs/2607.29213", "track": "recommendation", "organization": "Rajax Network Technology / Taobao Shangou / Alibaba", "published": "2026-07-31", "code_url": null, "topics": [ "multimodal-recommendation", "representation-alignment", "reinforcement-learning" ], "local_code_dir": "src/auto_research/reproductions/gala", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao Shangou", "metric": "order volume", "lift_percent": 0.55, "traffic": "randomized traffic split", "source_url": "https://arxiv.org/html/2607.29213v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "gatesid", "title": "GateSID: Adaptive Gating for Semantic-Collaborative Alignment in Cold-Start Recommendation", "paper_url": "https://arxiv.org/abs/2603.22916", "detail_path": "reproductions/2603.22916-gatesid/README.md", "topic": [ "semantic-id", "cold-start", "contrastive-learning" ], "first_author": null, "first_author_affiliation": "Alibaba International Digital Commerce", "published": "2026-03-24", "code": null, "adapter": { "adapter_key": "gatesid", "arxiv_id": "2603.22916", "title": "GateSID: Adaptive Gating for Semantic-Collaborative Alignment in Cold-Start Recommendation", "paper_url": "https://arxiv.org/abs/2603.22916", "track": "recommendation", "organization": "Alibaba International Digital Commerce", "published": "2026-03-24", "code_url": null, "topics": [ "semantic-id", "cold-start", "contrastive-learning" ], "local_code_dir": "src/auto_research/reproductions/gatesid", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "industrial recommendation", "metric": "GMV", "lift_percent": 2.6, "traffic": "20% traffic, two weeks", "source_url": "https://arxiv.org/html/2603.22916v1", "source_location": "Section 3.4", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "genfacet", "title": "GenFacet: End-to-End Generative Faceted Search via Multi-Task Preference Alignment in E-Commerce", "paper_url": "https://arxiv.org/abs/2603.19665", "detail_path": "reproductions/2603.19665-genfacet/README.md", "topic": [ "e-commerce-search", "facet-generation", "preference-alignment" ], "first_author": null, "first_author_affiliation": "JD.com", "published": "2026-03-20", "code": null, "adapter": { "adapter_key": "genfacet", "arxiv_id": "2603.19665", "title": "GenFacet: End-to-End Generative Faceted Search via Multi-Task Preference Alignment in E-Commerce", "paper_url": "https://arxiv.org/abs/2603.19665", "track": "recommendation", "organization": "JD.com", "published": "2026-03-20", "code_url": null, "topics": [ "e-commerce-search", "facet-generation", "preference-alignment" ], "local_code_dir": "src/auto_research/reproductions/genfacet", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "JD.com e-commerce search", "metric": "Facet CTR", "lift_percent": 42.0, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2603.19665v1", "source_location": "Abstract", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "reward:facet-preference" ] } }, { "domain": "recommendation", "key": "genpage", "title": "GenPage: Towards End-to-End Generative Homepage Construction at Netflix", "paper_url": "https://arxiv.org/abs/2606.31031", "detail_path": "reproductions/2606.31031-genpage/README.md", "topic": [ "generative-recommendation", "page-generation", "reinforcement-learning" ], "first_author": null, "first_author_affiliation": "Netflix", "published": "2026-06-30", "code": null, "adapter": { "adapter_key": "genpage", "arxiv_id": "2606.31031", "title": "GenPage: Towards End-to-End Generative Homepage Construction at Netflix", "paper_url": "https://arxiv.org/abs/2606.31031", "track": "recommendation", "organization": "Netflix", "published": "2026-06-30", "code_url": null, "topics": [ "generative-recommendation", "page-generation", "reinforcement-learning" ], "local_code_dir": "src/auto_research/reproductions/genpage", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Netflix homepage", "metric": "core engagement", "lift_percent": 0.24, "traffic": "online A/B; p<0.001", "source_url": "https://arxiv.org/html/2606.31031v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "genrank", "title": "Towards Large-scale Generative Ranking", "paper_url": "https://arxiv.org/abs/2505.04180", "detail_path": "reproductions/2505.04180-genrank/README.md", "topic": [ "generative-ranking", "action-modeling", "efficiency" ], "first_author": null, "first_author_affiliation": "Xiaohongshu", "published": "2025-05", "code": null, "adapter": { "adapter_key": "genrank", "arxiv_id": "2505.04180", "title": "Towards Large-scale Generative Ranking", "paper_url": "https://arxiv.org/abs/2505.04180", "track": "recommendation", "organization": "Xiaohongshu", "published": "2025-05", "code_url": null, "topics": [ "generative-ranking", "action-modeling", "efficiency" ], "local_code_dir": "src/auto_research/reproductions/genrank", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens 100K explicit action sequence" ], "baseline": "item/action 交错序列模型", "metrics": [ "auc.genrank.mean", "auc.genrank.std", "auc.interleaved.mean", "auc.interleaved.std", "auc.relative_change_percent", "latency_ms_per_example.genrank.mean", "latency_ms_per_example.genrank.std", "latency_ms_per_example.interleaved.mean", "latency_ms_per_example.interleaved.std", "latency_ms_per_example.relative_change_percent" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=120", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Explore Feed", "metric": "engagements", "lift_percent": 1.2474, "traffic": "10% treatment/control, 15 days", "source_url": "https://arxiv.org/abs/2505.04180", "source_location": "original paper online-result disclosure: product=Explore Feed; metric=engagements" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "genrec", "title": "GenRec: A Preference-Oriented Generative Framework for Large-Scale Recommendation", "paper_url": "https://arxiv.org/abs/2604.14878", "detail_path": "reproductions/2604.14878-genrec/README.md", "topic": [ "generative-recommendation", "grpo", "page-wise" ], "first_author": null, "first_author_affiliation": "JD.com", "published": "2026-04-16", "code": null, "adapter": { "adapter_key": "genrec", "arxiv_id": "2604.14878", "title": "GenRec: A Preference-Oriented Generative Framework for Large-Scale Recommendation", "paper_url": "https://arxiv.org/abs/2604.14878", "track": "recommendation", "organization": "JD.com", "published": "2026-04-16", "code_url": null, "topics": [ "generative-recommendation", "grpo", "page-wise" ], "local_code_dir": "src/auto_research/reproductions/genrec", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (260 users / 420 items)" ], "baseline": "point-wise ranker，实验组为 page-wise NTP + GRPO-SR", "metrics": [ "baseline.fresh_hit_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "JD recommendation", "metric": "Clicks", "lift_percent": 9.5, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2604.14878", "source_location": "original paper online-result disclosure: product=JD recommendation; metric=Clicks" }, { "product": "JD recommendation", "metric": "Transactions", "lift_percent": 8.7, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2604.14878", "source_location": "original paper online-result disclosure: product=JD recommendation; metric=Transactions" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "genrec-netflix", "title": "GenRec: An LLM-Backed Recommendation Ranker at Netflix", "paper_url": "https://arxiv.org/abs/2608.10257", "detail_path": "reproductions/2608.10257-genrec-netflix/README.md", "topic": [ "ranking", "llm-recommendation", "post-training", "reward-weighting", "context-engineering", "training-serving" ], "first_author": null, "first_author_affiliation": "Netflix", "published": "2026-08-10", "code": null, "adapter": { "adapter_key": "genrec-netflix", "arxiv_id": "2608.10257", "title": "GenRec: An LLM-Backed Recommendation Ranker at Netflix", "paper_url": "https://arxiv.org/abs/2608.10257", "track": "recommendation", "organization": "Netflix", "published": "2026-08-10", "code_url": null, "topics": [ "ranking", "llm-recommendation", "post-training", "reward-weighting", "context-engineering", "training-serving" ], "local_code_dir": "src/auto_research/reproductions/genrec_netflix", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "ID-only discriminative GRU ranker with matched Phase-2 examples", "metrics": [ "Hit@10", "NDCG@10", "MRR", "head share@10" ], "default_seeds": [ 42 ], "budget": "steps=120", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Netflix batch-compute recommendation surfaces", "metric": "short-term homepage engagement", "lift_percent": 0.115, "traffic": "approximately 10% traffic", "source_url": "https://arxiv.org/html/2608.10257v1#S5.SS1", "source_location": "Section 5.1 / Figure 3", "experiment_duration": "4 weeks", "significance": "P = 3.1e-10", "retrieved_at": "2026-08-13" }, { "product": "Netflix batch-compute recommendation surfaces", "metric": "long-term core metric", "lift_percent": 0.006, "traffic": "approximately 10% traffic", "source_url": "https://arxiv.org/html/2608.10257v1#S5.SS1", "source_location": "Section 5.1 / Figure 3", "experiment_duration": "4 weeks", "significance": "P = 0.025", "retrieved_at": "2026-08-13" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "gese", "title": "Generate to Explore, Select to Exploit: Aligning LLM-based Headline Generation with Personalized Recommendation", "paper_url": "https://arxiv.org/abs/2609.15094", "detail_path": "reproductions/2609.15094-gese/README.md", "topic": [ "presentation-layer", "headline-generation", "gspo", "bandit-selection" ], "first_author": null, "first_author_affiliation": "Baidu", "published": "2026-09-14", "code": null, "adapter": { "adapter_key": "gese", "arxiv_id": "2609.15094", "title": "Generate to Explore, Select to Exploit: Aligning LLM-based Headline Generation with Personalized Recommendation", "paper_url": "https://arxiv.org/abs/2609.15094", "track": "recommendation", "organization": "Baidu", "published": "2026-09-14", "code_url": null, "topics": [ "presentation-layer", "headline-generation", "gspo", "bandit-selection" ], "local_code_dir": "src/auto_research/reproductions/gese", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "commercial information feed with 100M+ DAU", "metric": "CTR", "lift_percent": 2.57, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2609.15094v1", "source_location": "Section 4.2; Table 2", "retrieved_at": "2026-09-16" }, { "product": "commercial information feed with 100M+ DAU", "metric": "dwell time", "lift_percent": 0.87, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2609.15094v1", "source_location": "Section 4.2; Table 2", "retrieved_at": "2026-09-16" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "glide", "title": "Deploying Semantic ID-based Generative Retrieval for Large-Scale Podcast Discovery at Spotify", "paper_url": "https://arxiv.org/abs/2603.17540", "detail_path": "reproductions/2603.17540-glide/README.md", "topic": [ "generative-retrieval", "semantic-id", "soft-prompt" ], "first_author": null, "first_author_affiliation": "Spotify", "published": "2026-03-18", "code": null, "adapter": { "adapter_key": "glide", "arxiv_id": "2603.17540", "title": "Deploying Semantic ID-based Generative Retrieval for Large-Scale Podcast Discovery at Spotify", "paper_url": "https://arxiv.org/abs/2603.17540", "track": "recommendation", "organization": "Spotify", "published": "2026-03-18", "code_url": null, "topics": [ "generative-retrieval", "semantic-id", "soft-prompt" ], "local_code_dir": "src/auto_research/reproductions/glide", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (260 users / 420 items)" ], "baseline": "传统 transition/content retrieval，实验组为 Semantic-ID 双 prompt 检索", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Spotify", "metric": "Non-habitual streaming", "lift_percent": 5.4, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2603.17540", "source_location": "original paper online-result disclosure: product=Spotify; metric=Non-habitual streaming" }, { "product": "Spotify", "metric": "New-show discovery", "lift_percent": 14.3, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2603.17540", "source_location": "original paper online-result disclosure: product=Spotify; metric=New-show discovery" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "glorank", "title": "From Local Indices to Global Identifiers: Generative Reranking via Global Action Space", "paper_url": "https://arxiv.org/abs/2604.25291", "detail_path": "reproductions/2604.25291-glorank/README.md", "topic": [ "reranking", "generative-recommendation", "semantic-id", "reinforcement-learning" ], "first_author": null, "first_author_affiliation": "City University of Hong Kong / Kuaishou / UC San Diego", "published": "2026-04-28", "code": null, "adapter": { "adapter_key": "glorank", "arxiv_id": "2604.25291", "title": "From Local Indices to Global Identifiers: Generative Reranking via Global Action Space", "paper_url": "https://arxiv.org/abs/2604.25291", "track": "recommendation", "organization": "City University of Hong Kong / Kuaishou / UC San Diego", "published": "2026-04-28", "code_url": null, "topics": [ "reranking", "generative-recommendation", "semantic-id", "reinforcement-learning" ], "local_code_dir": "src/auto_research/reproductions/glorank", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享 transition + content scorer，实验组只加入 glorank 核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.group_relative_reward_update", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou", "metric": "Watch Time", "lift_percent": 0.095, "traffic": "7.8% traffic, 14 days", "source_url": "https://arxiv.org/abs/2604.25291", "source_location": "original paper online-result disclosure: product=Kuaishou; metric=Watch Time" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "gpl-prerank", "title": "Generative Pseudo-Labeling for Pre-Ranking with LLMs", "paper_url": "https://arxiv.org/abs/2602.20995", "detail_path": "reproductions/2602.20995-gpl-prerank/README.md", "topic": [ "pre-ranking", "pseudo-labeling", "llm-recommendation" ], "first_author": null, "first_author_affiliation": "Alibaba Group / Renmin University", "published": "2026-02-24", "code": null, "adapter": { "adapter_key": "gpl-prerank", "arxiv_id": "2602.20995", "title": "Generative Pseudo-Labeling for Pre-Ranking with LLMs", "paper_url": "https://arxiv.org/abs/2602.20995", "track": "recommendation", "organization": "Alibaba Group / Renmin University", "published": "2026-02-24", "code_url": null, "topics": [ "pre-ranking", "pseudo-labeling", "llm-recommendation" ], "local_code_dir": "src/auto_research/reproductions/gpl_prerank", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao Guess What You Like", "metric": "CTR", "lift_percent": 3.07, "traffic": "hundreds of millions DAU, two weeks", "source_url": "https://arxiv.org/html/2602.20995v1", "source_location": "Section 4.5 / Table 4", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "gr4ad", "title": "Generative Recommendation for Large-Scale Advertising", "paper_url": "https://arxiv.org/abs/2602.22732", "detail_path": "reproductions/2602.22732-gr4ad/README.md", "topic": [ "generative-retrieval", "advertising", "reinforcement-learning", "serving" ], "first_author": null, "first_author_affiliation": "Kuaishou", "published": "2026-02", "code": null, "adapter": { "adapter_key": "gr4ad", "arxiv_id": "2602.22732", "title": "Generative Recommendation for Large-Scale Advertising", "paper_url": "https://arxiv.org/abs/2602.22732", "track": "recommendation", "organization": "Kuaishou", "published": "2026-02", "code_url": null, "topics": [ "generative-retrieval", "advertising", "reinforcement-learning", "serving" ], "local_code_dir": "src/auto_research/reproductions/gr4ad", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "DIN，GR4AD NDCG@10 相对 DIN +69.67%", "metrics": [ "paper_online_ab.ad_revenue_percent", "results.din.hit_at_10", "results.din.ndcg_at_10", "results.dlrm.hit_at_10", "results.dlrm.ndcg_at_10", "results.gr4ad.hit_at_10", "results.gr4ad.ndcg_at_10", "results.ndcg_ablation_percent", "results.ndcg_vs_din_percent" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=100", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou Ads", "metric": "ad revenue", "lift_percent": 4.2, "traffic": "large-scale online A/B", "source_url": "https://arxiv.org/abs/2602.22732", "source_location": "original paper online-result disclosure: product=Kuaishou Ads; metric=ad revenue" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "grc", "title": "Learning to Reflect and Correct: Towards Better Decoding Trajectories for Large-Scale Generative Recommendation", "paper_url": "https://arxiv.org/abs/2602.23639", "detail_path": "reproductions/2602.23639-grc/README.md", "topic": [ "generative-recommendation", "reinforcement-learning", "self-correction" ], "first_author": null, "first_author_affiliation": "Alibaba International / Wuhan University", "published": "2026-02-27", "code": null, "adapter": { "adapter_key": "grc", "arxiv_id": "2602.23639", "title": "Learning to Reflect and Correct: Towards Better Decoding Trajectories for Large-Scale Generative Recommendation", "paper_url": "https://arxiv.org/abs/2602.23639", "track": "recommendation", "organization": "Alibaba International / Wuhan University", "published": "2026-02-27", "code_url": null, "topics": [ "generative-recommendation", "reinforcement-learning", "self-correction" ], "local_code_dir": "src/auto_research/reproductions/grc", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "movielens 100k" ], "baseline": "同一 SID 的 GR backbone", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Alibaba International", "metric": "Revenue", "lift_percent": 1.79, "traffic": "2026-01-02 to 2026-01-12, 15%/15%", "source_url": "https://arxiv.org/abs/2602.23639", "source_location": "original paper online-result disclosure: product=Alibaba International; metric=Revenue" }, { "product": "Alibaba International", "metric": "CTR", "lift_percent": 2.11, "traffic": "2026-01-02 to 2026-01-12, 15%/15%", "source_url": "https://arxiv.org/abs/2602.23639", "source_location": "original paper online-result disclosure: product=Alibaba International; metric=CTR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "growthgr", "title": "Towards Sustainable Growth: A Multi-Value-Aware Retrieval Framework for E-Commerce Search", "paper_url": "https://arxiv.org/abs/2605.17994", "detail_path": "reproductions/2605.17994-growthgr/README.md", "topic": [ "generative-recommendation", "reinforcement-learning", "new-items" ], "first_author": null, "first_author_affiliation": "Alibaba Group / Taobao & Tmall", "published": "2026-05-18", "code": null, "adapter": { "adapter_key": "growthgr", "arxiv_id": "2605.17994", "title": "Towards Sustainable Growth: A Multi-Value-Aware Retrieval Framework for E-Commerce Search", "paper_url": "https://arxiv.org/abs/2605.17994", "track": "recommendation", "organization": "Alibaba Group / Taobao & Tmall", "published": "2026-05-18", "code_url": null, "topics": [ "generative-recommendation", "reinforcement-learning", "new-items" ], "local_code_dir": "src/auto_research/reproductions/growthgr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "movielens 100k" ], "baseline": "同一 transition policy 的 NTP retrieval", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao/Tmall Search", "metric": "new-item GMV", "lift_percent": 5.3, "traffic": "online A/B test", "source_url": "https://arxiv.org/abs/2605.17994", "source_location": "original paper online-result disclosure: product=Taobao/Tmall Search; metric=new-item GMV" }, { "product": "Taobao/Tmall Search", "metric": "overall search GMV", "lift_percent": 0.3, "traffic": "online A/B test", "source_url": "https://arxiv.org/abs/2605.17994", "source_location": "original paper online-result disclosure: product=Taobao/Tmall Search; metric=overall search GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "gryphon-v2", "title": "Gryphon-v2: One Model in Place of a Cascade — Generate-and-Rank Recommender with Rollout Distillation", "paper_url": "https://arxiv.org/abs/2608.06213", "detail_path": "reproductions/2608.06213-gryphon-v2/README.md", "topic": [ "generative-recommendation", "retrieval", "ranking", "rollout-distillation" ], "first_author": null, "first_author_affiliation": "Yandex", "published": "2026-08-06", "code": null, "adapter": { "adapter_key": "gryphon-v2", "arxiv_id": "2608.06213", "title": "Gryphon-v2: One Model in Place of a Cascade — Generate-and-Rank Recommender with Rollout Distillation", "paper_url": "https://arxiv.org/abs/2608.06213", "track": "recommendation", "organization": "Yandex", "published": "2026-08-06", "code_url": null, "topics": [ "generative-recommendation", "retrieval", "ranking", "rollout-distillation" ], "local_code_dir": "src/auto_research/reproductions/gryphon_v2", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "同预算生成式召回，实验组加入双来源蒸馏与 ranker", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "gryphon_v2.hit_at_10", "gryphon_v2.ndcg_at_10", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Yandex Music large-scale recommendation surface", "metric": "active users", "lift_percent": 1.41, "traffic": "single model replaces 15+ candidate generators, pre-ranking and final ranking at comparable latency", "source_url": "https://arxiv.org/abs/2608.06213", "source_location": "original paper online-result disclosure: product=Yandex Music large-scale recommendation surface; metric=active users" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "guess-where-you-go", "title": "Guess Where You Go: Generative Next Point-of-Interest Recommendation in Amap", "paper_url": "https://arxiv.org/abs/2607.26073", "detail_path": "reproductions/2607.26073-guess-where-you-go/README.md", "topic": [ "generative-recommendation", "next-poi", "semantic-id", "reinforcement-learning" ], "first_author": null, "first_author_affiliation": "Amap / Alibaba", "published": "2026-07-13", "code": "https://github.com/alibaba/SimCIT", "adapter": { "adapter_key": "guess-where-you-go", "arxiv_id": "2607.26073", "title": "Guess Where You Go: Generative Next Point-of-Interest Recommendation in Amap", "paper_url": "https://arxiv.org/abs/2607.26073", "track": "recommendation", "organization": "Amap / Alibaba", "published": "2026-07-13", "code_url": "https://github.com/alibaba/SimCIT", "topics": [ "generative-recommendation", "next-poi", "semantic-id", "reinforcement-learning" ], "local_code_dir": "src/auto_research/reproductions/guess_where_you_go", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Amap homepage", "metric": "P-CTR", "lift_percent": 5.83, "traffic": "one-month A/B", "source_url": "https://arxiv.org/html/2607.26073v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "ha-moe", "title": "Heterogeneous Ranking in Industrial-Scale Recommender Systems: A Case Study", "paper_url": "https://arxiv.org/abs/2607.27577", "detail_path": "reproductions/2607.27577-ha-moe/README.md", "topic": [ "ranking", "mixture-of-experts", "heterogeneity", "diversity" ], "first_author": null, "first_author_affiliation": "Google / Discover", "published": "2026-07-30", "code": null, "adapter": { "adapter_key": "ha-moe", "arxiv_id": "2607.27577", "title": "Heterogeneous Ranking in Industrial-Scale Recommender Systems: A Case Study", "paper_url": "https://arxiv.org/abs/2607.27577", "track": "recommendation", "organization": "Google / Discover", "published": "2026-07-30", "code_url": null, "topics": [ "ranking", "mixture-of-experts", "heterogeneity", "diversity" ], "local_code_dir": "src/auto_research/reproductions/ha_moe", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (220 users / 360 items)" ], "baseline": "同切分的 homogeneous transition-content ranker，实验组加入 HA-MoE", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Google Discover", "metric": "DAU", "lift_percent": 0.22, "traffic": "7-day 1% live-traffic A/B", "source_url": "https://arxiv.org/html/2607.27577v1#S4.SS4", "source_location": "Section 4.4 Table 4", "experiment_duration": "7 days", "significance": "±0.11%", "retrieved_at": "2026-08-09" }, { "product": "Google Discover", "metric": "Diverse Engagement Rate", "lift_percent": 0.54, "traffic": "7-day 1% live-traffic A/B", "source_url": "https://arxiv.org/html/2607.27577v1#S4.SS4", "source_location": "Section 4.4 Table 4", "experiment_duration": "7 days", "significance": "±0.07%", "retrieved_at": "2026-08-09" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "hap", "title": "Not All Candidates are Created Equal: Heterogeneity-Aware Pre-ranking", "paper_url": "https://arxiv.org/abs/2603.03770", "detail_path": "reproductions/2603.03770-hap/README.md", "topic": [ "pre-ranking", "contrastive-learning", "dynamic-routing" ], "first_author": null, "first_author_affiliation": "ByteDance / Toutiao", "published": "2026-03-04", "code": "https://github.com/Toutiao-Rec/HAP", "adapter": { "adapter_key": "hap", "arxiv_id": "2603.03770", "title": "Not All Candidates are Created Equal: Heterogeneity-Aware Pre-ranking", "paper_url": "https://arxiv.org/abs/2603.03770", "track": "recommendation", "organization": "ByteDance / Toutiao", "published": "2026-03-04", "code_url": "https://github.com/Toutiao-Rec/HAP", "topics": [ "pre-ranking", "contrastive-learning", "dynamic-routing" ], "local_code_dir": "src/auto_research/reproductions/hap", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享 transition + content scorer，实验组只加入 hap 核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Toutiao", "metric": "App Usage Duration", "lift_percent": 0.4, "traffic": "nine-month deployment", "source_url": "https://arxiv.org/abs/2603.03770", "source_location": "original paper online-result disclosure: product=Toutiao; metric=App Usage Duration" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "harmonrank", "title": "Rethinking Multi-objective Ranking Ensemble in Recommender System: From Score Fusion to Rank Consistency", "paper_url": "https://arxiv.org/abs/2601.02955", "detail_path": "reproductions/2601.02955-harmonrank/README.md", "topic": [ "multi-objective-ranking", "rank-consistency", "ensemble" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-01-06", "code": null, "adapter": { "adapter_key": "harmonrank", "arxiv_id": "2601.02955", "title": "Rethinking Multi-objective Ranking Ensemble in Recommender System: From Score Fusion to Rank Consistency", "paper_url": "https://arxiv.org/abs/2601.02955", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-01-06", "code_url": null, "topics": [ "multi-objective-ranking", "rank-consistency", "ensemble" ], "local_code_dir": "src/auto_research/reproductions/harmonrank", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou live e-commerce", "metric": "Purchase", "lift_percent": 2.635, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2601.02955v1", "source_location": "Section 5.7", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "reward:rank-consistency" ] } }, { "domain": "recommendation", "key": "harness-lm", "title": "HARNESS-LM: A Three-Phase Training Recipe for Harnessing SLMs in Sponsored Search Retrieval", "paper_url": "https://arxiv.org/abs/2605.23572", "detail_path": "reproductions/2605.23572-harness-lm/README.md", "topic": [ "llm-recommendation", "retrieval", "distillation" ], "first_author": null, "first_author_affiliation": "Microsoft AI / Bing Ads", "published": "2026-05-22", "code": null, "adapter": { "adapter_key": "harness-lm", "arxiv_id": "2605.23572", "title": "HARNESS-LM: A Three-Phase Training Recipe for Harnessing SLMs in Sponsored Search Retrieval", "paper_url": "https://arxiv.org/abs/2605.23572", "track": "recommendation", "organization": "Microsoft AI / Bing Ads", "published": "2026-05-22", "code_url": null, "topics": [ "llm-recommendation", "retrieval", "distillation" ], "local_code_dir": "src/auto_research/reproductions/harness_lm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "movielens 100k" ], "baseline": "小型对称内容 retriever", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Bing Ads", "metric": "Revenue", "lift_percent": 1.0, "traffic": "production deployment", "source_url": "https://arxiv.org/abs/2605.23572", "source_location": "original paper online-result disclosure: product=Bing Ads; metric=Revenue" }, { "product": "Bing Ads", "metric": "Clicks", "lift_percent": 0.4, "traffic": "production deployment", "source_url": "https://arxiv.org/abs/2605.23572", "source_location": "original paper online-result disclosure: product=Bing Ads; metric=Clicks" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "hcub", "title": "Hierarchical Contextual Uplift Bandits for Catalog Personalization", "paper_url": "https://arxiv.org/abs/2601.14333", "detail_path": "reproductions/2601.14333-hcub/README.md", "topic": [ "bandit", "uplift-modeling", "catalog-personalization" ], "first_author": null, "first_author_affiliation": "Dream11", "published": "2026-01-20", "code": null, "adapter": { "adapter_key": "hcub", "arxiv_id": "2601.14333", "title": "Hierarchical Contextual Uplift Bandits for Catalog Personalization", "paper_url": "https://arxiv.org/abs/2601.14333", "track": "recommendation", "organization": "Dream11", "published": "2026-01-20", "code_url": null, "topics": [ "bandit", "uplift-modeling", "catalog-personalization" ], "local_code_dir": "src/auto_research/reproductions/hcub", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Dream11 catalog personalization", "metric": "Revenue", "lift_percent": 0.51, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2601.14333v1", "source_location": "Introduction and online experiments", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "context:hierarchical-uplift" ] } }, { "domain": "recommendation", "key": "higr", "title": "HiGR: Industrial-Scale Hierarchical Generative Slate Recommendation Framework in Tencent", "paper_url": "https://arxiv.org/abs/2512.24787", "detail_path": "reproductions/2512.24787-higr/README.md", "topic": [ "generative-recommendation", "slate-recommendation", "semantic-id", "preference-optimization" ], "first_author": null, "first_author_affiliation": "Tencent", "published": "2025-12-31", "code": null, "adapter": { "adapter_key": "higr", "arxiv_id": "2512.24787", "title": "HiGR: Industrial-Scale Hierarchical Generative Slate Recommendation Framework in Tencent", "paper_url": "https://arxiv.org/abs/2512.24787", "track": "recommendation", "organization": "Tencent", "published": "2025-12-31", "code_url": null, "topics": [ "generative-recommendation", "slate-recommendation", "semantic-id", "preference-optimization" ], "local_code_dir": "src/auto_research/reproductions/higr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "flat item-level slate ranking", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.mrr_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.mrr_at_10", "method.ndcg_at_10", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.mrr_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Tencent video", "metric": "stay time", "lift_percent": 1.03, "traffic": "5% traffic", "source_url": "https://arxiv.org/abs/2512.24787", "source_location": "original paper online-result disclosure: product=Tencent video; metric=stay time" }, { "product": "Tencent video", "metric": "video views", "lift_percent": 1.73, "traffic": "5% traffic", "source_url": "https://arxiv.org/abs/2512.24787", "source_location": "original paper online-result disclosure: product=Tencent video; metric=video views" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "hill-index", "title": "Efficient Retrieval Scaling with Hierarchical Indexing for Large Scale Recommendation", "paper_url": "https://arxiv.org/abs/2604.12965", "detail_path": "reproductions/2604.12965-hill-index/README.md", "topic": [ "retrieval", "hierarchical-index", "residual-quantization", "test-time-training", "serving" ], "first_author": null, "first_author_affiliation": "Meta / Facebook and Instagram Ads", "published": "2026-04-14", "code": null, "adapter": { "adapter_key": "hill-index", "arxiv_id": "2604.12965", "title": "Efficient Retrieval Scaling with Hierarchical Indexing for Large Scale Recommendation", "paper_url": "https://arxiv.org/abs/2604.12965", "track": "recommendation", "organization": "Meta / Facebook and Instagram Ads", "published": "2026-04-14", "code_url": null, "topics": [ "retrieval", "hierarchical-index", "residual-quantization", "test-time-training", "serving" ], "local_code_dir": "src/auto_research/reproductions/hill_index", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (220 users / 360 items)" ], "baseline": "one-layer learned index with the same public item features and retrieval protocol", "metrics": [ "hit_at_10", "ndcg_at_10", "head_share_at_10", "mean_scored_items" ], "default_seeds": [ 42, 43, 44 ], "budget": "6 coarse nodes; up to 6 residual child nodes per parent", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta Ads Retrieval", "metric": "online ads metric", "lift_percent": 2.57, "traffic": "production online A/B: 2-layer MoNN Large/Small versus MoNN Small", "source_url": "https://arxiv.org/html/2604.12965v1#S4.SS5", "source_location": "Section 4.5, Table 8", "retrieved_at": "2026-09-01" } ], "selection_exception": null, "evolve_operators": [ "rankmixer_hill_index" ] } }, { "domain": "recommendation", "key": "hisac", "title": "HiSAC: Hierarchical Sparse Activation Compression for Ultra-long Sequence Modeling in Recommenders", "paper_url": "https://arxiv.org/abs/2602.21009", "detail_path": "reproductions/2602.21009-hisac/README.md", "topic": [ "long-sequence", "sparse-activation", "semantic-id" ], "first_author": null, "first_author_affiliation": "Alibaba / Taobao", "published": "2026-02-24", "code": null, "adapter": { "adapter_key": "hisac", "arxiv_id": "2602.21009", "title": "HiSAC: Hierarchical Sparse Activation Compression for Ultra-long Sequence Modeling in Recommenders", "paper_url": "https://arxiv.org/abs/2602.21009", "track": "recommendation", "organization": "Alibaba / Taobao", "published": "2026-02-24", "code_url": null, "topics": [ "long-sequence", "sparse-activation", "semantic-id" ], "local_code_dir": "src/auto_research/reproductions/hisac", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "recent-window dense proxy", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao recommendation", "metric": "CTR", "lift_percent": 1.65, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2602.21009", "source_location": "original paper online-result disclosure: product=Taobao recommendation; metric=CTR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "hpgr", "title": "Beyond the Flat Sequence: Hierarchical and Preference-Aware Generative Recommendations", "paper_url": "https://arxiv.org/abs/2603.00980", "detail_path": "reproductions/2603.00980-hpgr/README.md", "topic": [ "generative-recommendation", "hierarchical-sequence", "sparse-attention" ], "first_author": null, "first_author_affiliation": "Harbin Institute of Technology", "published": "2026-03-01", "code": null, "adapter": { "adapter_key": "hpgr", "arxiv_id": "2603.00980", "title": "Beyond the Flat Sequence: Hierarchical and Preference-Aware Generative Recommendations", "paper_url": "https://arxiv.org/abs/2603.00980", "track": "recommendation", "organization": "Harbin Institute of Technology", "published": "2026-03-01", "code_url": null, "topics": [ "generative-recommendation", "hierarchical-sequence", "sparse-attention" ], "local_code_dir": "src/auto_research/reproductions/hpgr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Huawei AppGallery recommendation", "metric": "eCPM", "lift_percent": 1.99, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2603.00980v1", "source_location": "Section 1 and online experiments", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "context:hierarchical-preference" ] } }, { "domain": "recommendation", "key": "hrpo", "title": "Hierarchical Residual Policy Optimization for Generative Recommendations", "paper_url": "https://arxiv.org/abs/2608.00750", "detail_path": "reproductions/2608.00750-hrpo/README.md", "topic": [ "generative-recommendation", "reinforcement-learning", "semantic-id", "credit-assignment" ], "first_author": null, "first_author_affiliation": "City University of Hong Kong / Kuaishou", "published": "2026-08-01", "code": "https://github.com/Applied-Machine-Learning-Lab/KDD2026-HRPO", "adapter": { "adapter_key": "hrpo", "arxiv_id": "2608.00750", "title": "Hierarchical Residual Policy Optimization for Generative Recommendations", "paper_url": "https://arxiv.org/abs/2608.00750", "track": "recommendation", "organization": "City University of Hong Kong / Kuaishou", "published": "2026-08-01", "code_url": "https://github.com/Applied-Machine-Learning-Lab/KDD2026-HRPO", "topics": [ "generative-recommendation", "reinforcement-learning", "semantic-id", "credit-assignment" ], "local_code_dir": "src/auto_research/reproductions/hrpo", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "去掉论文特有机制、其余数据切分与预算相同的 matched control", "metrics": [ "paper_results.fiction_target_cost_percent", "paper_results.mini_game_target_cost_percent", "paper_results.short_drama_target_cost_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "variants.hrpo residual-return policy.hit_at_10", "variants.hrpo residual-return policy.ndcg_at_10", "variants.sid sft.hit_at_10", "variants.sid sft.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou Fiction IAA", "metric": "Target Cost", "lift_percent": 3.49, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2608.00750", "source_location": "original paper online-result disclosure: product=Kuaishou Fiction IAA; metric=Target Cost" }, { "product": "Kuaishou Mini-Game IAA", "metric": "Target Cost", "lift_percent": 0.186, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2608.00750", "source_location": "original paper online-result disclosure: product=Kuaishou Mini-Game IAA; metric=Target Cost" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "hstu", "title": "Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations", "paper_url": "https://arxiv.org/abs/2402.17152", "detail_path": "reproductions/2402.17152-hstu/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": "https://github.com/meta-recsys/generative-recommenders", "adapter": { "adapter_key": "hstu", "arxiv_id": "2402.17152", "title": "Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations", "paper_url": "https://arxiv.org/abs/2402.17152", "track": "recommendation", "organization": null, "published": null, "code_url": "https://github.com/meta-recsys/generative-recommenders", "topics": [], "local_code_dir": "src/auto_research/reproductions/hstu", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "维度、层数、heads、loss 和训练步数匹配的 SASRec", "metrics": [ "hstu.hit_at_10", "hstu.hit_at_10_std", "hstu.ndcg_at_10", "hstu.ndcg_at_10_std", "ndcg_gain_percent", "sasrec.hit_at_10", "sasrec.hit_at_10_std", "sasrec.ndcg_at_10", "sasrec.ndcg_at_10_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=320", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "hyformer", "title": "HyFormer: Revisiting the Roles of Sequence Modeling and Feature Interaction in CTR Prediction", "paper_url": "https://arxiv.org/abs/2601.12681", "detail_path": "reproductions/2601.12681-hyformer/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "hyformer", "arxiv_id": "2601.12681", "title": "HyFormer: Revisiting the Roles of Sequence Modeling and Feature Interaction in CTR Prediction", "paper_url": "https://arxiv.org/abs/2601.12681", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/hyformer", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens 100K" ], "baseline": "独立 sequence Transformer+dense MLP 的 Late Fusion", "metrics": [ "ndcg_gain_percent", "results.hyformer.hit_at_10", "results.hyformer.hit_at_10_std", "results.hyformer.ndcg_at_10", "results.hyformer.ndcg_at_10_std", "results.late_fusion.hit_at_10", "results.late_fusion.hit_at_10_std", "results.late_fusion.ndcg_at_10", "results.late_fusion.ndcg_at_10_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=240", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "idproxy", "title": "IDProxy: Cold-Start CTR Prediction for Ads and Recommendation at Xiaohongshu with Multimodal LLMs", "paper_url": "https://arxiv.org/abs/2603.01590", "detail_path": "reproductions/2603.01590-idproxy/README.md", "topic": [ "multimodal-llm", "id-alignment", "ranking" ], "first_author": null, "first_author_affiliation": "Xiaohongshu / Shanghai Jiao Tong University / Fudan University", "published": "2026-03-02", "code": null, "adapter": { "adapter_key": "idproxy", "arxiv_id": "2603.01590", "title": "IDProxy: Cold-Start CTR Prediction for Ads and Recommendation at Xiaohongshu with Multimodal LLMs", "paper_url": "https://arxiv.org/abs/2603.01590", "track": "recommendation", "organization": "Xiaohongshu / Shanghai Jiao Tong University / Fudan University", "published": "2026-03-02", "code_url": null, "topics": [ "multimodal-llm", "id-alignment", "ranking" ], "local_code_dir": "src/auto_research/reproductions/idproxy", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (260 users / 420 items)" ], "baseline": "ID-only ranker，实验组为 coarse-to-fine IDProxy", "metrics": [ "alignment.contrastive_final_loss", "alignment.contrastive_initial_loss", "baseline.fresh_hit_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "steps=40", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Xiaohongshu content feed", "metric": "Engagement", "lift_percent": 0.5, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2603.01590", "source_location": "original paper online-result disclosure: product=Xiaohongshu content feed; metric=Engagement" }, { "product": "Xiaohongshu ads", "metric": "ADVV", "lift_percent": 1.93, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2603.01590", "source_location": "original paper online-result disclosure: product=Xiaohongshu ads; metric=ADVV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "incrementality", "title": "From Prediction to Incrementality: Causal Optimization for Large-Scale Targeting and Recommendation", "paper_url": "https://arxiv.org/abs/2608.10182", "detail_path": "reproductions/2608.10182-incrementality/README.md", "topic": [ "causal-recommendation", "uplift", "bandit", "constrained-allocation" ], "first_author": null, "first_author_affiliation": "LinkedIn", "published": "2026-08-10", "code": null, "adapter": { "adapter_key": "incrementality", "arxiv_id": "2608.10182", "title": "From Prediction to Incrementality: Causal Optimization for Large-Scale Targeting and Recommendation", "paper_url": "https://arxiv.org/abs/2608.10182", "track": "recommendation", "organization": "LinkedIn", "published": "2026-08-10", "code_url": null, "topics": [ "causal-recommendation", "uplift", "bandit", "constrained-allocation" ], "local_code_dir": "src/auto_research/reproductions/incrementality", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K causal targeting simulation" ], "baseline": "predictive treated-outcome targeting at the same budget", "metrics": [ "policy_value", "total_incremental_value", "uplift_rank_correlation" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users; 30% treatment budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "LinkedIn Feed marketing traffic", "metric": "primary long-term-value KPI", "lift_percent": 7.2, "traffic": "online A/B test", "source_url": "https://arxiv.org/html/2608.10182v1", "source_location": "Section 4 online evaluation", "significance": "p = 0.041", "retrieved_at": "2026-09-01" } ], "selection_exception": null, "evolve_operators": [ "reward:incrementality" ] } }, { "domain": "recommendation", "key": "inthq", "title": "IntHQ: Task-Interactive Hierarchical Query on Dual-Stream Representations for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2608.09634", "detail_path": "reproductions/2608.09634-inthq/README.md", "topic": [ "multi-task-learning", "generative-recommendation", "long-sequence" ], "first_author": null, "first_author_affiliation": "Amap / Alibaba", "published": "2026-08-10", "code": null, "adapter": { "adapter_key": "inthq", "arxiv_id": "2608.09634", "title": "IntHQ: Task-Interactive Hierarchical Query on Dual-Stream Representations for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2608.09634", "track": "recommendation", "organization": "Amap / Alibaba", "published": "2026-08-10", "code_url": null, "topics": [ "multi-task-learning", "generative-recommendation", "long-sequence" ], "local_code_dir": "src/auto_research/reproductions/inthq", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Amap", "metric": "UVCTR", "lift_percent": 1.6, "traffic": "20% per bucket, 7 days", "source_url": "https://arxiv.org/html/2608.09634v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "intsr", "title": "IntSR: An Integrated Generative Framework for Search and Recommendation", "paper_url": "https://arxiv.org/abs/2509.21179", "detail_path": "reproductions/2509.21179-intsr/README.md", "topic": [ "search", "recommendation", "generative-retrieval", "content-understanding" ], "first_author": null, "first_author_affiliation": "Alibaba / Amap", "published": "2025-09-25", "code": null, "adapter": { "adapter_key": "intsr", "arxiv_id": "2509.21179", "title": "IntSR: An Integrated Generative Framework for Search and Recommendation", "paper_url": "https://arxiv.org/abs/2509.21179", "track": "recommendation", "organization": "Alibaba / Amap", "published": "2025-09-25", "code_url": null, "topics": [ "search", "recommendation", "generative-retrieval", "content-understanding" ], "local_code_dir": "src/auto_research/reproductions/intsr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享 transition + content scorer，实验组只加入 intsr 核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "paper_results.gmv_percent", "paper_results.poi_ctr_percent", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Amap", "metric": "GMV", "lift_percent": 9.34, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2509.21179", "source_location": "original paper online-result disclosure: product=Amap; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "journeyformer", "title": "JourneyFormer: Encoding Airbnb Guest Journey with Sequence Modeling", "paper_url": "https://arxiv.org/abs/2606.19108", "detail_path": "reproductions/2606.19108-journeyformer/README.md", "topic": [ "search-ranking", "sequence-modeling", "long-sequence" ], "first_author": null, "first_author_affiliation": "Airbnb", "published": "2026-06-17", "code": null, "adapter": { "adapter_key": "journeyformer", "arxiv_id": "2606.19108", "title": "JourneyFormer: Encoding Airbnb Guest Journey with Sequence Modeling", "paper_url": "https://arxiv.org/abs/2606.19108", "track": "recommendation", "organization": "Airbnb", "published": "2026-06-17", "code_url": null, "topics": [ "search-ranking", "sequence-modeling", "long-sequence" ], "local_code_dir": "src/auto_research/reproductions/journeyformer", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Airbnb Search", "metric": "bookers", "lift_percent": 0.55, "traffic": "3-week A/B; p<0.01", "source_url": "https://arxiv.org/html/2606.19108v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "kar", "title": "Towards Open-World Recommendation with Knowledge Augmentation from Large Language Models", "paper_url": "https://arxiv.org/abs/2306.10933", "detail_path": "reproductions/2306.10933-kar/README.md", "topic": [ "llm-recommendation", "knowledge-augmentation", "ctr-ranking" ], "first_author": null, "first_author_affiliation": "Huawei", "published": "2023-06", "code": null, "adapter": { "adapter_key": "kar", "arxiv_id": "2306.10933", "title": "Towards Open-World Recommendation with Knowledge Augmentation from Large Language Models", "paper_url": "https://arxiv.org/abs/2306.10933", "track": "recommendation", "organization": "Huawei", "published": "2023-06", "code_url": null, "topics": [ "llm-recommendation", "knowledge-augmentation", "ctr-ranking" ], "local_code_dir": "src/auto_research/reproductions/kar", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens 100K title/genre text" ], "baseline": "纯 ID ranker", "metrics": [ "auc.id_ranker.mean", "auc.id_ranker.std", "auc.kar_hybrid_expert.mean", "auc.kar_hybrid_expert.std", "auc.mean_absolute_gain", "auc.mean_relative_gain_percent" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=100", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Huawei News", "metric": "recall", "lift_percent": 7.0, "traffic": "online A/B test", "source_url": "https://arxiv.org/abs/2306.10933", "source_location": "original paper online-result disclosure: product=Huawei News; metric=recall" }, { "product": "Huawei Music", "metric": "song play count", "lift_percent": 1.7, "traffic": "10% treatment/control, 7 days", "source_url": "https://arxiv.org/abs/2306.10933", "source_location": "original paper online-result disclosure: product=Huawei Music; metric=song play count" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "kgd", "title": "Knowledge–Geometry Decoupling: Refreshable Pretrained Transfer for Streaming Recommendation", "paper_url": "https://arxiv.org/abs/2608.02738", "detail_path": "reproductions/2608.02738-kgd/README.md", "topic": [ "sequential-recommendation", "pretraining", "streaming", "transfer" ], "first_author": null, "first_author_affiliation": "Xiamen University / Shopee", "published": "2026-08-03", "code": "https://github.com/FuCongResearchSquad/KGD4REC", "adapter": { "adapter_key": "kgd", "arxiv_id": "2608.02738", "title": "Knowledge–Geometry Decoupling: Refreshable Pretrained Transfer for Streaming Recommendation", "paper_url": "https://arxiv.org/abs/2608.02738", "track": "recommendation", "organization": "Xiamen University / Shopee", "published": "2026-08-03", "code_url": "https://github.com/FuCongResearchSquad/KGD4REC", "topics": [ "sequential-recommendation", "pretraining", "streaming", "transfer" ], "local_code_dir": "src/auto_research/reproductions/kgd", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "去掉论文特有机制、其余数据切分与预算相同的 matched control", "metrics": [ "paper_results.advertising_revenue_percent", "paper_results.gmv_per_user_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "variants.adjacent ntp transfer.hit_at_10", "variants.adjacent ntp transfer.ndcg_at_10", "variants.kgd bmtp + read-only transfer + acr.hit_at_10", "variants.kgd bmtp + read-only transfer + acr.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Shopee Homepage Search", "metric": "GMV per user", "lift_percent": 1.75, "traffic": "live A/B; fully deployed", "source_url": "https://arxiv.org/abs/2608.02738", "source_location": "original paper online-result disclosure: product=Shopee Homepage Search; metric=GMV per user" }, { "product": "Shopee Homepage Search", "metric": "advertising revenue", "lift_percent": 1.53, "traffic": "live A/B; fully deployed", "source_url": "https://arxiv.org/abs/2608.02738", "source_location": "original paper online-result disclosure: product=Shopee Homepage Search; metric=advertising revenue" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "kunlun", "title": "Kunlun: Establishing Scaling Laws for Massive-Scale Recommendation Systems through Unified Architecture Design", "paper_url": "https://arxiv.org/abs/2602.10016", "detail_path": "reproductions/2602.10016-kunlun/README.md", "topic": [ "ranking", "scaling-law", "long-sequence", "mixture-of-experts", "serving" ], "first_author": null, "first_author_affiliation": "Meta", "published": "2026-02-10", "code": null, "adapter": { "adapter_key": "kunlun", "arxiv_id": "2602.10016", "title": "Kunlun: Establishing Scaling Laws for Massive-Scale Recommendation Systems through Unified Architecture Design", "paper_url": "https://arxiv.org/abs/2602.10016", "track": "recommendation", "organization": "Meta", "published": "2026-02-10", "code_url": null, "topics": [ "ranking", "scaling-law", "long-sequence", "mixture-of-experts", "serving" ], "local_code_dir": "src/auto_research/reproductions/kunlun", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (220 users / 360 items)" ], "baseline": "shallow transition-content ranker，实验组执行 Kunlun compact core", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta Ads", "metric": "topline", "lift_percent": 1.2, "traffic": "deployed across major Ads models", "source_url": "https://arxiv.org/html/2602.10016v3#S5.SS5", "source_location": "Section 5.5", "retrieved_at": "2026-08-09" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "l2rec", "title": "L2Rec: Towards Dual-View Understanding of LLMs for Personalized Recommendation", "paper_url": "https://arxiv.org/abs/2605.26717", "detail_path": "reproductions/2605.26717-l2rec/README.md", "topic": [ "llm-recommendation", "parameter-efficient-tuning", "multi-view" ], "first_author": null, "first_author_affiliation": "NetEase Cloud Music", "published": "2026-05-26", "code": null, "adapter": { "adapter_key": "l2rec", "arxiv_id": "2605.26717", "title": "L2Rec: Towards Dual-View Understanding of LLMs for Personalized Recommendation", "paper_url": "https://arxiv.org/abs/2605.26717", "track": "recommendation", "organization": "NetEase Cloud Music", "published": "2026-05-26", "code_url": null, "topics": [ "llm-recommendation", "parameter-efficient-tuning", "multi-view" ], "local_code_dir": "src/auto_research/reproductions/l2rec", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "homepage feed", "metric": "CTR", "lift_percent": 9.24, "traffic": "6% user traffic, one month; p<0.01", "source_url": "https://arxiv.org/html/2605.26717v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "lazformer", "title": "LazFormer: Scaling Transformers for Industrial Recommendation via Transferable Generative Pre-training", "paper_url": "https://arxiv.org/abs/2609.14978", "detail_path": "reproductions/2609.14978-lazformer/README.md", "topic": [ "ranking", "generative-pretraining", "residual-adapter", "hybrid-sparse-attention" ], "first_author": null, "first_author_affiliation": "Alibaba International Digital Commerce Group", "published": "2026-09-14", "code": null, "adapter": { "adapter_key": "lazformer", "arxiv_id": "2609.14978", "title": "LazFormer: Scaling Transformers for Industrial Recommendation via Transferable Generative Pre-training", "paper_url": "https://arxiv.org/abs/2609.14978", "track": "recommendation", "organization": "Alibaba International Digital Commerce Group", "published": "2026-09-14", "code_url": null, "topics": [ "ranking", "generative-pretraining", "residual-adapter", "hybrid-sparse-attention" ], "local_code_dir": "src/auto_research/reproductions/lazformer", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "large-scale e-commerce homepage ranking", "metric": "IPV", "lift_percent": 5.21, "traffic": "two-week production A/B", "source_url": "https://arxiv.org/html/2609.14978v1", "source_location": "Section 4.7; Table 7", "experiment_duration": "two weeks", "retrieved_at": "2026-09-16" }, { "product": "large-scale e-commerce homepage ranking", "metric": "orders", "lift_percent": 3.38, "traffic": "two-week production A/B", "source_url": "https://arxiv.org/html/2609.14978v1", "source_location": "Section 4.7; Table 7", "experiment_duration": "two weeks", "retrieved_at": "2026-09-16" }, { "product": "large-scale e-commerce homepage ranking", "metric": "buyers", "lift_percent": 3.7, "traffic": "two-week production A/B", "source_url": "https://arxiv.org/html/2609.14978v1", "source_location": "Section 4.7; Table 7", "experiment_duration": "two weeks", "retrieved_at": "2026-09-16" }, { "product": "large-scale e-commerce homepage ranking", "metric": "GMV", "lift_percent": 9.85, "traffic": "two-week production A/B", "source_url": "https://arxiv.org/html/2609.14978v1", "source_location": "Section 4.7; Table 7", "experiment_duration": "two weeks", "retrieved_at": "2026-09-16" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "leadre", "title": "LEADRE: Multi-Faceted Knowledge Enhanced LLM Empowered Display Advertisement Recommender System", "paper_url": "https://arxiv.org/abs/2411.13789", "detail_path": "reproductions/2411.13789-leadre/README.md", "topic": [ "llm-recommendation", "generative-retrieval", "advertising", "preference-alignment" ], "first_author": null, "first_author_affiliation": "Tencent / WeChat", "published": "2024-11", "code": null, "adapter": { "adapter_key": "leadre", "arxiv_id": "2411.13789", "title": "LEADRE: Multi-Faceted Knowledge Enhanced LLM Empowered Display Advertisement Recommender System", "paper_url": "https://arxiv.org/abs/2411.13789", "track": "recommendation", "organization": "Tencent / WeChat", "published": "2024-11", "code_url": null, "topics": [ "llm-recommendation", "generative-retrieval", "advertising", "preference-alignment" ], "local_code_dir": "src/auto_research/reproductions/leadre", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens-100K" ], "baseline": "DIN，LEADRE NDCG@10 相对 DIN +12.94%", "metrics": [ "paper_online_ab.channels_gmv_percent", "paper_online_ab.moments_gmv_percent", "results.din.hit_at_10", "results.din.ndcg_at_10", "results.leadre_dpo.hit_at_10", "results.leadre_dpo.ndcg_at_10", "results.ndcg_ablation_percent", "results.ndcg_vs_din_percent", "results.sft.hit_at_10", "results.sft.ndcg_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=100", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "WeChat Channels", "metric": "GMV", "lift_percent": 1.57, "traffic": "production online A/B", "source_url": "https://arxiv.org/abs/2411.13789", "source_location": "original paper online-result disclosure: product=WeChat Channels; metric=GMV" }, { "product": "WeChat Moments", "metric": "GMV", "lift_percent": 1.17, "traffic": "production online A/B", "source_url": "https://arxiv.org/abs/2411.13789", "source_location": "original paper online-result disclosure: product=WeChat Moments; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "learn", "title": "Knowledge Adaptation from Large Language Model to Recommendation for Practical Industrial Application", "paper_url": "https://arxiv.org/abs/2405.03988", "detail_path": "reproductions/2405.03988-learn/README.md", "topic": [ "llm-recommendation", "knowledge-adaptation", "cold-start" ], "first_author": null, "first_author_affiliation": "Kuaishou", "published": "2024-05", "code": null, "adapter": { "adapter_key": "learn", "arxiv_id": "2405.03988", "title": "Knowledge Adaptation from Large Language Model to Recommendation for Practical Industrial Application", "paper_url": "https://arxiv.org/abs/2405.03988", "track": "recommendation", "organization": "Kuaishou", "published": "2024-05", "code_url": null, "topics": [ "llm-recommendation", "knowledge-adaptation", "cold-start" ], "local_code_dir": "src/auto_research/reproductions/learn", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens 100K title/genre text" ], "baseline": "frozen LLM semantic mean", "metrics": [ "hit_at_10.frozen_llm_semantic_mean.mean", "hit_at_10.frozen_llm_semantic_mean.std", "hit_at_10.learn_ceg_pch.mean", "hit_at_10.learn_ceg_pch.std", "hit_at_10.relative_change_percent", "ndcg_at_10.frozen_llm_semantic_mean.mean", "ndcg_at_10.frozen_llm_semantic_mean.std", "ndcg_at_10.learn_ceg_pch.mean", "ndcg_at_10.learn_ceg_pch.std", "ndcg_at_10.relative_change_percent" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=140", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou e-commerce ads", "metric": "cold-start item revenue", "lift_percent": 8.77, "traffic": "20% traffic, 9 days", "source_url": "https://arxiv.org/abs/2405.03988", "source_location": "original paper online-result disclosure: product=Kuaishou e-commerce ads; metric=cold-start item revenue" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "linkedin-feed-sr", "title": "An Industrial-Scale Sequential Recommender for LinkedIn Feed Ranking", "paper_url": "https://arxiv.org/abs/2602.12354", "detail_path": "reproductions/2602.12354-linkedin-feed-sr/README.md", "topic": [ "ranking", "sequential-recommendation", "long-sequence" ], "first_author": null, "first_author_affiliation": "LinkedIn", "published": "2026-02-12", "code": null, "adapter": { "adapter_key": "linkedin-feed-sr", "arxiv_id": "2602.12354", "title": "An Industrial-Scale Sequential Recommender for LinkedIn Feed Ranking", "paper_url": "https://arxiv.org/abs/2602.12354", "track": "recommendation", "organization": "LinkedIn", "published": "2026-02-12", "code_url": null, "topics": [ "ranking", "sequential-recommendation", "long-sequence" ], "local_code_dir": "src/auto_research/reproductions/linkedin_feed_sr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "LinkedIn Feed", "metric": "time spent", "lift_percent": 2.1, "traffic": "production A/B", "source_url": "https://arxiv.org/html/2602.12354v1", "source_location": "Section 7 / Table 5", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "llatte", "title": "LLaTTE: Scaling Laws for Multi-Stage Sequence Modeling in Large-Scale Ads Recommendation", "paper_url": "https://arxiv.org/abs/2601.20083", "detail_path": "reproductions/2601.20083-llatte/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "llatte", "arxiv_id": "2601.20083", "title": "LLaTTE: Scaling Laws for Multi-Stage Sequence Modeling in Large-Scale Ads Recommendation", "paper_url": "https://arxiv.org/abs/2601.20083", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/llatte", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "相同 ID/语义 embedding 的 short online sequence", "metrics": [ "relative_ndcg_change_percent", "results.llatte_mla_dhen.hit_at_10", "results.llatte_mla_dhen.hit_at_10_std", "results.llatte_mla_dhen.ndcg_at_10", "results.llatte_mla_dhen.ndcg_at_10_std", "results.short_online_sequence.hit_at_10", "results.short_online_sequence.hit_at_10_std", "results.short_online_sequence.ndcg_at_10", "results.short_online_sequence.ndcg_at_10_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=75", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "llm-ad-retrieval", "title": "LLM Retrieval for Stable and Predictable Ad Recommendations", "paper_url": "https://arxiv.org/abs/2605.21969", "detail_path": "reproductions/2605.21969-llm-ad-retrieval/README.md", "topic": [ "llm-recommendation", "retrieval", "stability" ], "first_author": null, "first_author_affiliation": "Meta", "published": "2026-05", "code": null, "adapter": { "adapter_key": "llm-ad-retrieval", "arxiv_id": "2605.21969", "title": "LLM Retrieval for Stable and Predictable Ad Recommendations", "paper_url": "https://arxiv.org/abs/2605.21969", "track": "recommendation", "organization": "Meta", "published": "2026-05", "code_url": null, "topics": [ "llm-recommendation", "retrieval", "stability" ], "local_code_dir": "src/auto_research/reproductions/llm_ad_retrieval", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K titles, genres, and chronological positive feedback" ], "baseline": "Collaborative Item Graph", "metrics": [ "paper_online_ab.final_stage_recall_lift_percent", "results.collaborative_graph.ndcg_at_20", "results.collaborative_graph.recall_at_20", "results.collaborative_plus_llm_graph.ndcg_at_20", "results.collaborative_plus_llm_graph.recall_at_20", "results.ndcg_relative_gain_percent", "results.recall_relative_gain_percent", "training.attribute_head_final_loss", "training.attribute_head_initial_loss", "training.domain_sft_final_loss", "training.domain_sft_initial_loss" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=160", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta Ads", "metric": "top-line performance", "lift_percent": 0.45, "traffic": "large-scale industrial online A/B", "source_url": "https://arxiv.org/abs/2605.21969", "source_location": "original paper online-result disclosure: product=Meta Ads; metric=top-line performance" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "llm-ts-prior", "title": "LLM-Derived Priors for Thompson Sampling in Cold-Start Comment Recommendation", "paper_url": "https://arxiv.org/abs/2608.03382", "detail_path": "reproductions/2608.03382-llm-ts-prior/README.md", "topic": [ "llm-recommendation", "bandit", "cold-start", "comment-ranking" ], "first_author": null, "first_author_affiliation": "NAVER WEBTOON", "published": "2026-08-04", "code": null, "adapter": { "adapter_key": "llm-ts-prior", "arxiv_id": "2608.03382", "title": "LLM-Derived Priors for Thompson Sampling in Cold-Start Comment Recommendation", "paper_url": "https://arxiv.org/abs/2608.03382", "track": "recommendation", "organization": "NAVER WEBTOON", "published": "2026-08-04", "code_url": null, "topics": [ "llm-recommendation", "bandit", "cold-start", "comment-ranking" ], "local_code_dir": "src/auto_research/reproductions/llm_ts_prior", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "去掉论文特有机制、其余数据切分与预算相同的 matched control", "metrics": [ "paper_results.content_prior_overall_ctr_percent", "paper_results.gender_prior_cold_10_49_ctr_percent", "paper_results.gender_prior_overall_ctr_percent", "relative.cold_start_ctr_percent", "relative.ctr_percent", "variants.llm semantic prior.cold_start_ctr", "variants.llm semantic prior.ctr", "variants.uniform thompson prior.cold_start_ctr", "variants.uniform thompson prior.ctr" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "WEBTOON comment recommendation", "metric": "Gender-prior overall CTR", "lift_percent": 1.48, "traffic": "four-week A/B/C; about 595K users per arm", "significance": "p=0.144, not significant", "source_url": "https://arxiv.org/abs/2608.03382", "source_location": "original paper online-result disclosure: product=WEBTOON comment recommendation; metric=Gender-prior overall CTR" }, { "product": "WEBTOON 10-49 exposure bucket", "metric": "Gender-prior CTR", "lift_percent": 9.51, "traffic": "A/B/C cold-start subgroup", "significance": "p<0.001", "source_url": "https://arxiv.org/abs/2608.03382", "source_location": "original paper online-result disclosure: product=WEBTOON 10-49 exposure bucket; metric=Gender-prior CTR" }, { "product": "WEBTOON overall", "metric": "Content-prior CTR", "lift_percent": -5.68, "traffic": "four-week A/B/C", "significance": "p<0.001", "source_url": "https://arxiv.org/abs/2608.03382", "source_location": "original paper online-result disclosure: product=WEBTOON overall; metric=Content-prior CTR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "long-history-transformer", "title": "Long-History User Transformers for Real-Time Ad Ranking", "paper_url": "https://arxiv.org/abs/2607.14331", "detail_path": "reproductions/2607.14331-long-history-transformer/README.md", "topic": [ "advertising", "long-sequence", "cached-user-representation", "serving" ], "first_author": null, "first_author_affiliation": "Yandex", "published": "2026-07-15", "code": null, "adapter": { "adapter_key": "long-history-transformer", "arxiv_id": "2607.14331", "title": "Long-History User Transformers for Real-Time Ad Ranking", "paper_url": "https://arxiv.org/abs/2607.14331", "track": "recommendation", "organization": "Yandex", "published": "2026-07-15", "code_url": null, "topics": [ "advertising", "long-sequence", "cached-user-representation", "serving" ], "local_code_dir": "src/auto_research/reproductions/long_history_transformer", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "request-time recent-history Transformer，实验组为缓存全历史加轻量在线 Transformer", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Yandex Search Ads", "metric": "Primary ranking metric", "lift_percent": 2.77, "traffic": "production A/B; p<0.05", "source_url": "https://arxiv.org/abs/2607.14331", "source_location": "original paper online-result disclosure: product=Yandex Search Ads; metric=Primary ranking metric" }, { "product": "Yandex Search Ads", "metric": "Revenue", "lift_percent": 2.26, "traffic": "production A/B; p<0.05", "source_url": "https://arxiv.org/abs/2607.14331", "source_location": "original paper online-result disclosure: product=Yandex Search Ads; metric=Revenue" }, { "product": "Yandex Advertising Network", "metric": "Primary ranking metric", "lift_percent": 2.1, "traffic": "production A/B; p<0.05", "source_url": "https://arxiv.org/abs/2607.14331", "source_location": "original paper online-result disclosure: product=Yandex Advertising Network; metric=Primary ranking metric" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "longer", "title": "LONGER: Scaling Up Long Sequence Modeling in Industrial Recommenders", "paper_url": "https://arxiv.org/abs/2505.04421", "detail_path": "reproductions/2505.04421-longer/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "longer", "arxiv_id": "2505.04421", "title": "LONGER: Scaling Up Long Sequence Modeling in Industrial Recommenders", "paper_url": "https://arxiv.org/abs/2505.04421", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/longer", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "recent sequence transformer", "metrics": [ "relative_ndcg_change_percent", "results.longer_hybrid_attention.hit_at_10", "results.longer_hybrid_attention.hit_at_10_std", "results.longer_hybrid_attention.ndcg_at_10", "results.longer_hybrid_attention.ndcg_at_10_std", "results.recent_sequence_transformer.hit_at_10", "results.recent_sequence_transformer.hit_at_10_std", "results.recent_sequence_transformer.ndcg_at_10", "results.recent_sequence_transformer.ndcg_at_10_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=90", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "lsvcr", "title": "A Large Language Model Enhanced Sequential Recommender for Joint Video and Comment Recommendation", "paper_url": "https://arxiv.org/abs/2403.13574", "detail_path": "reproductions/2403.13574-lsvcr/README.md", "topic": [ "LLM recommendation", "comment recommendation", "preference alignment" ], "first_author": null, "first_author_affiliation": "Kuaishou", "published": "2024-03", "code": "https://github.com/RUCAIBox/LSVCR", "adapter": { "adapter_key": "lsvcr", "arxiv_id": "2403.13574", "title": "A Large Language Model Enhanced Sequential Recommender for Joint Video and Comment Recommendation", "paper_url": "https://arxiv.org/abs/2403.13574", "track": "recommendation", "organization": "Kuaishou", "published": "2024-03", "code_url": "https://github.com/RUCAIBox/LSVCR", "topics": [ "LLM recommendation", "comment recommendation", "preference alignment" ], "local_code_dir": "src/auto_research/reproductions/lsvcr", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "Amazon Beauty 5-core product interactions and real review text" ], "baseline": "相同 SR 网络但不做 LLM preference alignment", "metrics": [ "aggregate.baseline_comment_ndcg_mean", "aggregate.baseline_item_ndcg_mean", "aggregate.lsvcr_comment_ndcg_mean", "aggregate.lsvcr_item_ndcg_mean", "per_seed.42.baseline_comment_ndcg_at_10", "per_seed.42.baseline_item_ndcg_at_10", "per_seed.42.lsvcr_comment_ndcg_at_10", "per_seed.42.lsvcr_item_ndcg_at_10", "per_seed.43.baseline_comment_ndcg_at_10", "per_seed.43.baseline_item_ndcg_at_10", "per_seed.43.lsvcr_comment_ndcg_at_10", "per_seed.43.lsvcr_item_ndcg_at_10", "per_seed.44.baseline_comment_ndcg_at_10", "per_seed.44.baseline_item_ndcg_at_10", "per_seed.44.lsvcr_comment_ndcg_at_10", "per_seed.44.lsvcr_item_ndcg_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=120", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou video", "metric": "Watch Time", "lift_percent": 0.3649, "traffic": "20K users, 2 weeks", "source_url": "https://arxiv.org/abs/2403.13574", "source_location": "original paper online-result disclosure: product=Kuaishou video; metric=Watch Time" }, { "product": "Kuaishou comments", "metric": "Watch Time", "lift_percent": 4.1264, "traffic": "20K users, 2 weeks", "source_url": "https://arxiv.org/abs/2403.13574", "source_location": "original paper online-result disclosure: product=Kuaishou comments; metric=Watch Time" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "ltv-video-ranking", "title": "A Long-term Value Prediction Framework In Video Ranking", "paper_url": "https://arxiv.org/abs/2602.17058", "detail_path": "reproductions/2602.17058-ltv-video-ranking/README.md", "topic": [ "ranking", "long-term-value", "debiasing" ], "first_author": null, "first_author_affiliation": "Alibaba Group / Tsinghua University", "published": "2026-02-19", "code": null, "adapter": { "adapter_key": "ltv-video-ranking", "arxiv_id": "2602.17058", "title": "A Long-term Value Prediction Framework In Video Ranking", "paper_url": "https://arxiv.org/abs/2602.17058", "track": "recommendation", "organization": "Alibaba Group / Tsinghua University", "published": "2026-02-19", "code_url": null, "topics": [ "ranking", "long-term-value", "debiasing" ], "local_code_dir": "src/auto_research/reproductions/ltv_video_ranking", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao video ranking", "metric": "VV", "lift_percent": 2.49, "traffic": "multi-day production A/B", "source_url": "https://arxiv.org/html/2602.17058v1", "source_location": "Section 5.3 / Table 6", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "lum", "title": "Unlocking Scaling Law in Industrial Recommendation Systems with a Three-step Paradigm based Large User Model", "paper_url": "https://arxiv.org/abs/2502.08309", "detail_path": "reproductions/2502.08309-lum/README.md", "topic": [ "large user model", "generative pretraining", "knowledge querying", "CTR" ], "first_author": null, "first_author_affiliation": "Alibaba", "published": "2025-02", "code": null, "adapter": { "adapter_key": "lum", "arxiv_id": "2502.08309", "title": "Unlocking Scaling Law in Industrial Recommendation Systems with a Three-step Paradigm based Large User Model", "paper_url": "https://arxiv.org/abs/2502.08309", "track": "recommendation", "organization": "Alibaba", "published": "2025-02", "code_url": null, "topics": [ "large user model", "generative pretraining", "knowledge querying", "CTR" ], "local_code_dir": "src/auto_research/reproductions/lum", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens-1M ratings, genres, chronological behavior" ], "baseline": "ID-DLRM", "metrics": [ "aggregate.dlrm_auc_mean", "aggregate.dlrm_auc_std", "aggregate.lum_auc_mean", "aggregate.lum_auc_std", "per_seed.42.dlrm_auc", "per_seed.42.lum_auc", "per_seed.43.dlrm_auc", "per_seed.43.lum_auc", "per_seed.44.dlrm_auc", "per_seed.44.lum_auc" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=200", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao sponsored search", "metric": "CTR", "lift_percent": 2.9, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2502.08309", "source_location": "original paper online-result disclosure: product=Taobao sponsored search; metric=CTR" }, { "product": "Taobao sponsored search", "metric": "RPM", "lift_percent": 1.2, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2502.08309", "source_location": "original paper online-result disclosure: product=Taobao sponsored search; metric=RPM" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "lwgr", "title": "LWGR: Lagrangian-Constrained Personalized World Knowledge for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2605.18771", "detail_path": "reproductions/2605.18771-lwgr/README.md", "topic": [ "generative-recommendation", "soft-prompt", "world-knowledge", "lagrangian" ], "first_author": null, "first_author_affiliation": "Alibaba International / CAS", "published": "2026-05", "code": null, "adapter": { "adapter_key": "lwgr", "arxiv_id": "2605.18771", "title": "LWGR: Lagrangian-Constrained Personalized World Knowledge for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2605.18771", "track": "recommendation", "organization": "Alibaba International / CAS", "published": "2026-05", "code_url": null, "topics": [ "generative-recommendation", "soft-prompt", "world-knowledge", "lagrangian" ], "local_code_dir": "src/auto_research/reproductions/lwgr", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MiniOneRec Office" ], "baseline": "相同 Office SID、相同 48 维 GR backbone、seeds 42/43/44 的冻结 reference", "metrics": [ "selected_relative_to_reference_percent.ndcg_at_10", "selected_relative_to_reference_percent.recall_at_10", "test.lwgr.ndcg_at_10", "test.lwgr.ndcg_at_5", "test.lwgr.recall_at_10", "test.lwgr.recall_at_5", "test.reference.ndcg_at_10", "test.reference.ndcg_at_5", "test.reference.recall_at_10", "test.reference.recall_at_5", "test.unconstrained.ndcg_at_10", "test.unconstrained.ndcg_at_5", "test.unconstrained.recall_at_10", "test.unconstrained.recall_at_5" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=180", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Southeast Asia Ads", "metric": "Revenue", "lift_percent": 1.35, "traffic": "15% users, 2025-12-19 to 2025-12-30", "source_url": "https://arxiv.org/abs/2605.18771", "source_location": "original paper online-result disclosure: product=Southeast Asia Ads; metric=Revenue" }, { "product": "Southeast Asia Ads", "metric": "CTR", "lift_percent": 1.17, "traffic": "p<0.05", "source_url": "https://arxiv.org/abs/2605.18771", "source_location": "original paper online-result disclosure: product=Southeast Asia Ads; metric=CTR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "m6rec", "title": "M6-Rec: Generative Pretrained Language Models are Open-Ended Recommender Systems", "paper_url": "https://arxiv.org/abs/2205.08084", "detail_path": "reproductions/2205.08084-m6rec/README.md", "topic": [ "llm-recommendation", "foundation-model", "parameter-efficient-tuning" ], "first_author": null, "first_author_affiliation": "Alibaba", "published": "2022-05", "code": null, "adapter": { "adapter_key": "m6rec", "arxiv_id": "2205.08084", "title": "M6-Rec: Generative Pretrained Language Models are Open-Ended Recommender Systems", "paper_url": "https://arxiv.org/abs/2205.08084", "track": "recommendation", "organization": "Alibaba", "published": "2022-05", "code_url": null, "topics": [ "llm-recommendation", "foundation-model", "parameter-efficient-tuning" ], "local_code_dir": "src/auto_research/reproductions/m6rec", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K title and genre text" ], "baseline": "只训练 option prototype 的 Option Tuning", "metrics": [ "summary.option_adapter_auc_mean", "summary.option_adapter_auc_std", "summary.option_tuning_auc_mean", "summary.option_tuning_auc_std", "summary.relative_auc_gain_percent_mean", "summary.relative_auc_gain_percent_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=80", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Alipay", "metric": "mini-app retrieval CTR", "lift_percent": 1.0, "traffic": "live replacement; fully deployed", "source_url": "https://arxiv.org/abs/2205.08084", "source_location": "original paper online-result disclosure: product=Alipay; metric=mini-app retrieval CTR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "marc", "title": "Modular Representation Compression: Adapting LLMs for Efficient and Effective Recommendations", "paper_url": "https://arxiv.org/abs/2604.18146", "detail_path": "reproductions/2604.18146-marc/README.md", "topic": [ "llm-recommendation", "representation-compression", "modular-learning" ], "first_author": null, "first_author_affiliation": "Shanghai Jiao Tong University", "published": "2026-04-20", "code": null, "adapter": { "adapter_key": "marc", "arxiv_id": "2604.18146", "title": "Modular Representation Compression: Adapting LLMs for Efficient and Effective Recommendations", "paper_url": "https://arxiv.org/abs/2604.18146", "track": "recommendation", "organization": "Shanghai Jiao Tong University", "published": "2026-04-20", "code_url": null, "topics": [ "llm-recommendation", "representation-compression", "modular-learning" ], "local_code_dir": "src/auto_research/reproductions/marc", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Commercial search advertising", "metric": "eCPM", "lift_percent": 2.82, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2604.18146v1", "source_location": "Abstract and experiments", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "head:modular-compression" ] } }, { "domain": "recommendation", "key": "mbgr", "title": "MBGR: Multi-Business Generative Recommendation", "paper_url": "https://arxiv.org/abs/2604.02684", "detail_path": "reproductions/2604.02684-mbgr/README.md", "topic": [ "generative-recommendation", "multi-domain", "semantic-id" ], "first_author": null, "first_author_affiliation": "Meituan", "published": "2026-04-03", "code": null, "adapter": { "adapter_key": "mbgr", "arxiv_id": "2604.02684", "title": "MBGR: Multi-Business Generative Recommendation", "paper_url": "https://arxiv.org/abs/2604.02684", "track": "recommendation", "organization": "Meituan", "published": "2026-04-03", "code_url": null, "topics": [ "generative-recommendation", "multi-domain", "semantic-id" ], "local_code_dir": "src/auto_research/reproductions/mbgr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "movielens 100k" ], "baseline": "共享 next-item 模型", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meituan multi-business recommendation", "metric": "CTCVR", "lift_percent": 3.98, "traffic": "online A/B test", "source_url": "https://arxiv.org/abs/2604.02684", "source_location": "original paper online-result disclosure: product=Meituan multi-business recommendation; metric=CTCVR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "mdcns", "title": "Divergence Meets Consensus: A Multi-Source Negative Sampling Framework for Sequential Recommendation", "paper_url": "https://arxiv.org/abs/2605.19651", "detail_path": "reproductions/2605.19651-mdcns/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": "https://github.com/Lyz103/SIGIR26-MDCNS", "adapter": { "adapter_key": "mdcns", "arxiv_id": "2605.19651", "title": "Divergence Meets Consensus: A Multi-Source Negative Sampling Framework for Sequential Recommendation", "paper_url": "https://arxiv.org/abs/2605.19651", "track": "recommendation", "organization": null, "published": null, "code_url": "https://github.com/Lyz103/SIGIR26-MDCNS", "topics": [], "local_code_dir": "src/auto_research/reproductions/mdcns", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "Amazon Beauty 5-core authors' split" ], "baseline": "Uniform negative sampling，辅助基线是 DNS", "metrics": [ "ndcg_gain_vs_uniform_percent", "results.dns.hit_at_10", "results.dns.ndcg_at_10", "results.mdcns.hit_at_10", "results.mdcns.ndcg_at_10", "results.uniform.hit_at_10", "results.uniform.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "epochs=4", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "mdl", "title": "MDL: A Unified Multi-Distribution Learner in Large-scale Industrial Recommendation through Tokenization", "paper_url": "https://arxiv.org/abs/2602.07520", "detail_path": "reproductions/2602.07520-mdl/README.md", "topic": [ "ranking", "multi-scenario", "multi-task", "tokenization" ], "first_author": null, "first_author_affiliation": "ByteDance / Douyin", "published": "2026-02-07", "code": null, "adapter": { "adapter_key": "mdl", "arxiv_id": "2602.07520", "title": "MDL: A Unified Multi-Distribution Learner in Large-scale Industrial Recommendation through Tokenization", "paper_url": "https://arxiv.org/abs/2602.07520", "track": "recommendation", "organization": "ByteDance / Douyin", "published": "2026-02-07", "code_url": null, "topics": [ "ranking", "multi-scenario", "multi-task", "tokenization" ], "local_code_dir": "src/auto_research/reproductions/mdl", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享但非 tokenized ranker", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Douyin search", "metric": "LT30", "lift_percent": 0.0626, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2602.07520", "source_location": "original paper online-result disclosure: product=Douyin search; metric=LT30" }, { "product": "Douyin search", "metric": "query rewrite rate", "lift_percent": -0.3267, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2602.07520", "source_location": "original paper online-result disclosure: product=Douyin search; metric=query rewrite rate" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "melo", "title": "Melo: A Production LLM-Powered Music Recommendation Agent", "paper_url": "https://arxiv.org/abs/2607.23718", "detail_path": "reproductions/2607.23718-melo/README.md", "topic": [ "llm-recommendation", "agent", "entity-grounding", "reflection" ], "first_author": null, "first_author_affiliation": "NetEase Cloud Music / Zhejiang University of Technology", "published": "2026-07-26", "code": null, "adapter": { "adapter_key": "melo", "arxiv_id": "2607.23718", "title": "Melo: A Production LLM-Powered Music Recommendation Agent", "paper_url": "https://arxiv.org/abs/2607.23718", "track": "recommendation", "organization": "NetEase Cloud Music / Zhejiang University of Technology", "published": "2026-07-26", "code_url": null, "topics": [ "llm-recommendation", "agent", "entity-grounding", "reflection" ], "local_code_dir": "src/auto_research/reproductions/melo", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "无运行时修复的 catalog recommender，实验组为 grounding + retry", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.validation_only_model_selection.fresh_hit_at_10", "stages.validation_only_model_selection.hit_at_10", "stages.validation_only_model_selection.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "NetEase Cloud Music playlist surfaces", "metric": "playlist retention lower bound (pp)", "lift_percent": 2.0, "traffic": "one-month randomized system-level A/B", "source_url": "https://arxiv.org/abs/2607.23718", "source_location": "original paper online-result disclosure: product=NetEase Cloud Music playlist surfaces; metric=playlist retention lower bound (pp)" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "memento", "title": "Memento: Personalized RAG-Style Long-Retention Data Scaling for Online Ads Recommendation", "paper_url": "https://arxiv.org/abs/2605.24051", "detail_path": "reproductions/2605.24051-memento/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "memento", "arxiv_id": "2605.24051", "title": "Memento: Personalized RAG-Style Long-Retention Data Scaling for Online Ads Recommendation", "paper_url": "https://arxiv.org/abs/2605.24051", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/memento", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "LastN 历史截断", "metrics": [ "relative_ndcg_change_percent", "results.last_n.hit_at_10", "results.last_n.ndcg_at_10", "results.memento_mmr.hit_at_10", "results.memento_mmr.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "memory-layer", "title": "Memory Layer: Train the In-Model Cache for Recommendation Models", "paper_url": "https://arxiv.org/abs/2607.25110", "detail_path": "reproductions/2607.25110-memory-layer/README.md", "topic": [ "ranking", "training-serving-consistency", "cold-start", "cache", "serving" ], "first_author": null, "first_author_affiliation": "Meta / Instagram Reels", "published": "2026-07-27", "code": null, "adapter": { "adapter_key": "memory-layer", "arxiv_id": "2607.25110", "title": "Memory Layer: Train the In-Model Cache for Recommendation Models", "paper_url": "https://arxiv.org/abs/2607.25110", "track": "recommendation", "organization": "Meta / Instagram Reels", "published": "2026-07-27", "code_url": null, "topics": [ "ranking", "training-serving-consistency", "cold-start", "cache", "serving" ], "local_code_dir": "src/auto_research/reproductions/memory_layer", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (220 users / 360 items)" ], "baseline": "same item tower with an external early frozen cache and no always-on miss path", "metrics": [ "hit_at_10", "ndcg_at_10", "head_share_at_10", "snapshot_coverage", "memory_coverage" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; one chronological writeback pass", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Instagram Reels early-stage ranking", "metric": "cold-start reshare rate", "lift_percent": 5.0, "traffic": "post-launch A/B backtest; paper reports a statistically significant +5% to +6% range", "source_url": "https://arxiv.org/html/2607.25110v1#S6.SS3", "source_location": "Section 6.3, Table 1", "significance": "all Table 1 A/B improvements statistically significant", "retrieved_at": "2026-09-01" }, { "product": "Instagram Reels early-stage ranking", "metric": "video views for media younger than one hour", "lift_percent": 6.0, "traffic": "post-launch A/B backtest; paper reports a +6% to +7% range", "source_url": "https://arxiv.org/html/2607.25110v1#S6.SS3", "source_location": "Section 6.3, Table 1", "significance": "statistically significant", "retrieved_at": "2026-09-01" } ], "selection_exception": null, "evolve_operators": [ "rankmixer_memory_layer" ] } }, { "domain": "recommendation", "key": "mesh", "title": "MESH: Scaling Up Retrieval with Heterogeneous Content Unification", "paper_url": "https://arxiv.org/abs/2607.12392", "detail_path": "reproductions/2607.12392-mesh/README.md", "topic": [ "retrieval", "fresh-content", "serving" ], "first_author": null, "first_author_affiliation": "Pinterest", "published": "2026-07-14", "code": null, "adapter": { "adapter_key": "mesh", "arxiv_id": "2607.12392", "title": "MESH: Scaling Up Retrieval with Heterogeneous Content Unification", "paper_url": "https://arxiv.org/abs/2607.12392", "track": "recommendation", "organization": "Pinterest", "published": "2026-07-14", "code_url": null, "topics": [ "retrieval", "fresh-content", "serving" ], "local_code_dir": "src/auto_research/reproductions/mesh", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "movielens 100k" ], "baseline": "flat retrieval", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Pinterest", "metric": "fresh-item repins", "lift_percent": 5.5, "traffic": "online A/B test", "source_url": "https://arxiv.org/abs/2607.12392", "source_location": "original paper online-result disclosure: product=Pinterest; metric=fresh-item repins" }, { "product": "Pinterest", "metric": "retention", "lift_percent": 0.46, "traffic": "online A/B test", "source_url": "https://arxiv.org/abs/2607.12392", "source_location": "original paper online-result disclosure: product=Pinterest; metric=retention" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "metastrategy", "title": "MetaStrategy: Generative Ranking with Executable LLM Strategies", "paper_url": "https://arxiv.org/abs/2608.09440", "detail_path": "reproductions/2608.09440-metastrategy/README.md", "topic": [ "ranking", "llm-recommendation", "executable-strategy", "online-policy" ], "first_author": null, "first_author_affiliation": "Alibaba / Taobao", "published": "2026-08-10", "code": null, "adapter": { "adapter_key": "metastrategy", "arxiv_id": "2608.09440", "title": "MetaStrategy: Generative Ranking with Executable LLM Strategies", "paper_url": "https://arxiv.org/abs/2608.09440", "track": "recommendation", "organization": "Alibaba / Taobao", "published": "2026-08-10", "code_url": null, "topics": [ "ranking", "llm-recommendation", "executable-strategy", "online-policy" ], "local_code_dir": "src/auto_research/reproductions/metastrategy", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "同预算单目标排序器，实验组为 typed strategy generator + deterministic compiler", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "metastrategy.hit_at_10", "metastrategy.ndcg_at_10", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao Homepage Guess You Like", "metric": "click PV", "lift_percent": 2.11, "traffic": "seven-day user-randomized online A/B", "source_url": "https://arxiv.org/pdf/2608.09440", "source_location": "Section 5.4 / online A/B table", "experiment_duration": "7 days", "retrieved_at": "2026-08-13" }, { "product": "Taobao Homepage Guess You Like", "metric": "transaction amount", "lift_percent": 2.83, "traffic": "seven-day user-randomized online A/B", "source_url": "https://arxiv.org/pdf/2608.09440", "source_location": "Section 5.4 / online A/B table", "experiment_duration": "7 days", "retrieved_at": "2026-08-13" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "mfli", "title": "Rethinking ANN-based Retrieval: Multifaceted Learnable Index for Large-scale Recommendation System", "paper_url": "https://arxiv.org/abs/2602.16124", "detail_path": "reproductions/2602.16124-mfli/README.md", "topic": [ "retrieval", "learnable-index", "fresh-content", "diversity", "serving" ], "first_author": null, "first_author_affiliation": "Meta", "published": "2026-02-18", "code": null, "adapter": { "adapter_key": "mfli", "arxiv_id": "2602.16124", "title": "Rethinking ANN-based Retrieval: Multifaceted Learnable Index for Large-scale Recommendation System", "paper_url": "https://arxiv.org/abs/2602.16124", "track": "recommendation", "organization": "Meta", "published": "2026-02-18", "code_url": null, "topics": [ "retrieval", "learnable-index", "fresh-content", "diversity", "serving" ], "local_code_dir": "src/auto_research/reproductions/mfli", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (220 users / 360 items)" ], "baseline": "同切分的 single-space ANN proxy，实验组加入四 facet MFLI", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta recommendation", "metric": "low-view exposure", "lift_percent": 279.0, "traffic": "7-day online A/B", "source_url": "https://arxiv.org/html/2602.16124v1#S4.SS6", "source_location": "Section 4.6", "experiment_duration": "7 days", "retrieved_at": "2026-08-09" }, { "product": "Meta recommendation", "metric": "diversity", "lift_percent": 0.3, "traffic": "7-day online A/B", "source_url": "https://arxiv.org/html/2602.16124v1#S4.SS6", "source_location": "Section 4.6", "experiment_duration": "7 days", "retrieved_at": "2026-08-09" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "mgoe", "title": "Macro Graph of Experts for Billion-Scale Multi-Task Recommendation", "paper_url": "https://arxiv.org/abs/2506.10520", "detail_path": "reproductions/2506.10520-mgoe/README.md", "topic": [ "multi-task-learning", "mixture-of-experts", "graph-neural-network", "ranking" ], "first_author": null, "first_author_affiliation": "Alibaba", "published": "2025-06-12", "code": "https://github.com/RainmannnnN/MGOE", "adapter": { "adapter_key": "mgoe", "arxiv_id": "2506.10520", "title": "Macro Graph of Experts for Billion-Scale Multi-Task Recommendation", "paper_url": "https://arxiv.org/abs/2506.10520", "track": "recommendation", "organization": "Alibaba", "published": "2025-06-12", "code_url": "https://github.com/RainmannnnN/MGOE", "topics": [ "multi-task-learning", "mixture-of-experts", "graph-neural-network", "ranking" ], "local_code_dir": "src/auto_research/reproductions/mgoe", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "MMoE independent task gates", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.mrr_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.mrr_at_10", "method.ndcg_at_10", "paper_results.cvr_percent", "paper_results.gmv_percent", "paper_results.pctr_percent", "paper_results.uctr_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.mrr_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Alibaba", "metric": "PCTR", "lift_percent": 2.16, "traffic": "production A/B vs MMoE", "source_url": "https://arxiv.org/abs/2506.10520", "source_location": "original paper online-result disclosure: product=Alibaba; metric=PCTR" }, { "product": "Alibaba", "metric": "GMV", "lift_percent": 16.46, "traffic": "production A/B vs MMoE", "source_url": "https://arxiv.org/abs/2506.10520", "source_location": "original paper online-result disclosure: product=Alibaba; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "mim", "title": "MIM: Multi-modal Content Interest Modeling Paradigm for User Behavior Modeling", "paper_url": "https://arxiv.org/abs/2502.00321", "detail_path": "reproductions/2502.00321-mim/README.md", "topic": [ "multimodal-recommendation", "user-modeling", "pretraining", "ranking" ], "first_author": null, "first_author_affiliation": "Alibaba / Taobao", "published": "2025-02-01", "code": "https://pan.quark.cn/s/8fc8ec3e74f3", "adapter": { "adapter_key": "mim", "arxiv_id": "2502.00321", "title": "MIM: Multi-modal Content Interest Modeling Paradigm for User Behavior Modeling", "paper_url": "https://arxiv.org/abs/2502.00321", "track": "recommendation", "organization": "Alibaba / Taobao", "published": "2025-02-01", "code_url": "https://pan.quark.cn/s/8fc8ec3e74f3", "topics": [ "multimodal-recommendation", "user-modeling", "pretraining", "ranking" ], "local_code_dir": "src/auto_research/reproductions/mim", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "ID co-occurrence interest model", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.mrr_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.mrr_at_10", "method.ndcg_at_10", "paper_results.ctr_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.mrr_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao", "metric": "CTR", "lift_percent": 14.14, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2502.00321", "source_location": "original paper online-result disclosure: product=Taobao; metric=CTR" }, { "product": "Taobao", "metric": "RPM", "lift_percent": 4.12, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2502.00321", "source_location": "original paper online-result disclosure: product=Taobao; metric=RPM" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "mima", "title": "MIMA: Multi-Interest Recommendation via Multi-Positive Exclusive Assignment", "paper_url": "https://arxiv.org/abs/2609.12842", "detail_path": "reproductions/2609.12842-mima/README.md", "topic": [ "retrieval", "multi-interest", "exclusive-assignment", "industrial-recommendation" ], "first_author": null, "first_author_affiliation": "Alibaba International Digital Commerce Group", "published": "2026-09-11", "code": null, "adapter": { "adapter_key": "mima", "arxiv_id": "2609.12842", "title": "MIMA: Multi-Interest Recommendation via Multi-Positive Exclusive Assignment", "paper_url": "https://arxiv.org/abs/2609.12842", "track": "recommendation", "organization": "Alibaba International Digital Commerce Group", "published": "2026-09-11", "code_url": null, "topics": [ "retrieval", "multi-interest", "exclusive-assignment", "industrial-recommendation" ], "local_code_dir": "src/auto_research/reproductions/mima", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "large-scale e-commerce homepage retrieval", "metric": "transaction count", "lift_percent": 5.6, "traffic": "seven-day production A/B", "source_url": "https://arxiv.org/html/2609.12842v1", "source_location": "Section 5.5.2, Table 6", "experiment_duration": "seven days", "retrieved_at": "2026-09-15" }, { "product": "large-scale e-commerce homepage retrieval", "metric": "transaction amount", "lift_percent": 5.44, "traffic": "seven-day production A/B", "source_url": "https://arxiv.org/html/2609.12842v1", "source_location": "Section 5.5.2, Table 6", "experiment_duration": "seven days", "retrieved_at": "2026-09-15" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "mixformer", "title": "MixFormer: Co-Scaling Up Dense and Sequence in Industrial Recommenders", "paper_url": "https://arxiv.org/abs/2602.14110", "detail_path": "reproductions/2602.14110-mixformer/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "mixformer", "arxiv_id": "2602.14110", "title": "MixFormer: Co-Scaling Up Dense and Sequence in Industrial Recommenders", "paper_url": "https://arxiv.org/abs/2602.14110", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/mixformer", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens 100K" ], "baseline": "参数量匹配的 Stacked Dense+Sequence", "metrics": [ "ndcg_gain_percent", "results.stacked_dense_plus_sequence.hit_at_10", "results.stacked_dense_plus_sequence.hit_at_10_std", "results.stacked_dense_plus_sequence.ndcg_at_10", "results.stacked_dense_plus_sequence.ndcg_at_10_std", "results.unified_mixformer.hit_at_10", "results.unified_mixformer.hit_at_10_std", "results.unified_mixformer.ndcg_at_10", "results.unified_mixformer.ndcg_at_10_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=240", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "ml-dcn", "title": "ML-DCN: Masked Low-Rank Deep Crossing Network Towards Scalable Ads Click-through Rate Prediction at Pinterest", "paper_url": "https://arxiv.org/abs/2602.09194", "detail_path": "reproductions/2602.09194-ml-dcn/README.md", "topic": [ "advertising", "feature-crossing", "serving-efficiency" ], "first_author": null, "first_author_affiliation": "Pinterest", "published": "2026-02-09", "code": null, "adapter": { "adapter_key": "ml-dcn", "arxiv_id": "2602.09194", "title": "ML-DCN: Masked Low-Rank Deep Crossing Network Towards Scalable Ads Click-through Rate Prediction at Pinterest", "paper_url": "https://arxiv.org/abs/2602.09194", "track": "recommendation", "organization": "Pinterest", "published": "2026-02-09", "code_url": null, "topics": [ "advertising", "feature-crossing", "serving-efficiency" ], "local_code_dir": "src/auto_research/reproductions/ml_dcn", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Pinterest Ads", "metric": "platform-wide CTR", "lift_percent": 1.89, "traffic": "production A/B at neutral cost", "source_url": "https://arxiv.org/html/2602.09194v1", "source_location": "Section 4.2 / Table 6", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "mlcc", "title": "Compress, Cross and Scale: Multi-Level Compression Cross Networks for Efficient Scaling in Recommender Systems", "paper_url": "https://arxiv.org/abs/2602.12041", "detail_path": "reproductions/2602.12041-mlcc/README.md", "topic": [ "feature-interaction", "compression", "scaling-law" ], "first_author": null, "first_author_affiliation": "Bilibili", "published": "2026-02-12", "code": null, "adapter": { "adapter_key": "mlcc", "arxiv_id": "2602.12041", "title": "Compress, Cross and Scale: Multi-Level Compression Cross Networks for Efficient Scaling in Recommender Systems", "paper_url": "https://arxiv.org/abs/2602.12041", "track": "recommendation", "organization": "Bilibili", "published": "2026-02-12", "code_url": null, "topics": [ "feature-interaction", "compression", "scaling-law" ], "local_code_dir": "src/auto_research/reproductions/mlcc", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Bilibili advertising", "metric": "Advertiser value", "lift_percent": 32.0, "traffic": "production online A/B under latency constraint", "source_url": "https://arxiv.org/html/2602.12041v1", "source_location": "Online experiments", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "head:multi-level-cross" ] } }, { "domain": "recommendation", "key": "mm-llm", "title": "A General Framework for Multimodal LLM-Based Multimedia Understanding in Large-Scale Recommendation Systems", "paper_url": "https://arxiv.org/abs/2605.09338", "detail_path": "reproductions/2605.09338-mm-llm/README.md", "topic": [ "multimodal", "llm-recommendation", "ranking" ], "first_author": null, "first_author_affiliation": "Meta", "published": "2026-05", "code": null, "adapter": { "adapter_key": "mm-llm", "arxiv_id": "2605.09338", "title": "A General Framework for Multimodal LLM-Based Multimedia Understanding in Large-Scale Recommendation Systems", "paper_url": "https://arxiv.org/abs/2605.09338", "track": "recommendation", "organization": "Meta", "published": "2026-05", "code_url": null, "topics": [ "multimodal", "llm-recommendation", "ranking" ], "local_code_dir": "src/auto_research/reproductions/mm_llm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K content proxy" ], "baseline": "DIN，MM-LLM NDCG@10 相对 DIN -13.23%", "metrics": [ "paper_online_ab.offline_auc_percent", "results.din.hit_at_10", "results.din.ndcg_at_10", "results.mm_llm_features.hit_at_10", "results.mm_llm_features.ndcg_at_10", "results.ndcg_ablation_percent", "results.ndcg_vs_din_percent", "results.visual_id_baseline.hit_at_10", "results.visual_id_baseline.ndcg_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=100", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta recommendation", "metric": "engagement", "lift_percent": 0.02, "traffic": "large-scale online A/B", "source_url": "https://arxiv.org/abs/2605.09338", "source_location": "original paper online-result disclosure: product=Meta recommendation; metric=engagement" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "mmoe", "title": "Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts", "paper_url": "https://research.google/pubs/modeling-task-relationships-in-multi-task-learning-with-multi-gate-mixture-of-experts/", "detail_path": "reproductions/kdd2018-mmoe-mmoe/README.md", "topic": [ "ranking", "classic", "multi-task", "mixture-of-experts" ], "first_author": null, "first_author_affiliation": "Google", "published": "2018-08-19", "code": null, "adapter": { "adapter_key": "mmoe", "arxiv_id": "kdd2018-mmoe", "title": "Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts", "paper_url": "https://research.google/pubs/modeling-task-relationships-in-multi-task-learning-with-multi-gate-mixture-of-experts/", "track": "recommendation", "organization": "Google", "published": "2018-08-19", "code_url": null, "topics": [ "ranking", "classic", "multi-task", "mixture-of-experts" ], "local_code_dir": "src/auto_research/reproductions/mmoe", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K entire-space multitask construction" ], "baseline": "shared-bottom 双任务网络，实验组为每任务独立 gate 的共享 experts", "metrics": [ "baseline.click_auc", "baseline.click_auc_std", "baseline.click_logloss", "baseline.click_logloss_std", "baseline.conversion_auc", "baseline.conversion_auc_std", "baseline.conversion_logloss", "baseline.conversion_logloss_std", "baseline.mean_auc", "baseline.mean_auc_std", "method.click_auc", "method.click_auc_std", "method.click_logloss", "method.click_logloss_std", "method.conversion_auc", "method.conversion_auc_std", "method.conversion_logloss", "method.conversion_logloss_std", "method.mean_auc", "method.mean_auc_std", "relative.conversion_auc_percent", "relative.mean_auc_percent" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=200", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": "用户明确批准 MMoE 为经典例外；原文生产任务未披露当前门槛要求的量化线上 A/B lift。", "evolve_operators": [] } }, { "domain": "recommendation", "key": "mosaic", "title": "Mosaic: A Fleet of User Embedding Specialists for Recommendation at Meta", "paper_url": "https://arxiv.org/abs/2607.24015", "detail_path": "reproductions/2607.24015-mosaic/README.md", "topic": [ "user-embedding", "multi-task-learning", "sequential-recommendation", "mixture-of-experts" ], "first_author": null, "first_author_affiliation": "Meta", "published": "2026-07-27", "code": null, "adapter": { "adapter_key": "mosaic", "arxiv_id": "2607.24015", "title": "Mosaic: A Fleet of User Embedding Specialists for Recommendation at Meta", "paper_url": "https://arxiv.org/abs/2607.24015", "track": "recommendation", "organization": "Meta", "published": "2026-07-27", "code_url": null, "topics": [ "user-embedding", "multi-task-learning", "sequential-recommendation", "mixture-of-experts" ], "local_code_dir": "src/auto_research/reproductions/mosaic", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "同一 MovieLens-1M 切分、同一 full-catalog 候选集上的单 GRU sequential embedding", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "training.baseline.final_loss", "training.baseline.initial_loss", "training.mosaic.final_loss", "training.mosaic.final_mrm_loss", "training.mosaic.initial_loss" ], "default_seeds": [ 42 ], "budget": "steps=120", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta Surface 1", "metric": "topline engagement", "lift_percent": 0.1, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2607.24015", "source_location": "original paper online-result disclosure: product=Meta Surface 1; metric=topline engagement" }, { "product": "Meta Surface 2", "metric": "topline engagement", "lift_percent": 0.15, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2607.24015", "source_location": "original paper online-result disclosure: product=Meta Surface 2; metric=topline engagement" }, { "product": "Meta Surface 3", "metric": "topline engagement", "lift_percent": 0.28, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2607.24015", "source_location": "original paper online-result disclosure: product=Meta Surface 3; metric=topline engagement" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "mpformer", "title": "MPFormer: Adaptive Framework for Industrial Multi-Task Personalized Sequential Retriever", "paper_url": "https://arxiv.org/abs/2508.20400", "detail_path": "reproductions/2508.20400-mpformer/README.md", "topic": [ "retrieval", "multi-task", "sequence", "serving" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2025-08-28", "code": null, "adapter": { "adapter_key": "mpformer", "arxiv_id": "2508.20400", "title": "MPFormer: Adaptive Framework for Industrial Multi-Task Personalized Sequential Retriever", "paper_url": "https://arxiv.org/abs/2508.20400", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2025-08-28", "code_url": null, "topics": [ "retrieval", "multi-task", "sequence", "serving" ], "local_code_dir": "src/auto_research/reproductions/mpformer", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享 transition + content scorer，实验组只加入 mpformer 核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou", "metric": "Watch Time", "lift_percent": 0.426, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2508.20400", "source_location": "original paper online-result disclosure: product=Kuaishou; metric=Watch Time" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "msd", "title": "Balancing Efficiency and Effectiveness: An LLM-Infused Approach for Optimized CTR Prediction", "paper_url": "https://arxiv.org/abs/2412.06860", "detail_path": "reproductions/2412.06860-msd/README.md", "topic": [ "LLM distillation", "CTR", "serving efficiency" ], "first_author": null, "first_author_affiliation": "Meituan", "published": "2024-12", "code": null, "adapter": { "adapter_key": "msd", "arxiv_id": "2412.06860", "title": "Balancing Efficiency and Effectiveness: An LLM-Infused Approach for Optimized CTR Prediction", "paper_url": "https://arxiv.org/abs/2412.06860", "track": "recommendation", "organization": "Meituan", "published": "2024-12", "code_url": null, "topics": [ "LLM distillation", "CTR", "serving efficiency" ], "local_code_dir": "src/auto_research/reproductions/msd", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens-100K title/genre CTR" ], "baseline": "ID-only CTR ranker", "metrics": [ "aggregate.id_auc_mean", "aggregate.id_auc_std", "aggregate.msd_auc_mean", "aggregate.msd_auc_std", "per_seed.42.distill_loss_final", "per_seed.42.distill_loss_initial", "per_seed.42.id_auc", "per_seed.42.msd_auc", "per_seed.43.distill_loss_final", "per_seed.43.distill_loss_initial", "per_seed.43.id_auc", "per_seed.43.msd_auc", "per_seed.44.distill_loss_final", "per_seed.44.distill_loss_initial", "per_seed.44.id_auc", "per_seed.44.msd_auc" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=100", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meituan sponsored search", "metric": "CTR", "lift_percent": 2.12, "traffic": "2024-10-20 to 2024-10-30", "source_url": "https://arxiv.org/abs/2412.06860", "source_location": "original paper online-result disclosure: product=Meituan sponsored search; metric=CTR" }, { "product": "Meituan sponsored search", "metric": "CPM", "lift_percent": 2.59, "traffic": "2024-10-20 to 2024-10-30", "source_url": "https://arxiv.org/abs/2412.06860", "source_location": "original paper online-result disclosure: product=Meituan sponsored search; metric=CPM" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "msn", "title": "MSN: A Memory-based Sparse Activation Scaling Framework for Large-scale Industrial Recommendation", "paper_url": "https://arxiv.org/abs/2602.07526", "detail_path": "reproductions/2602.07526-msn/README.md", "topic": [ "search-ranking", "sparse-memory", "product-key-memory" ], "first_author": null, "first_author_affiliation": "ByteDance / Douyin Search", "published": "2026-02-07", "code": null, "adapter": { "adapter_key": "msn", "arxiv_id": "2602.07526", "title": "MSN: A Memory-based Sparse Activation Scaling Framework for Large-scale Industrial Recommendation", "paper_url": "https://arxiv.org/abs/2602.07526", "track": "recommendation", "organization": "ByteDance / Douyin Search", "published": "2026-02-07", "code_url": null, "topics": [ "search-ranking", "sparse-memory", "product-key-memory" ], "local_code_dir": "src/auto_research/reproductions/msn", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (260 users / 420 items)" ], "baseline": "dense transition/content ranker，实验组为 gated top-k PKM", "metrics": [ "baseline.fresh_hit_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "memory.active_slots_per_request", "memory.occupied_slots", "memory.slots", "method.fresh_hit_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Douyin Search", "metric": "Watch time", "lift_percent": 0.2958, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2602.07526", "source_location": "original paper online-result disclosure: product=Douyin Search; metric=Watch time" }, { "product": "Douyin Search", "metric": "Finish rate", "lift_percent": 0.2071, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2602.07526", "source_location": "original paper online-result disclosure: product=Douyin Search; metric=Finish rate" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "muchator", "title": "MuChator: Enabling Active Music Discovery via Conversational Music LLMs in Douyin Music", "paper_url": "https://arxiv.org/abs/2605.27103", "detail_path": "reproductions/2605.27103-muchator/README.md", "topic": [ "music-recommendation", "conversational-llm", "active-discovery" ], "first_author": null, "first_author_affiliation": "ByteDance", "published": "2026-05-26", "code": null, "adapter": { "adapter_key": "muchator", "arxiv_id": "2605.27103", "title": "MuChator: Enabling Active Music Discovery via Conversational Music LLMs in Douyin Music", "paper_url": "https://arxiv.org/abs/2605.27103", "track": "recommendation", "organization": "ByteDance", "published": "2026-05-26", "code_url": null, "topics": [ "music-recommendation", "conversational-llm", "active-discovery" ], "local_code_dir": "src/auto_research/reproductions/muchator", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Douyin Music", "metric": "Active days", "lift_percent": 46.49, "traffic": "approximately one month; >0.1B daily users", "source_url": "https://arxiv.org/html/2605.27103v1", "source_location": "Section 4.2, Table 2", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "context:conversational-intent" ] } }, { "domain": "recommendation", "key": "netflix-mediafm", "title": "Multimedia Asset Personalization via Multimodal Embeddings at Netflix", "paper_url": "https://arxiv.org/abs/2608.18322", "detail_path": "reproductions/2608.18322-netflix-mediafm/README.md", "topic": [ "multimodal-recommendation", "content-understanding", "cold-start" ], "first_author": null, "first_author_affiliation": "Netflix", "published": "2026-08-18", "code": null, "adapter": { "adapter_key": "netflix-mediafm", "arxiv_id": "2608.18322", "title": "Multimedia Asset Personalization via Multimodal Embeddings at Netflix", "paper_url": "https://arxiv.org/abs/2608.18322", "track": "recommendation", "organization": "Netflix", "published": "2026-08-18", "code_url": null, "topics": [ "multimodal-recommendation", "content-understanding", "cold-start" ], "local_code_dir": "src/auto_research/reproductions/netflix_mediafm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Netflix Search", "metric": "playthrough rate", "lift_percent": 0.36, "traffic": "Search canvas A/B", "source_url": "https://arxiv.org/html/2608.18322v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "next-vlm", "title": "NEXT: Reasoning-Driven Video Recommendation via a Vision-Language Model", "paper_url": "https://arxiv.org/abs/2607.24789", "detail_path": "reproductions/2607.24789-next-vlm/README.md", "topic": [ "multimodal", "reasoning", "retrieval", "knowledge-graph", "vlm" ], "first_author": null, "first_author_affiliation": "Meta", "published": "2026-06-27", "code": null, "adapter": { "adapter_key": "next-vlm", "arxiv_id": "2607.24789", "title": "NEXT: Reasoning-Driven Video Recommendation via a Vision-Language Model", "paper_url": "https://arxiv.org/abs/2607.24789", "track": "recommendation", "organization": "Meta", "published": "2026-06-27", "code_url": null, "topics": [ "multimodal", "reasoning", "retrieval", "knowledge-graph", "vlm" ], "local_code_dir": "src/auto_research/reproductions/next_vlm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "multi-path transition/semantic ranker", "metrics": [ "Hit@10", "NDCG@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "commercial short-form video platform", "metric": "watch time", "lift_percent": 0.53, "traffic": "multi-week A/B at approximately 100M users", "source_url": "https://arxiv.org/html/2607.24789v1#S5.SS2", "source_location": "Section 5.2 Table 5", "experiment_duration": "multiple weeks", "retrieved_at": "2026-08-24" }, { "product": "commercial short-form video platform", "metric": "distinct video exposure", "lift_percent": 0.51, "traffic": "multi-week A/B at approximately 100M users", "source_url": "https://arxiv.org/html/2607.24789v1#S5.SS2", "source_location": "Section 5.2 Table 5", "experiment_duration": "multiple weeks", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "nontp", "title": "Not Only NTP: Extending Training Signal Coverage for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2607.12277", "detail_path": "reproductions/2607.12277-nontp/README.md", "topic": [ "generative-recommendation", "contrastive-learning", "multi-domain" ], "first_author": null, "first_author_affiliation": "Meituan", "published": "2026-07-14", "code": null, "adapter": { "adapter_key": "nontp", "arxiv_id": "2607.12277", "title": "Not Only NTP: Extending Training Signal Coverage for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2607.12277", "track": "recommendation", "organization": "Meituan", "published": "2026-07-14", "code_url": null, "topics": [ "generative-recommendation", "contrastive-learning", "multi-domain" ], "local_code_dir": "src/auto_research/reproductions/nontp", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K chronological leave-two-out; primary genre domain" ], "baseline": "相同 causal backbone、相同初始化和训练预算的 ntp", "metrics": [ "relative_nontp_vs_ntp_percent.hit_at_10", "relative_nontp_vs_ntp_percent.ndcg_at_10", "results.nontp.hit_at_10", "results.nontp.hit_at_10_std", "results.nontp.ndcg_at_10", "results.nontp.ndcg_at_10_std", "results.ntp.hit_at_10", "results.ntp.hit_at_10_std", "results.ntp.ndcg_at_10", "results.ntp.ndcg_at_10_std", "results.tcl.hit_at_10", "results.tcl.hit_at_10_std", "results.tcl.ndcg_at_10", "results.tcl.ndcg_at_10_std", "results.tdl.hit_at_10", "results.tdl.hit_at_10_std", "results.tdl.ndcg_at_10", "results.tdl.ndcg_at_10_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=80", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meituan DSP recall", "metric": "CTR", "lift_percent": 1.8, "traffic": "5% traffic each, 14 days, p<0.01", "source_url": "https://arxiv.org/abs/2607.12277", "source_location": "original paper online-result disclosure: product=Meituan DSP recall; metric=CTR" }, { "product": "Meituan DSP recall", "metric": "GMV", "lift_percent": 2.1, "traffic": "5% traffic each, 14 days, p<0.01", "source_url": "https://arxiv.org/abs/2607.12277", "source_location": "original paper online-result disclosure: product=Meituan DSP recall; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "notellm", "title": "NoteLLM: A Retrievable Large Language Model for Note Recommendation", "paper_url": "https://arxiv.org/abs/2403.01744", "detail_path": "reproductions/2403.01744-notellm/README.md", "topic": [ "llm-recommendation", "contrastive-learning", "i2i-retrieval" ], "first_author": null, "first_author_affiliation": "Xiaohongshu", "published": "2024-03", "code": null, "adapter": { "adapter_key": "notellm", "arxiv_id": "2403.01744", "title": "NoteLLM: A Retrievable Large Language Model for Note Recommendation", "paper_url": "https://arxiv.org/abs/2403.01744", "track": "recommendation", "organization": "Xiaohongshu", "published": "2024-03", "code_url": null, "topics": [ "llm-recommendation", "contrastive-learning", "i2i-retrieval" ], "local_code_dir": "src/auto_research/reproductions/notellm", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens 100K title/genre and co-occurrence pairs" ], "baseline": "冻结 compression-state embedding", "metrics": [ "hit_at_10.frozen_compression_embedding.mean", "hit_at_10.frozen_compression_embedding.std", "hit_at_10.notellm_gcl_csft.mean", "hit_at_10.notellm_gcl_csft.std", "hit_at_10.relative_change_percent", "ndcg_at_10.frozen_compression_embedding.mean", "ndcg_at_10.frozen_compression_embedding.std", "ndcg_at_10.notellm_gcl_csft.mean", "ndcg_at_10.notellm_gcl_csft.std", "ndcg_at_10.relative_change_percent" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=40", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Xiaohongshu I2I", "metric": "CTR", "lift_percent": 16.2, "traffic": "week-long online experiment", "source_url": "https://arxiv.org/abs/2403.01744", "source_location": "original paper online-result disclosure: product=Xiaohongshu I2I; metric=CTR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "nova", "title": "NOVA: A Verification-Aware Agent Harness for Architecture Evolution in Industrial Recommender Systems", "paper_url": "https://arxiv.org/abs/2606.27243", "detail_path": "reproductions/2606.27243-nova/README.md", "topic": [ "auto-research", "architecture-evolution", "verification" ], "first_author": null, "first_author_affiliation": "Tencent", "published": "2026-06-25", "code": null, "adapter": { "adapter_key": "nova", "arxiv_id": "2606.27243", "title": "NOVA: A Verification-Aware Agent Harness for Architecture Evolution in Industrial Recommender Systems", "paper_url": "https://arxiv.org/abs/2606.27243", "track": "recommendation", "organization": "Tencent", "published": "2026-06-25", "code_url": null, "topics": [ "auto-research", "architecture-evolution", "verification" ], "local_code_dir": "src/auto_research/reproductions/nova", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (260 users / 420 items)" ], "baseline": "固定启发式 ensemble，实验组为经过验证级联选择的 specialist", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Tencent advertising", "metric": "GMV objective 1", "lift_percent": 1.25, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2606.27243", "source_location": "original paper online-result disclosure: product=Tencent advertising; metric=GMV objective 1" }, { "product": "Tencent advertising", "metric": "GMV objective 2", "lift_percent": 1.7, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2606.27243", "source_location": "original paper online-result disclosure: product=Tencent advertising; metric=GMV objective 2" }, { "product": "Tencent advertising", "metric": "GMV objective 3", "lift_percent": 2.02, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2606.27243", "source_location": "original paper online-result disclosure: product=Tencent advertising; metric=GMV objective 3" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "ogr", "title": "Once Generated, Ranked: End-to-End Generative Slate Recommendation with Unified Semantic-Collaborative IDs", "paper_url": "https://arxiv.org/abs/2608.17613", "detail_path": "reproductions/2608.17613-ogr/README.md", "topic": [ "generative-recommendation", "slate-ranking", "semantic-id", "reinforcement-learning" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-08-18", "code": null, "adapter": { "adapter_key": "ogr", "arxiv_id": "2608.17613", "title": "Once Generated, Ranked: End-to-End Generative Slate Recommendation with Unified Semantic-Collaborative IDs", "paper_url": "https://arxiv.org/abs/2608.17613", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-08-18", "code_url": null, "topics": [ "generative-recommendation", "slate-ranking", "semantic-id", "reinforcement-learning" ], "local_code_dir": "src/auto_research/reproductions/ogr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou", "metric": "Effective Views", "lift_percent": 1.12, "traffic": "3% production traffic", "source_url": "https://arxiv.org/html/2608.17613v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "onemall", "title": "OneMall: One Model, More Scenarios -- End-to-End Generative Recommender Family at Kuaishou E-Commerce", "paper_url": "https://arxiv.org/abs/2601.21770", "detail_path": "reproductions/2601.21770-onemall/README.md", "topic": [ "generative-recommendation", "multi-scenario", "semantic-id" ], "first_author": null, "first_author_affiliation": "Kuaishou", "published": "2026-01-29", "code": null, "adapter": { "adapter_key": "onemall", "arxiv_id": "2601.21770", "title": "OneMall: One Model, More Scenarios -- End-to-End Generative Recommender Family at Kuaishou E-Commerce", "paper_url": "https://arxiv.org/abs/2601.21770", "track": "recommendation", "organization": "Kuaishou", "published": "2026-01-29", "code_url": null, "topics": [ "generative-recommendation", "multi-scenario", "semantic-id" ], "local_code_dir": "src/auto_research/reproductions/onemall", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享 transition/content ensemble", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou product-card", "metric": "GMV", "lift_percent": 13.01, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2601.21770", "source_location": "original paper online-result disclosure: product=Kuaishou product-card; metric=GMV" }, { "product": "Kuaishou short-video", "metric": "orders", "lift_percent": 15.32, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2601.21770", "source_location": "original paper online-result disclosure: product=Kuaishou short-video; metric=orders" }, { "product": "Kuaishou live", "metric": "orders", "lift_percent": 2.78, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2601.21770", "source_location": "original paper online-result disclosure: product=Kuaishou live; metric=orders" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "onemodel", "title": "OneModel: A Unified Foundation for Platform-Scale Multi-Scenario Ranking", "paper_url": "https://arxiv.org/abs/2608.18606", "detail_path": "reproductions/2608.18606-onemodel/README.md", "topic": [ "ranking", "multi-scenario", "foundation-model", "long-sequence" ], "first_author": null, "first_author_affiliation": "Xiaohongshu", "published": "2026-08-19", "code": null, "adapter": { "adapter_key": "onemodel", "arxiv_id": "2608.18606", "title": "OneModel: A Unified Foundation for Platform-Scale Multi-Scenario Ranking", "paper_url": "https://arxiv.org/abs/2608.18606", "track": "recommendation", "organization": "Xiaohongshu", "published": "2026-08-19", "code_url": null, "topics": [ "ranking", "multi-scenario", "foundation-model", "long-sequence" ], "local_code_dir": "src/auto_research/reproductions/onemodel", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "shared global sequential ranker", "metrics": [ "Hit@10", "NDCG@10", "head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Xiaohongshu Explore Feed", "metric": "time spent", "lift_percent": 0.33, "traffic": "production A/B", "source_url": "https://arxiv.org/html/2608.18606v1#S5.SS3", "source_location": "Section 5.3 Table 4", "retrieved_at": "2026-08-24" }, { "product": "Xiaohongshu Feed Ads", "metric": "advertising value", "lift_percent": 3.43, "traffic": "production A/B", "source_url": "https://arxiv.org/html/2608.18606v1#S5.SS3", "source_location": "Section 5.3 Table 4", "retrieved_at": "2026-08-24" }, { "product": "Xiaohongshu Merchant Recommendation", "metric": "DGMV", "lift_percent": 1.1867, "traffic": "production A/B", "source_url": "https://arxiv.org/html/2608.18606v1#S5.SS3", "source_location": "Section 5.3 Table 4", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "onepiece", "title": "OnePiece: Bringing Context Engineering and Reasoning to Industrial Cascade Ranking", "paper_url": "https://arxiv.org/abs/2509.18091", "detail_path": "reproductions/2509.18091-onepiece/README.md", "topic": [ "ranking", "retrieval", "latent-reasoning", "multi-task" ], "first_author": null, "first_author_affiliation": "Shopee", "published": "2025-09-22", "code": null, "adapter": { "adapter_key": "onepiece", "arxiv_id": "2509.18091", "title": "OnePiece: Bringing Context Engineering and Reasoning to Industrial Cascade Ranking", "paper_url": "https://arxiv.org/abs/2509.18091", "track": "recommendation", "organization": "Shopee", "published": "2025-09-22", "code_url": null, "topics": [ "ranking", "retrieval", "latent-reasoning", "multi-task" ], "local_code_dir": "src/auto_research/reproductions/onepiece", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享 transition + content scorer，实验组只加入 onepiece 核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "paper_results.ad_revenue_percent", "paper_results.ranking_gmv_uu_percent", "paper_results.retrieval_gmv_uu_percent", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "steps=3", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Shopee", "metric": "Advertising Revenue", "lift_percent": 2.9, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2509.18091", "source_location": "original paper online-result disclosure: product=Shopee; metric=Advertising Revenue" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "onerank", "title": "OneRank: Unified Transformer-Native Ranking Architecture for Multi-Task Recommendation", "paper_url": "https://arxiv.org/abs/2606.16838", "detail_path": "reproductions/2606.16838-onerank/README.md", "topic": [ "ranking", "multi-task", "transformer" ], "first_author": null, "first_author_affiliation": "Renmin University of China", "published": "2026-06-15", "code": null, "adapter": { "adapter_key": "onerank", "arxiv_id": "2606.16838", "title": "OneRank: Unified Transformer-Native Ranking Architecture for Multi-Task Recommendation", "paper_url": "https://arxiv.org/abs/2606.16838", "track": "recommendation", "organization": "Renmin University of China", "published": "2026-06-15", "code_url": null, "topics": [ "ranking", "multi-task", "transformer" ], "local_code_dir": "src/auto_research/reproductions/onerank", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "Transformer encoder plus separate task predictor", "metrics": [ "hit_at_10", "ndcg_at_10", "task_private_channels", "gradient_detach_boundaries" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Shopee personalized ranking", "metric": "GMV per user", "lift_percent": 1.01, "traffic": "seven-day test; 10% treatment and 10% control", "source_url": "https://arxiv.org/html/2606.16838v1", "source_location": "Section 4.5, Table 3", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "head:unified-ranker" ] } }, { "domain": "recommendation", "key": "oneranker", "title": "OneRanker: Unified Generation and Ranking with One Model in Industrial Advertising Recommendation", "paper_url": "https://arxiv.org/abs/2603.02999", "detail_path": "reproductions/2603.02999-oneranker/README.md", "topic": [ "advertising", "generative-recommendation", "ranking", "multi-task" ], "first_author": null, "first_author_affiliation": "Tencent", "published": "2026-03-03", "code": null, "adapter": { "adapter_key": "oneranker", "arxiv_id": "2603.02999", "title": "OneRanker: Unified Generation and Ranking with One Model in Industrial Advertising Recommendation", "paper_url": "https://arxiv.org/abs/2603.02999", "track": "recommendation", "organization": "Tencent", "published": "2026-03-03", "code_url": null, "topics": [ "advertising", "generative-recommendation", "ranking", "multi-task" ], "local_code_dir": "src/auto_research/reproductions/oneranker", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享 transition + content scorer，实验组只加入 oneranker 核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "paper_results.gmv_percent", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.distribution_consistency_loss", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Weixin Channels Ads", "metric": "GMV", "lift_percent": 1.34, "traffic": "full deployment", "source_url": "https://arxiv.org/abs/2603.02999", "source_location": "original paper online-result disclosure: product=Weixin Channels Ads; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "onerec", "title": "OneRec: Unifying Retrieve and Rank with Generative Recommender and Iterative Preference Alignment", "paper_url": "https://arxiv.org/abs/2502.18965", "detail_path": "reproductions/2502.18965-onerec/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "onerec", "arxiv_id": "2502.18965", "title": "OneRec: Unifying Retrieve and Rank with Generative Recommender and Iterative Preference Alignment", "paper_url": "https://arxiv.org/abs/2502.18965", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/onerec", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens 1M" ], "baseline": "Session Generator SFT", "metrics": [ "configuration.reward_steps", "results.onerec_iterative_preference_alignment.hit_at_10", "results.onerec_iterative_preference_alignment.ndcg_at_10", "results.session_generator_sft.hit_at_10", "results.session_generator_sft.ndcg_at_10", "training.dpo_final_loss", "training.dpo_initial_loss", "training.reward_final_loss", "training.sft_final_loss" ], "default_seeds": [ 42 ], "budget": "steps=240", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "onerec-v2", "title": "OneRec-V2 Technical Report", "paper_url": "https://arxiv.org/abs/2508.20900", "detail_path": "reproductions/2508.20900-onerec-v2/README.md", "topic": [ "llm-recommendation", "generative-retrieval", "reinforcement-learning" ], "first_author": null, "first_author_affiliation": "Kuaishou", "published": "2025-08", "code": null, "adapter": { "adapter_key": "onerec-v2", "arxiv_id": "2508.20900", "title": "OneRec-V2 Technical Report", "paper_url": "https://arxiv.org/abs/2508.20900", "track": "recommendation", "organization": "Kuaishou", "published": "2025-08", "code_url": null, "topics": [ "llm-recommendation", "generative-retrieval", "reinforcement-learning" ], "local_code_dir": "src/auto_research/reproductions/onerec_v2", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "KuaiRand-Pure official standard feed logs" ], "baseline": "Encoder-Decoder，实验组 Lazy SFT 的延迟 -54.78%、validation loss -0.34%", "metrics": [ "summary.encoder_loss_mean", "summary.lazy_latency_reduction_percent", "summary.lazy_loss_mean" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=80", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou", "metric": "app stay time", "lift_percent": 0.467, "traffic": "5% traffic, one week", "source_url": "https://arxiv.org/abs/2508.20900", "source_location": "original paper online-result disclosure: product=Kuaishou; metric=app stay time" }, { "product": "Kuaishou Lite", "metric": "app stay time", "lift_percent": 0.741, "traffic": "5% traffic, one week", "source_url": "https://arxiv.org/abs/2508.20900", "source_location": "original paper online-result disclosure: product=Kuaishou Lite; metric=app stay time" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "oneshot-index", "title": "OneShot: Index-in-Ranking with Neural Scoring for Large-Scale Retrieval", "paper_url": "https://arxiv.org/abs/2607.27475", "detail_path": "reproductions/2607.27475-oneshot-index/README.md", "topic": [ "retrieval", "learnable-index", "neural-scoring", "serving" ], "first_author": null, "first_author_affiliation": "Meta / Instagram", "published": "2026-07-29", "code": null, "adapter": { "adapter_key": "oneshot-index", "arxiv_id": "2607.27475", "title": "OneShot: Index-in-Ranking with Neural Scoring for Large-Scale Retrieval", "paper_url": "https://arxiv.org/abs/2607.27475", "track": "recommendation", "organization": "Meta / Instagram", "published": "2026-07-29", "code_url": null, "topics": [ "retrieval", "learnable-index", "neural-scoring", "serving" ], "local_code_dir": "src/auto_research/reproductions/oneshot_index", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "two-tower dot-product retrieval", "metrics": [ "Hit@10", "NDCG@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Instagram short video", "metric": "daily sessions", "lift_percent": 0.035, "traffic": "fully deployed to global traffic", "source_url": "https://arxiv.org/html/2607.27475v2#S5.SS1.SSS2", "source_location": "Section 5.1.2 Table 3", "retrieved_at": "2026-08-24" }, { "product": "Instagram short video", "metric": "watch time", "lift_percent": 0.136, "traffic": "fully deployed to global traffic", "source_url": "https://arxiv.org/html/2607.27475v2#S5.SS1.SSS2", "source_location": "Section 5.1.2 Table 3", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "onetrans", "title": "OneTrans: Unified Feature Interaction and Sequence Modeling with One Transformer in Industrial Recommender", "paper_url": "https://arxiv.org/abs/2510.26104", "detail_path": "reproductions/2510.26104-onetrans/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "onetrans", "arxiv_id": "2510.26104", "title": "OneTrans: Unified Feature Interaction and Sequence Modeling with One Transformer in Industrial Recommender", "paper_url": "https://arxiv.org/abs/2510.26104", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/onetrans", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens 100K" ], "baseline": "独立 sequence encoder 后与 candidate 交互的 Encode-then-Interact", "metrics": [ "ndcg_gain_percent", "results.encode_then_interact.hit_at_10", "results.encode_then_interact.hit_at_10_std", "results.encode_then_interact.ndcg_at_10", "results.encode_then_interact.ndcg_at_10_std", "results.onetrans.hit_at_10", "results.onetrans.hit_at_10_std", "results.onetrans.ndcg_at_10", "results.onetrans.ndcg_at_10_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=240", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "open-web-ufm", "title": "Building a User Foundation Model for the Open Web", "paper_url": "https://arxiv.org/abs/2607.28019", "detail_path": "reproductions/2607.28019-open-web-ufm/README.md", "topic": [ "advertising", "user-foundation-model", "self-supervised-learning", "llm-optimizer" ], "first_author": null, "first_author_affiliation": "Teads", "published": "2026-07-30", "code": null, "adapter": { "adapter_key": "open-web-ufm", "arxiv_id": "2607.28019", "title": "Building a User Foundation Model for the Open Web", "paper_url": "https://arxiv.org/abs/2607.28019", "track": "recommendation", "organization": "Teads", "published": "2026-07-30", "code_url": null, "topics": [ "advertising", "user-foundation-model", "self-supervised-learning", "llm-optimizer" ], "local_code_dir": "src/auto_research/reproductions/open_web_ufm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M (240 users / 400 items)" ], "baseline": "matched control without the open-web-ufm mechanism", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "method.pretraining_updates", "relative_ndcg_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Teads click-optimized RTB", "metric": "CTR", "lift_percent": 2.13, "traffic": "full traffic, 50/50 user-level split, 7 days", "source_url": "https://arxiv.org/abs/2607.28019", "source_location": "original paper online-result disclosure: product=Teads click-optimized RTB; metric=CTR" }, { "product": "Teads visit-optimized RTB", "metric": "visit rate", "lift_percent": 2.37, "traffic": "full traffic, 50/50 user-level split, 7 days", "source_url": "https://arxiv.org/abs/2607.28019", "source_location": "original paper online-result disclosure: product=Teads visit-optimized RTB; metric=visit rate" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "oxygenrec-v2", "title": "OxygenREC-v2: Internalizing Discrimination into Generative Recommendation", "paper_url": "https://arxiv.org/abs/2607.24255", "detail_path": "reproductions/2607.24255-oxygenrec-v2/README.md", "topic": [ "generative-recommendation", "privileged-distillation", "behavior-instruction" ], "first_author": null, "first_author_affiliation": "JD.COM", "published": "2026-07-27", "code": null, "adapter": { "adapter_key": "oxygenrec-v2", "arxiv_id": "2607.24255", "title": "OxygenREC-v2: Internalizing Discrimination into Generative Recommendation", "paper_url": "https://arxiv.org/abs/2607.24255", "track": "recommendation", "organization": "JD.COM", "published": "2026-07-27", "code_url": null, "topics": [ "generative-recommendation", "privileged-distillation", "behavior-instruction" ], "local_code_dir": "src/auto_research/reproductions/oxygenrec_v2", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "matched control without the oxygenrec-v2 mechanism", "metrics": [ "paper_results.offline_recall_at_512", "paper_results.online_gmv_best_lift_percent", "relative.entropy_routed_positions_percent", "relative.final_loss_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.initial_loss_percent", "relative.ndcg_at_10_percent", "relative.parameters_percent", "variants.OxygenREC-v2 core.final_loss", "variants.OxygenREC-v2 core.hit_at_10", "variants.OxygenREC-v2 core.initial_loss", "variants.OxygenREC-v2 core.ndcg_at_10", "variants.PT-only generator.final_loss", "variants.PT-only generator.hit_at_10", "variants.PT-only generator.initial_loss", "variants.PT-only generator.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "steps=60", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "JD.COM six production surfaces", "metric": "UCTCVR", "lift_percent": 4.44, "traffic": "5–20% traffic, 5–8 days", "source_url": "https://arxiv.org/abs/2607.24255", "source_location": "original paper online-result disclosure: product=JD.COM six production surfaces; metric=UCTCVR" }, { "product": "JD.COM homepage floor", "metric": "GMV", "lift_percent": 21.21, "traffic": "bucketed A/B, p<0.05", "source_url": "https://arxiv.org/abs/2607.24255", "source_location": "original paper online-result disclosure: product=JD.COM homepage floor; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "pa-bridge", "title": "Bridging Passive and Active: Enhancing Conversation Starter Recommendation via Active Expression Modeling", "paper_url": "https://arxiv.org/abs/2605.05855", "detail_path": "reproductions/2605.05855-pa-bridge/README.md", "topic": [ "conversation-starter", "distribution-alignment", "debiasing" ], "first_author": null, "first_author_affiliation": "ByteDance", "published": "2026-05-07", "code": null, "adapter": { "adapter_key": "pa-bridge", "arxiv_id": "2605.05855", "title": "Bridging Passive and Active: Enhancing Conversation Starter Recommendation via Active Expression Modeling", "paper_url": "https://arxiv.org/abs/2605.05855", "track": "recommendation", "organization": "ByteDance", "published": "2026-05-07", "code_url": null, "topics": [ "conversation-starter", "distribution-alignment", "debiasing" ], "local_code_dir": "src/auto_research/reproductions/pa_bridge", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "ByteDance conversation starter", "metric": "Feature penetration", "lift_percent": 0.54, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2605.05855v1", "source_location": "Section 3.3, Table 2", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "context:active-expression" ] } }, { "domain": "recommendation", "key": "pearl-percentile", "title": "PEARL: Unbiased Percentile Estimation via Contrastive Learning for Industrial-Scale Livestream Recommendation", "paper_url": "https://arxiv.org/abs/2605.21752", "detail_path": "reproductions/2605.21752-pearl-percentile/README.md", "topic": [ "ranking", "contrastive-learning", "debiasing" ], "first_author": null, "first_author_affiliation": "TikTok", "published": "2026-05-20", "code": null, "adapter": { "adapter_key": "pearl-percentile", "arxiv_id": "2605.21752", "title": "PEARL: Unbiased Percentile Estimation via Contrastive Learning for Industrial-Scale Livestream Recommendation", "paper_url": "https://arxiv.org/abs/2605.21752", "track": "recommendation", "organization": "TikTok", "published": "2026-05-20", "code_url": null, "topics": [ "ranking", "contrastive-learning", "debiasing" ], "local_code_dir": "src/auto_research/reproductions/pearl_percentile", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "livestream platform", "metric": "Watch Duration", "lift_percent": 2.1, "traffic": "week-long online A/B", "source_url": "https://arxiv.org/html/2605.21752v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "piano", "title": "PIANO: Personalized Reranking via Information Aggregation Node for Music Search Optimization", "paper_url": "https://arxiv.org/abs/2606.16641", "detail_path": "reproductions/2606.16641-piano/README.md", "topic": [ "search", "reranking", "listwise" ], "first_author": null, "first_author_affiliation": "NetEase Cloud Music", "published": "2026-06-15", "code": null, "adapter": { "adapter_key": "piano", "arxiv_id": "2606.16641", "title": "PIANO: Personalized Reranking via Information Aggregation Node for Music Search Optimization", "paper_url": "https://arxiv.org/abs/2606.16641", "track": "recommendation", "organization": "NetEase Cloud Music", "published": "2026-06-15", "code_url": null, "topics": [ "search", "reranking", "listwise" ], "local_code_dir": "src/auto_research/reproductions/piano", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "pointwise personalized reranker", "metrics": [ "hit_at_10", "ndcg_at_10", "interest_refiner", "information_nodes" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "NetEase Cloud Music search", "metric": "CVR", "lift_percent": 4.45, "traffic": "multi-week stable-bucket experiment", "source_url": "https://arxiv.org/html/2606.16641v1", "source_location": "Section 5.1, Table 6", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "head:listwise-node" ] } }, { "domain": "recommendation", "key": "pin-scale", "title": "Pin-SCALE: Semantic Cascading and Alignment Learning for Engagement-Aware IDs in Cold-Start Recommendations", "paper_url": "https://sigir2026.org/SIGIR2026_program.pdf", "detail_path": "reproductions/sigir2026-pin-scale-pin-scale/README.md", "topic": [ "semantic-id", "engagement-aware", "generative-retrieval" ], "first_author": null, "first_author_affiliation": "Pinterest", "published": "2026-07-19", "code": null, "adapter": { "adapter_key": "pin-scale", "arxiv_id": "sigir2026-pin-scale", "title": "Pin-SCALE: Semantic Cascading and Alignment Learning for Engagement-Aware IDs in Cold-Start Recommendations", "paper_url": "https://sigir2026.org/SIGIR2026_program.pdf", "track": "recommendation", "organization": "Pinterest", "published": "2026-07-19", "code_url": null, "topics": [ "semantic-id", "engagement-aware", "generative-retrieval" ], "local_code_dir": "src/auto_research/reproductions/pin_scale", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "unweighted semantic/content ranker", "metrics": [ "baseline.fresh_hit_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Pinterest", "metric": "Repin", "lift_percent": 3.67, "traffic": "online A/B", "source_url": "https://sigir2026.org/SIGIR2026_program.pdf", "source_location": "original paper online-result disclosure: product=Pinterest; metric=Repin" }, { "product": "Pinterest", "metric": "DAU", "lift_percent": 0.05, "traffic": "online A/B", "source_url": "https://sigir2026.org/SIGIR2026_program.pdf", "source_location": "original paper online-result disclosure: product=Pinterest; metric=DAU" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "pinclip", "title": "PinCLIP: Large-scale Foundational Multimodal Representation at Pinterest", "paper_url": "https://arxiv.org/abs/2603.03544", "detail_path": "reproductions/2603.03544-pinclip/README.md", "topic": [ "multimodal", "contrastive-learning", "cold-start" ], "first_author": null, "first_author_affiliation": "Pinterest", "published": "2026-03-03", "code": null, "adapter": { "adapter_key": "pinclip", "arxiv_id": "2603.03544", "title": "PinCLIP: Large-scale Foundational Multimodal Representation at Pinterest", "paper_url": "https://arxiv.org/abs/2603.03544", "track": "recommendation", "organization": "Pinterest", "published": "2026-03-03", "code_url": null, "topics": [ "multimodal", "contrastive-learning", "cold-start" ], "local_code_dir": "src/auto_research/reproductions/pinclip", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "raw genre/content similarity", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Pinterest organic fresh content", "metric": "Repin", "lift_percent": 15.0, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2603.03544", "source_location": "original paper online-result disclosure: product=Pinterest organic fresh content; metric=Repin" }, { "product": "Pinterest new Ads", "metric": "clicks", "lift_percent": 8.7, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2603.03544", "source_location": "original paper online-result disclosure: product=Pinterest new Ads; metric=clicks" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "pindco", "title": "PinDCO: Whole-Page Aware Dynamic Creative Optimization at Scale", "paper_url": "https://arxiv.org/abs/2609.11943", "detail_path": "reproductions/2609.11943-pindco/README.md", "topic": [ "advertising", "creative-optimization", "whole-page-optimization", "multi-tower" ], "first_author": null, "first_author_affiliation": "Pinterest", "published": "2026-07-21", "code": null, "adapter": { "adapter_key": "pindco", "arxiv_id": "2609.11943", "title": "PinDCO: Whole-Page Aware Dynamic Creative Optimization at Scale", "paper_url": "https://arxiv.org/abs/2609.11943", "track": "recommendation", "organization": "Pinterest", "published": "2026-07-21", "code_url": null, "topics": [ "advertising", "creative-optimization", "whole-page-optimization", "multi-tower" ], "local_code_dir": "src/auto_research/reproductions/pindco", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Pinterest Ads", "metric": "ad CTR", "lift_percent": 3.09, "traffic": "production A/B followed by platform launch", "source_url": "https://arxiv.org/html/2609.11943v1", "source_location": "Abstract; Section 4.2", "retrieved_at": "2026-09-15" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "pinequalizer", "title": "PinEqualizer: Full Funnel Content Exploration and Debiasing System at Pinterest", "paper_url": "https://arxiv.org/abs/2607.22518", "detail_path": "reproductions/2607.22518-pinequalizer/README.md", "topic": [ "cold-start", "exploration", "debiasing", "ranking", "retrieval" ], "first_author": null, "first_author_affiliation": "Pinterest", "published": "2026-07-23", "code": null, "adapter": { "adapter_key": "pinequalizer", "arxiv_id": "2607.22518", "title": "PinEqualizer: Full Funnel Content Exploration and Debiasing System at Pinterest", "paper_url": "https://arxiv.org/abs/2607.22518", "track": "recommendation", "organization": "Pinterest", "published": "2026-07-23", "code_url": null, "topics": [ "cold-start", "exploration", "debiasing", "ranking", "retrieval" ], "local_code_dir": "src/auto_research/reproductions/pinequalizer", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "matched control without the pinequalizer mechanism", "metrics": [ "baseline.fresh_ndcg_at_10", "baseline.ndcg_at_10", "baseline.underexplored_exposure_at_10", "pinequalizer.fresh_ndcg_at_10", "pinequalizer.ndcg_at_10", "relative_percent.fresh_ndcg_at_10", "relative_percent.ndcg_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=100", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Related Pins", "metric": "fresh-content engagement volume, ranking architecture", "lift_percent": 8.63, "traffic": "component-level user A/B", "source_url": "https://arxiv.org/abs/2607.22518", "source_location": "original paper online-result disclosure: product=Related Pins; metric=fresh-content engagement volume, ranking architecture" }, { "product": "Related Pins", "metric": "underexplored-content engagement volume, ranking architecture", "lift_percent": 6.57, "traffic": "component-level user A/B", "source_url": "https://arxiv.org/abs/2607.22518", "source_location": "original paper online-result disclosure: product=Related Pins; metric=underexplored-content engagement volume, ranking architecture" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "pinfm", "title": "PinFM: Foundation Model for User Activity Sequences at a Billion-scale Visual Discovery Platform", "paper_url": "https://arxiv.org/abs/2507.12704", "detail_path": "reproductions/2507.12704-pinfm/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "pinfm", "arxiv_id": "2507.12704", "title": "PinFM: Foundation Model for User Activity Sequences at a Billion-scale Visual Discovery Platform", "paper_url": "https://arxiv.org/abs/2507.12704", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/pinfm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "scratch DCAT", "metrics": [ "test.pinfm.hit_at_10", "test.pinfm.hit_at_10_std", "test.pinfm.ndcg_at_10", "test.pinfm.ndcg_at_10_std", "test.scratch_dcat.hit_at_10", "test.scratch_dcat.hit_at_10_std", "test.scratch_dcat.ndcg_at_10", "test.scratch_dcat.ndcg_at_10_std", "test_ndcg_gain_percent", "validation.ndcg_gain_percent", "validation.pinfm.ndcg_at_10", "validation.pinfm.ndcg_at_10_std", "validation.scratch_dcat.ndcg_at_10", "validation.scratch_dcat.ndcg_at_10_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=160", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "pinrec", "title": "PinRec: Outcome-Conditioned, Multi-Token Generative Retrieval for Industry-Scale Recommendation Systems", "paper_url": "https://arxiv.org/abs/2504.10507", "detail_path": "reproductions/2504.10507-pinrec/README.md", "topic": [ "generative-retrieval", "outcome-conditioning", "multi-token" ], "first_author": null, "first_author_affiliation": "Pinterest", "published": "2025-04", "code": null, "adapter": { "adapter_key": "pinrec", "arxiv_id": "2504.10507", "title": "PinRec: Outcome-Conditioned, Multi-Token Generative Retrieval for Industry-Scale Recommendation Systems", "paper_url": "https://arxiv.org/abs/2504.10507", "track": "recommendation", "organization": "Pinterest", "published": "2025-04", "code_url": null, "topics": [ "generative-retrieval", "outcome-conditioning", "multi-token" ], "local_code_dir": "src/auto_research/reproductions/pinrec", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens 100K rating outcomes" ], "baseline": "unconditioned next-token 模型", "metrics": [ "ndcg_at_10.outcome_conditioned_multi_token.mean", "ndcg_at_10.outcome_conditioned_multi_token.std", "ndcg_at_10.relative_change_percent", "ndcg_at_10.unconditioned.mean", "ndcg_at_10.unconditioned.std", "unordered_recall_at_10.outcome_conditioned_multi_token.mean", "unordered_recall_at_10.outcome_conditioned_multi_token.std", "unordered_recall_at_10.relative_change_percent", "unordered_recall_at_10.unconditioned.mean", "unordered_recall_at_10.unconditioned.std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=200", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Pinterest Homefeed", "metric": "grid clicks", "lift_percent": 4.01, "traffic": "global traffic, CUPED adjusted", "source_url": "https://arxiv.org/abs/2504.10507", "source_location": "original paper online-result disclosure: product=Pinterest Homefeed; metric=grid clicks" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "pinterest-ads-llm", "title": "Fine-Tuned LLM as a Complementary Predictor Improving Ads System", "paper_url": "https://arxiv.org/abs/2605.27856", "detail_path": "reproductions/2605.27856-pinterest-ads-llm/README.md", "topic": [ "advertising", "llm-sft", "grpo", "candidate-generation" ], "first_author": null, "first_author_affiliation": "Pinterest", "published": "2026-05", "code": null, "adapter": { "adapter_key": "pinterest-ads-llm", "arxiv_id": "2605.27856", "title": "Fine-Tuned LLM as a Complementary Predictor Improving Ads System", "paper_url": "https://arxiv.org/abs/2605.27856", "track": "recommendation", "organization": "Pinterest", "published": "2026-05", "code_url": null, "topics": [ "advertising", "llm-sft", "grpo", "candidate-generation" ], "local_code_dir": "src/auto_research/reproductions/pinterest_ads_llm", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MiniOneRec Office" ], "baseline": "同一 SmolLM2-135M、同一公开 advertiser 集合下的 SFT", "metrics": [ "downstream.ranking_auc_baseline", "downstream.ranking_auc_relative_percent", "downstream.ranking_auc_with_llm_feature", "downstream.retrieval_recall_at_50_baseline", "downstream.retrieval_recall_at_50_complementary", "test.sft.recall_at_1", "test.sft.recall_at_20", "test.sft.recall_at_5", "test.sft_grpo.recall_at_1", "test.sft_grpo.recall_at_20", "test.sft_grpo.recall_at_5" ], "default_seeds": [ 42 ], "budget": "steps=160", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "US Shopping Ads", "metric": "RoAS", "lift_percent": 4.94, "traffic": "p=0.021", "source_url": "https://arxiv.org/abs/2605.27856", "source_location": "original paper online-result disclosure: product=US Shopping Ads; metric=RoAS" }, { "product": "Opt-in US Shopping Ads", "metric": "RoAS", "lift_percent": 6.69, "traffic": "p=0.029", "source_url": "https://arxiv.org/abs/2605.27856", "source_location": "original paper online-result disclosure: product=Opt-in US Shopping Ads; metric=RoAS" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "pit", "title": "PIT: A Dynamic Personalized Item Tokenizer for End-to-End Generative Recommendation", "paper_url": "https://arxiv.org/abs/2602.08530", "detail_path": "reproductions/2602.08530-pit/README.md", "topic": [ "generative-recommendation", "semantic-id", "dynamic-tokenizer" ], "first_author": null, "first_author_affiliation": "Beijing University of Posts and Telecommunications", "published": "2026-02-09", "code": null, "adapter": { "adapter_key": "pit", "arxiv_id": "2602.08530", "title": "PIT: A Dynamic Personalized Item Tokenizer for End-to-End Generative Recommendation", "paper_url": "https://arxiv.org/abs/2602.08530", "track": "recommendation", "organization": "Beijing University of Posts and Telecommunications", "published": "2026-02-09", "code_url": null, "topics": [ "generative-recommendation", "semantic-id", "dynamic-tokenizer" ], "local_code_dir": "src/auto_research/reproductions/pit", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou generative recommendation", "metric": "Online business metric", "lift_percent": 0.402, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2602.08530v1", "source_location": "Online experiments", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "head:personalized-tokenizer" ] } }, { "domain": "recommendation", "key": "ple", "title": "Progressive Layered Extraction (PLE): A Novel Multi-Task Learning Model for Personalized Recommendations", "paper_url": "https://doi.org/10.1145/3383313.3412236", "detail_path": "reproductions/recsys2020-ple-ple/README.md", "topic": [ "ranking", "classic", "multi-task", "mixture-of-experts" ], "first_author": null, "first_author_affiliation": "Tencent", "published": "2020-09-22", "code": null, "adapter": { "adapter_key": "ple", "arxiv_id": "recsys2020-ple", "title": "Progressive Layered Extraction (PLE): A Novel Multi-Task Learning Model for Personalized Recommendations", "paper_url": "https://doi.org/10.1145/3383313.3412236", "track": "recommendation", "organization": "Tencent", "published": "2020-09-22", "code_url": null, "topics": [ "ranking", "classic", "multi-task", "mixture-of-experts" ], "local_code_dir": "src/auto_research/reproductions/ple", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K entire-space multitask construction" ], "baseline": "MMoE，实验组为共享与任务专属 experts 的 CGC/PLE", "metrics": [ "baseline.click_auc", "baseline.click_auc_std", "baseline.click_logloss", "baseline.click_logloss_std", "baseline.conversion_auc", "baseline.conversion_auc_std", "baseline.conversion_logloss", "baseline.conversion_logloss_std", "baseline.mean_auc", "baseline.mean_auc_std", "method.click_auc", "method.click_auc_std", "method.click_logloss", "method.click_logloss_std", "method.conversion_auc", "method.conversion_auc_std", "method.conversion_logloss", "method.conversion_logloss_std", "method.mean_auc", "method.mean_auc_std", "relative.conversion_auc_percent", "relative.mean_auc_percent" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=200", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": "用户明确批准 PLE 为经典例外；原文腾讯视频部署证据不按当前新论文的量化线上门槛计入。", "evolve_operators": [] } }, { "domain": "recommendation", "key": "plum", "title": "PLUM: Adapting Pre-trained Language Models for Industrial-scale Generative Recommendations", "paper_url": "https://arxiv.org/abs/2510.07784", "detail_path": "reproductions/2510.07784-plum/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "plum", "arxiv_id": "2510.07784", "title": "PLUM: Adapting Pre-trained Language Models for Industrial-scale Generative Recommendations", "paper_url": "https://arxiv.org/abs/2510.07784", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/plum", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens 1M" ], "baseline": "随机初始化无 CPT（R1），实验组是随机初始化有 CPT（CR1）", "metrics": [ "variants.CR1.cpt_final_loss", "variants.CR1.ndcg_at_10", "variants.CR1.recall_at_10", "variants.CR1.sft_final_loss", "variants.CR2.cpt_final_loss", "variants.CR2.ndcg_at_10", "variants.CR2.recall_at_10", "variants.CR2.sft_final_loss", "variants.R1.ndcg_at_10", "variants.R1.recall_at_10", "variants.R1.sft_final_loss", "variants.R2.ndcg_at_10", "variants.R2.recall_at_10", "variants.R2.sft_final_loss" ], "default_seeds": [ 42 ], "budget": "steps=240", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "podcast-mtl", "title": "Cold-Starting Podcast Ads and Promotions with Multi-Task Learning on Spotify", "paper_url": "https://arxiv.org/abs/2601.02306", "detail_path": "reproductions/2601.02306-podcast-mtl/README.md", "topic": [ "cold-start", "multi-task", "ads" ], "first_author": null, "first_author_affiliation": "Spotify", "published": "2026-01-05", "code": null, "adapter": { "adapter_key": "podcast-mtl", "arxiv_id": "2601.02306", "title": "Cold-Starting Podcast Ads and Promotions with Multi-Task Learning on Spotify", "paper_url": "https://arxiv.org/abs/2601.02306", "track": "recommendation", "organization": "Spotify", "published": "2026-01-05", "code_url": null, "topics": [ "cold-start", "multi-task", "ads" ], "local_code_dir": "src/auto_research/reproductions/podcast_mtl", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "single-task popularity/content ranker", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Spotify podcast ads", "metric": "effective cost per stream", "lift_percent": -22.0, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2601.02306", "source_location": "original paper online-result disclosure: product=Spotify podcast ads; metric=effective cost per stream" }, { "product": "Spotify podcast promotions", "metric": "stream rate", "lift_percent": 24.0, "traffic": "online A/B, upper reported lift", "source_url": "https://arxiv.org/abs/2601.02306", "source_location": "original paper online-result disclosure: product=Spotify podcast promotions; metric=stream rate" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "poem", "title": "POEM: Partial-Order Enhanced Real-Time Sequential Modeling for Recommendation", "paper_url": "https://arxiv.org/abs/2606.29946", "detail_path": "reproductions/2606.29946-poem/README.md", "topic": [ "sequential-recommendation", "partial-order", "real-time" ], "first_author": null, "first_author_affiliation": "Kuaishou", "published": "2026-06-29", "code": null, "adapter": { "adapter_key": "poem", "arxiv_id": "2606.29946", "title": "POEM: Partial-Order Enhanced Real-Time Sequential Modeling for Recommendation", "paper_url": "https://arxiv.org/abs/2606.29946", "track": "recommendation", "organization": "Kuaishou", "published": "2026-06-29", "code_url": null, "topics": [ "sequential-recommendation", "partial-order", "real-time" ], "local_code_dir": "src/auto_research/reproductions/poem", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "chronological sequential ranking", "metrics": [ "hit_at_10", "ndcg_at_10", "ranking_signals", "partial_order_pairs" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou main page", "metric": "Usage Time per User", "lift_percent": 0.249, "traffic": "seven-day test; 5% treatment and 5% control", "source_url": "https://arxiv.org/html/2606.29946v1", "source_location": "Section 4.2.3, Table 5", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "context:partial-order" ] } }, { "domain": "recommendation", "key": "policy-facet", "title": "Policy-Grounded Dynamic Facet Suggestions for Job Search", "paper_url": "https://arxiv.org/abs/2605.16479", "detail_path": "reproductions/2605.16479-policy-facet/README.md", "topic": [ "job-search", "facet-retrieval", "policy-ranking" ], "first_author": null, "first_author_affiliation": "LinkedIn", "published": "2026-05-15", "code": null, "adapter": { "adapter_key": "policy-facet", "arxiv_id": "2605.16479", "title": "Policy-Grounded Dynamic Facet Suggestions for Job Search", "paper_url": "https://arxiv.org/abs/2605.16479", "track": "recommendation", "organization": "LinkedIn", "published": "2026-05-15", "code_url": null, "topics": [ "job-search", "facet-retrieval", "policy-ranking" ], "local_code_dir": "src/auto_research/reproductions/policy_facet", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "LinkedIn Job Search", "metric": "Facet CTR", "lift_percent": 34.8, "traffic": "production A/B; p<0.0001", "source_url": "https://arxiv.org/html/2605.16479v1", "source_location": "Section 4.2, Table 5", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "head:policy-facet" ] } }, { "domain": "recommendation", "key": "precise", "title": "PRECISE: Pre-training Sequential Recommenders with Collaborative and Semantic Information", "paper_url": "https://arxiv.org/abs/2412.06308", "detail_path": "reproductions/2412.06308-precise/README.md", "topic": [ "LLM semantic embedding", "sequential pre-training", "MoE", "targeted training" ], "first_author": null, "first_author_affiliation": "Tencent / WeChat", "published": "2024-12", "code": null, "adapter": { "adapter_key": "precise", "arxiv_id": "2412.06308", "title": "PRECISE: Pre-training Sequential Recommenders with Collaborative and Semantic Information", "paper_url": "https://arxiv.org/abs/2412.06308", "track": "recommendation", "organization": "Tencent / WeChat", "published": "2024-12", "code_url": null, "topics": [ "LLM semantic embedding", "sequential pre-training", "MoE", "targeted training" ], "local_code_dir": "src/auto_research/reproductions/precise", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens-1M" ], "baseline": "只做目标任务训练的 PRECISE-TT", "metrics": [ "aggregate.relative_lift.cold_recall_at_10_percent", "aggregate.relative_lift.ndcg_at_10_percent", "aggregate.relative_lift.recall_at_10_percent", "aggregate.tt_only.cold_recall_at_10_mean", "aggregate.tt_only.cold_recall_at_10_std", "aggregate.tt_only.ndcg_at_10_mean", "aggregate.tt_only.ndcg_at_10_std", "aggregate.tt_only.recall_at_10_mean", "aggregate.tt_only.recall_at_10_std", "aggregate.ut_plus_tt.cold_recall_at_10_mean", "aggregate.ut_plus_tt.cold_recall_at_10_std", "aggregate.ut_plus_tt.ndcg_at_10_mean", "aggregate.ut_plus_tt.ndcg_at_10_std", "aggregate.ut_plus_tt.recall_at_10_mean", "aggregate.ut_plus_tt.recall_at_10_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=240", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "WeChat article ranking", "metric": "Clicks", "lift_percent": 1.961, "traffic": "180M participants", "source_url": "https://arxiv.org/abs/2412.06308", "source_location": "original paper online-result disclosure: product=WeChat article ranking; metric=Clicks" }, { "product": "WeChat article ranking", "metric": "Shares", "lift_percent": 1.433, "traffic": "180M participants", "source_url": "https://arxiv.org/abs/2412.06308", "source_location": "original paper online-result disclosure: product=WeChat article ranking; metric=Shares" }, { "product": "WeChat share U2I recall", "metric": "Shares", "lift_percent": 2.56, "traffic": "26M participants", "source_url": "https://arxiv.org/abs/2412.06308", "source_location": "original paper online-result disclosure: product=WeChat share U2I recall; metric=Shares" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "primal-dual-decoding", "title": "Stochastic Primal-Dual Decoding for Multiobjective Generative Recommender Systems", "paper_url": "https://arxiv.org/abs/2607.19357", "detail_path": "reproductions/2607.19357-primal-dual-decoding/README.md", "topic": [ "generative-recommendation", "multiobjective", "constrained-decoding" ], "first_author": null, "first_author_affiliation": "Spotify", "published": "2026-05-26", "code": null, "adapter": { "adapter_key": "primal-dual-decoding", "arxiv_id": "2607.19357", "title": "Stochastic Primal-Dual Decoding for Multiobjective Generative Recommender Systems", "paper_url": "https://arxiv.org/abs/2607.19357", "track": "recommendation", "organization": "Spotify", "published": "2026-05-26", "code_url": null, "topics": [ "generative-recommendation", "multiobjective", "constrained-decoding" ], "local_code_dir": "src/auto_research/reproductions/primal_dual_decoding", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Spotify recommender", "metric": "Auxiliary objective", "lift_percent": 1.8, "traffic": "large-scale online A/B; zero user-satisfaction cost", "source_url": "https://arxiv.org/html/2607.19357v1", "source_location": "Section 5.2", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "reward:primal-dual" ] } }, { "domain": "recommendation", "key": "prl-puts", "title": "A Production-Ready RL Framework for Personalized Utility Tuning with Pareto Sweeping in Pinterest Recommender Systems", "paper_url": "https://arxiv.org/abs/2605.16344", "detail_path": "reproductions/2605.16344-prl-puts/README.md", "topic": [ "reinforcement-learning", "multi-objective-ranking", "homefeed" ], "first_author": null, "first_author_affiliation": "Pinterest", "published": "2026-05-08", "code": null, "adapter": { "adapter_key": "prl-puts", "arxiv_id": "2605.16344", "title": "A Production-Ready RL Framework for Personalized Utility Tuning with Pareto Sweeping in Pinterest Recommender Systems", "paper_url": "https://arxiv.org/abs/2605.16344", "track": "recommendation", "organization": "Pinterest", "published": "2026-05-08", "code_url": null, "topics": [ "reinforcement-learning", "multi-objective-ranking", "homefeed" ], "local_code_dir": "src/auto_research/reproductions/prl_puts", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Pinterest Homefeed", "metric": "P2P impressions", "lift_percent": 0.3, "traffic": "1% per arm, two weeks", "source_url": "https://arxiv.org/html/2605.16344v1", "source_location": "Section 6.4 / Table 1", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "promise", "title": "PROMISE: Process Reward Models Unlock Test-Time Scaling Laws in Generative Recommendations", "paper_url": "https://arxiv.org/abs/2601.04674", "detail_path": "reproductions/2601.04674-promise/README.md", "topic": [ "generative-recommendation", "process-reward", "test-time-scaling" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-01-08", "code": null, "adapter": { "adapter_key": "promise", "arxiv_id": "2601.04674", "title": "PROMISE: Process Reward Models Unlock Test-Time Scaling Laws in Generative Recommendations", "paper_url": "https://arxiv.org/abs/2601.04674", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-01-08", "code_url": null, "topics": [ "generative-recommendation", "process-reward", "test-time-scaling" ], "local_code_dir": "src/auto_research/reproductions/promise", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou generative recommendation", "metric": "Online core metric", "lift_percent": 5.0, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2601.04674v1", "source_location": "Section 4.3", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "reward:process-reward" ] } }, { "domain": "recommendation", "key": "prompt-generation", "title": "Prompt Generation Technical Report", "paper_url": "https://arxiv.org/abs/2607.11326", "detail_path": "reproductions/2607.11326-prompt-generation/README.md", "topic": [ "llm-recommendation", "generative-retrieval", "autoresearch", "serving" ], "first_author": null, "first_author_affiliation": "Alibaba / Taobao Search", "published": "2026-07", "code": null, "adapter": { "adapter_key": "prompt-generation", "arxiv_id": "2607.11326", "title": "Prompt Generation Technical Report", "paper_url": "https://arxiv.org/abs/2607.11326", "track": "recommendation", "organization": "Alibaba / Taobao Search", "published": "2026-07", "code_url": null, "topics": [ "llm-recommendation", "generative-retrieval", "autoresearch", "serving" ], "local_code_dir": "src/auto_research/reproductions/prompt_generation", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MiniOneRec Amazon Office_Products" ], "baseline": "相同 Qwen2.5-0.5B、相同步数和候选集下的 SID only", "metrics": [ "variants.sid_only.final_loss", "variants.sid_only.initial_loss", "variants.sid_plus_merged_title_brand.final_loss", "variants.sid_plus_merged_title_brand.initial_loss", "variants.sid_plus_title.final_loss", "variants.sid_plus_title.initial_loss" ], "default_seeds": [ 42 ], "budget": "steps=20", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao Search", "metric": "transaction count", "lift_percent": 0.47, "traffic": "1% traffic for 14 days", "source_url": "https://arxiv.org/abs/2607.11326", "source_location": "original paper online-result disclosure: product=Taobao Search; metric=transaction count" }, { "product": "Taobao Search", "metric": "GMV", "lift_percent": 0.51, "traffic": "1% traffic for 14 days", "source_url": "https://arxiv.org/abs/2607.11326", "source_location": "original paper online-result disclosure: product=Taobao Search; metric=GMV" }, { "product": "Taobao Recommendation Newdetail", "metric": "IPV", "lift_percent": 0.66, "traffic": "2% traffic for 12 days", "source_url": "https://arxiv.org/abs/2607.11326", "source_location": "original paper online-result disclosure: product=Taobao Recommendation Newdetail; metric=IPV" }, { "product": "Taobao Shop Search", "metric": "transaction count", "lift_percent": 4.01, "traffic": "10% traffic for over two weeks", "source_url": "https://arxiv.org/abs/2607.11326", "source_location": "original paper online-result disclosure: product=Taobao Shop Search; metric=transaction count" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "proximity-features", "title": "Proximity Features: Privacy-Compliant Cold-Start Personalization at Airbnb", "paper_url": "https://arxiv.org/abs/2607.12246", "detail_path": "reproductions/2607.12246-proximity-features/README.md", "topic": [ "cold-start", "privacy", "features" ], "first_author": null, "first_author_affiliation": "Airbnb", "published": "2026-07-14", "code": null, "adapter": { "adapter_key": "proximity-features", "arxiv_id": "2607.12246", "title": "Proximity Features: Privacy-Compliant Cold-Start Personalization at Airbnb", "paper_url": "https://arxiv.org/abs/2607.12246", "track": "recommendation", "organization": "Airbnb", "published": "2026-07-14", "code_url": null, "topics": [ "cold-start", "privacy", "features" ], "local_code_dir": "src/auto_research/reproductions/proximity_features", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "movielens 100k" ], "baseline": "global cold-start aggregate", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Airbnb marketing landing", "metric": "first-time bookers", "lift_percent": 2.0, "traffic": "experiments 2024-12 to 2026-03", "source_url": "https://arxiv.org/abs/2607.12246", "source_location": "original paper online-result disclosure: product=Airbnb marketing landing; metric=first-time bookers" }, { "product": "Airbnb AutoSuggest", "metric": "global bookings", "lift_percent": 0.16, "traffic": "experiments 2024-12 to 2026-03", "source_url": "https://arxiv.org/abs/2607.12246", "source_location": "original paper online-result disclosure: product=Airbnb AutoSuggest; metric=global bookings" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "psg", "title": "PSG: Pair-Space Generation for Efficient Generative Reranking", "paper_url": "https://arxiv.org/abs/2607.26427", "detail_path": "reproductions/2607.26427-psg/README.md", "topic": [ "reranking", "generative-recommendation", "pair-token", "serving" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-07-29", "code": null, "adapter": { "adapter_key": "psg", "arxiv_id": "2607.26427", "title": "PSG: Pair-Space Generation for Efficient Generative Reranking", "paper_url": "https://arxiv.org/abs/2607.26427", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-07-29", "code_url": null, "topics": [ "reranking", "generative-recommendation", "pair-token", "serving" ], "local_code_dir": "src/auto_research/reproductions/psg", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "item-space autoregressive scorer", "metrics": [ "hit_at_10", "ndcg_at_10", "decode_step_reduction", "duplicate_items" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; six-item slate", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou single-column feed", "metric": "Stay Time per user", "lift_percent": 0.178, "traffic": "7-day test; two disjoint 10% traffic buckets", "source_url": "https://arxiv.org/html/2607.26427v1", "source_location": "Section 5 online A/B test, Table 8", "retrieved_at": "2026-09-01" } ], "selection_exception": null, "evolve_operators": [ "head:pair-space" ] } }, { "domain": "recommendation", "key": "pushdualgen", "title": "PushDualGen: Enabling LLMs to Generate Semantic IDs with Interpretable Copy for Industrial Push Recommendation", "paper_url": "https://arxiv.org/abs/2608.07989", "detail_path": "reproductions/2608.07989-pushdualgen/README.md", "topic": [ "generative-recommendation", "semantic-id", "llm-recommendation", "push" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-08-08", "code": null, "adapter": { "adapter_key": "pushdualgen", "arxiv_id": "2608.07989", "title": "PushDualGen: Enabling LLMs to Generate Semantic IDs with Interpretable Copy for Industrial Push Recommendation", "paper_url": "https://arxiv.org/abs/2608.07989", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-08-08", "code_url": null, "topics": [ "generative-recommendation", "semantic-id", "llm-recommendation", "push" ], "local_code_dir": "src/auto_research/reproductions/pushdualgen", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou Push", "metric": "effective play rate", "lift_percent": 8.5, "traffic": "15% traffic, 150M users, 14 days", "source_url": "https://arxiv.org/html/2608.07989v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "qgs", "title": "From Item-Only to Query-Item: Query-Conditioned Generative Search with QGS in Quark", "paper_url": "https://arxiv.org/abs/2605.25514", "detail_path": "reproductions/2605.25514-qgs/README.md", "topic": [ "generative-search", "search-ranking", "long-sequence" ], "first_author": null, "first_author_affiliation": "University of Science and Technology of China / Alibaba", "published": "2026-05-25", "code": null, "adapter": { "adapter_key": "qgs", "arxiv_id": "2605.25514", "title": "From Item-Only to Query-Item: Query-Conditioned Generative Search with QGS in Quark", "paper_url": "https://arxiv.org/abs/2605.25514", "track": "recommendation", "organization": "University of Science and Technology of China / Alibaba", "published": "2026-05-25", "code_url": null, "topics": [ "generative-search", "search-ranking", "long-sequence" ], "local_code_dir": "src/auto_research/reproductions/qgs", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Quark Search", "metric": "CTR", "lift_percent": 0.62, "traffic": "2% traffic, 7 days", "source_url": "https://arxiv.org/html/2605.25514v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "quasid", "title": "Stop Treating Collisions Equally: Qualification-Aware Semantic ID Learning for Recommendation at Industrial Scale", "paper_url": "https://arxiv.org/abs/2603.00632", "detail_path": "reproductions/2603.00632-quasid/README.md", "topic": [ "semantic-id", "collision-handling", "e-commerce" ], "first_author": null, "first_author_affiliation": "University of Electronic Science and Technology of China / Kuaishou", "published": "2026-02-28", "code": null, "adapter": { "adapter_key": "quasid", "arxiv_id": "2603.00632", "title": "Stop Treating Collisions Equally: Qualification-Aware Semantic ID Learning for Recommendation at Industrial Scale", "paper_url": "https://arxiv.org/abs/2603.00632", "track": "recommendation", "organization": "University of Electronic Science and Technology of China / Kuaishou", "published": "2026-02-28", "code_url": null, "topics": [ "semantic-id", "collision-handling", "e-commerce" ], "local_code_dir": "src/auto_research/reproductions/quasid", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou e-commerce", "metric": "GMV-S2", "lift_percent": 2.38, "traffic": "5% traffic, five days", "source_url": "https://arxiv.org/html/2603.00632v1", "source_location": "Section 5.4", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "radar", "title": "RADAR: Recall Augmentation through Deferred Asynchronous Retrieval", "paper_url": "https://arxiv.org/abs/2506.07261", "detail_path": "reproductions/2506.07261-radar/README.md", "topic": [ "retrieval", "asynchronous-serving", "ranking" ], "first_author": null, "first_author_affiliation": "Meta", "published": "2025-06-08", "code": null, "adapter": { "adapter_key": "radar", "arxiv_id": "2506.07261", "title": "RADAR: Recall Augmentation through Deferred Asynchronous Retrieval", "paper_url": "https://arxiv.org/abs/2506.07261", "track": "recommendation", "organization": "Meta", "published": "2025-06-08", "code_url": null, "topics": [ "retrieval", "asynchronous-serving", "ranking" ], "local_code_dir": "src/auto_research/reproductions/radar", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享 transition + content scorer，实验组只加入 radar 核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "paper_results.recall_at_200_multiplier", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta recommendation product", "metric": "Engagement", "lift_percent": 0.8, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2506.07261", "source_location": "original paper online-result disclosure: product=Meta recommendation product; metric=Engagement" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "rag-generation", "title": "Recommendation as Generation: Unifying Personalized Video Generation and Recommendation at Industrial Scale", "paper_url": "https://arxiv.org/abs/2606.25496", "detail_path": "reproductions/2606.25496-rag-generation/README.md", "topic": [ "generative-recommendation", "semantic-id", "video-generation" ], "first_author": null, "first_author_affiliation": "Kuaishou / Beihang University", "published": "2026-06-24", "code": null, "adapter": { "adapter_key": "rag-generation", "arxiv_id": "2606.25496", "title": "Recommendation as Generation: Unifying Personalized Video Generation and Recommendation at Industrial Scale", "paper_url": "https://arxiv.org/abs/2606.25496", "track": "recommendation", "organization": "Kuaishou / Beihang University", "published": "2026-06-24", "code_url": null, "topics": [ "generative-recommendation", "semantic-id", "video-generation" ], "local_code_dir": "src/auto_research/reproductions/rag_generation", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "fixed-catalog generative recommendation", "metrics": [ "hit_at_10", "ndcg_at_10", "sid_levels", "sid_width" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou advertising", "metric": "Ad Revenue vs GRM", "lift_percent": 1.87, "traffic": "large-scale production A/B test", "source_url": "https://arxiv.org/html/2606.25496v1", "source_location": "Section 4.1, Table 1", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "head:disentangled-sid" ] } }, { "domain": "recommendation", "key": "rag-qac", "title": "Unifying Ranking and Generation in Query Auto-Completion via Retrieval-Augmented Generation and Multi-Objective Alignment", "paper_url": "https://arxiv.org/abs/2602.01023", "detail_path": "reproductions/2602.01023-rag-qac/README.md", "topic": [ "query-auto-completion", "retrieval-augmented-generation", "preference-alignment" ], "first_author": null, "first_author_affiliation": "Apple", "published": "2026-02-01", "code": null, "adapter": { "adapter_key": "rag-qac", "arxiv_id": "2602.01023", "title": "Unifying Ranking and Generation in Query Auto-Completion via Retrieval-Augmented Generation and Multi-Objective Alignment", "paper_url": "https://arxiv.org/abs/2602.01023", "track": "recommendation", "organization": "Apple", "published": "2026-02-01", "code_url": null, "topics": [ "query-auto-completion", "retrieval-augmented-generation", "preference-alignment" ], "local_code_dir": "src/auto_research/reproductions/rag_qac", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "production query auto-completion", "metric": "suggestions taken", "lift_percent": 3.46, "traffic": "10% production traffic", "source_url": "https://arxiv.org/html/2602.01023v1", "source_location": "Section 7 / Table 3", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "ramp", "title": "RAMP: Robust Ad Recommendation Under Limited Personalized-Feature Availability via Masking and Alignment Pathways", "paper_url": "https://arxiv.org/abs/2607.17473", "detail_path": "reproductions/2607.17473-ramp/README.md", "topic": [ "advertising", "privacy", "distillation", "missing-features" ], "first_author": null, "first_author_affiliation": "Huawei Ireland Research Center / University College Dublin", "published": "2026-07-20", "code": "https://github.com/Ruixinhua/RAMP", "adapter": { "adapter_key": "ramp", "arxiv_id": "2607.17473", "title": "RAMP: Robust Ad Recommendation Under Limited Personalized-Feature Availability via Masking and Alignment Pathways", "paper_url": "https://arxiv.org/abs/2607.17473", "track": "recommendation", "organization": "Huawei Ireland Research Center / University College Dublin", "published": "2026-07-20", "code_url": "https://github.com/Ruixinhua/RAMP", "topics": [ "advertising", "privacy", "distillation", "missing-features" ], "local_code_dir": "src/auto_research/reproductions/ramp", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "缺失个性化字段的共享 CTR tower，实验组为 RAMP masked dual tower", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Huawei industrial CVR prediction", "metric": "Total advertiser value", "lift_percent": 3.0, "traffic": "production A/B; paper reports over 3%", "source_url": "https://arxiv.org/abs/2607.17473", "source_location": "original paper online-result disclosure: product=Huawei industrial CVR prediction; metric=Total advertiser value" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "rankgraph2", "title": "RankGraph-2: Lifecycle Co-Design for Billion-Node Graph Learning in Recommendation", "paper_url": "https://arxiv.org/abs/2606.18379", "detail_path": "reproductions/2606.18379-rankgraph2/README.md", "topic": [ "graph-retrieval", "ppr", "residual-quantization" ], "first_author": null, "first_author_affiliation": "Meta", "published": "2026-06-16", "code": null, "adapter": { "adapter_key": "rankgraph2", "arxiv_id": "2606.18379", "title": "RankGraph-2: Lifecycle Co-Design for Billion-Node Graph Learning in Recommendation", "paper_url": "https://arxiv.org/abs/2606.18379", "track": "recommendation", "organization": "Meta", "published": "2026-06-16", "code_url": null, "topics": [ "graph-retrieval", "ppr", "residual-quantization" ], "local_code_dir": "src/auto_research/reproductions/rankgraph2", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (260 users / 420 items)" ], "baseline": "one-hop retrieval，实验组为 debiased PPR + 两级 residual cluster index", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta recommendation products", "metric": "CTR", "lift_percent": 0.96, "traffic": "20+ production launches", "source_url": "https://arxiv.org/abs/2606.18379", "source_location": "original paper online-result disclosure: product=Meta recommendation products; metric=CTR" }, { "product": "Meta recommendation products", "metric": "CVR", "lift_percent": 2.75, "traffic": "20+ production launches", "source_url": "https://arxiv.org/abs/2606.18379", "source_location": "original paper online-result disclosure: product=Meta recommendation products; metric=CVR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "rankmixer", "title": "RankMixer: Scaling Up Ranking Models in Industrial Recommenders", "paper_url": "https://arxiv.org/abs/2507.15551", "detail_path": "reproductions/2507.15551-rankmixer/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "rankmixer", "arxiv_id": "2507.15551", "title": "RankMixer: Scaling Up Ranking Models in Industrial Recommenders", "paper_url": "https://arxiv.org/abs/2507.15551", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/rankmixer", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens 100K" ], "baseline": "Shared FFN", "metrics": [ "results.rankmixer_dense.hit_at_10", "results.rankmixer_dense.ndcg_at_10", "results.rankmixer_smoe.hit_at_10", "results.rankmixer_smoe.ndcg_at_10", "results.shared_ffn.hit_at_10", "results.shared_ffn.ndcg_at_10", "sparse_vs_shared_ndcg_gain_percent" ], "default_seeds": [ 42 ], "budget": "steps=240", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "rankup", "title": "RankUp: Towards High-rank Representations for Large Scale Advertising Recommender Systems", "paper_url": "https://arxiv.org/abs/2604.17878", "detail_path": "reproductions/2604.17878-rankup/README.md", "topic": [ "advertising", "representation-collapse", "feature-interaction" ], "first_author": null, "first_author_affiliation": "Tencent", "published": "2026-04-20", "code": null, "adapter": { "adapter_key": "rankup", "arxiv_id": "2604.17878", "title": "RankUp: Towards High-rank Representations for Large Scale Advertising Recommender Systems", "paper_url": "https://arxiv.org/abs/2604.17878", "track": "recommendation", "organization": "Tencent", "published": "2026-04-20", "code_url": null, "topics": [ "advertising", "representation-collapse", "feature-interaction" ], "local_code_dir": "src/auto_research/reproductions/rankup", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Weixin advertising", "metric": "GMV", "lift_percent": 4.81, "traffic": "fully deployed across three products", "source_url": "https://arxiv.org/html/2604.17878v1", "source_location": "Abstract and Section 1", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "head:high-rank-representation" ] } }, { "domain": "recommendation", "key": "rclrec", "title": "RCLRec: Reverse Curriculum Learning for Modeling Sparse Conversions in Generative Recommendation", "paper_url": "https://arxiv.org/abs/2603.28124", "detail_path": "reproductions/2603.28124-rclrec/README.md", "topic": [ "generative-recommendation", "conversion-modeling", "curriculum-learning" ], "first_author": null, "first_author_affiliation": "Alibaba International Digital Commerce Group", "published": "2026-03-30", "code": null, "adapter": { "adapter_key": "rclrec", "arxiv_id": "2603.28124", "title": "RCLRec: Reverse Curriculum Learning for Modeling Sparse Conversions in Generative Recommendation", "paper_url": "https://arxiv.org/abs/2603.28124", "track": "recommendation", "organization": "Alibaba International Digital Commerce Group", "published": "2026-03-30", "code_url": null, "topics": [ "generative-recommendation", "conversion-modeling", "curriculum-learning" ], "local_code_dir": "src/auto_research/reproductions/rclrec", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Alibaba advertising recommendation", "metric": "Advertising revenue", "lift_percent": 2.09, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2603.28124v1", "source_location": "Abstract", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "context:reverse-curriculum" ] } }, { "domain": "recommendation", "key": "real-estate-rerank", "title": "LLM-Based Re-Ranking for Real Estate Search", "paper_url": "https://arxiv.org/abs/2607.14835", "detail_path": "reproductions/2607.14835-real-estate-rerank/README.md", "topic": [ "search-ranking", "llm-reranking", "content-understanding" ], "first_author": null, "first_author_affiliation": "QuintoAndar", "published": "2026-07-16", "code": null, "adapter": { "adapter_key": "real-estate-rerank", "arxiv_id": "2607.14835", "title": "LLM-Based Re-Ranking for Real Estate Search", "paper_url": "https://arxiv.org/abs/2607.14835", "track": "recommendation", "organization": "QuintoAndar", "published": "2026-07-16", "code_url": null, "topics": [ "search-ranking", "llm-reranking", "content-understanding" ], "local_code_dir": "src/auto_research/reproductions/real_estate_rerank", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "QuintoAndar conversational search", "metric": "CTR", "lift_percent": 5.3, "traffic": "production A/B", "source_url": "https://arxiv.org/html/2607.14835v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "rec-distill", "title": "Rec-Distill: An Industrial Distillation Pipeline for Large-Scale Recommendation Models", "paper_url": "https://arxiv.org/abs/2605.29755", "detail_path": "reproductions/2605.29755-rec-distill/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "rec-distill", "arxiv_id": "2605.29755", "title": "Rec-Distill: An Industrial Distillation Pipeline for Large-Scale Recommendation Models", "paper_url": "https://arxiv.org/abs/2605.29755", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/rec_distill", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens 100K" ], "baseline": "相同日程训练的 Raw Student", "metrics": [ "results.large_teacher.hit_at_10", "results.large_teacher.ndcg_at_10", "results.raw_student.hit_at_10", "results.raw_student.ndcg_at_10", "results.rec_distill_student.hit_at_10", "results.rec_distill_student.ndcg_at_10", "validation_ndcg_by_alpha.0.1", "validation_ndcg_by_alpha.0.3", "validation_ndcg_by_alpha.1.0", "validation_ndcg_by_alpha.3.0" ], "default_seeds": [ 42 ], "budget": "steps=160", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "recap", "title": "RECAP: Feedback-Driven Streaming Semantic User Profiles for Short-Video Recommendation", "paper_url": "https://arxiv.org/abs/2607.15730", "detail_path": "reproductions/2607.15730-recap/README.md", "topic": [ "llm-recommendation", "semantic-profile", "grpo" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology / USTC", "published": "2026-07-17", "code": null, "adapter": { "adapter_key": "recap", "arxiv_id": "2607.15730", "title": "RECAP: Feedback-Driven Streaming Semantic User Profiles for Short-Video Recommendation", "paper_url": "https://arxiv.org/abs/2607.15730", "track": "recommendation", "organization": "Kuaishou Technology / USTC", "published": "2026-07-17", "code_url": null, "topics": [ "llm-recommendation", "semantic-profile", "grpo" ], "local_code_dir": "src/auto_research/reproductions/recap", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "open-loop SFT 流式画像", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "ndcg_at_10_relative_percent", "training.grpo_reward_first", "training.grpo_reward_last", "training.sft_final_loss" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou short-video recommendation", "metric": "average application usage time per user", "lift_percent": 0.139, "traffic": "7-day online A/B; statistically significant", "source_url": "https://arxiv.org/abs/2607.15730", "source_location": "original paper online-result disclosure: product=Kuaishou short-video recommendation; metric=average application usage time per user" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "recevolve", "title": "RecEvolve: A Knowledge-Driven Autonomous Agent System for Recommender Systems", "paper_url": "https://arxiv.org/abs/2609.01622", "detail_path": "reproductions/2609.01622-recevolve/README.md", "topic": [ "auto-research", "agent", "two-tower", "reward-hacking" ], "first_author": null, "first_author_affiliation": "Google", "published": "2026-07-20", "code": null, "adapter": { "adapter_key": "recevolve", "arxiv_id": "2609.01622", "title": "RecEvolve: A Knowledge-Driven Autonomous Agent System for Recommender Systems", "paper_url": "https://arxiv.org/abs/2609.01622", "track": "recommendation", "organization": "Google", "published": "2026-07-20", "code_url": null, "topics": [ "auto-research", "agent", "two-tower", "reward-hacking" ], "local_code_dir": "src/auto_research/reproductions/recevolve", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "fixed Two-Tower-style similarity", "metrics": [ "hit_at_10", "ndcg_at_10", "rollbacks", "reward hacks blocked" ], "default_seeds": [ 42, 43, 44 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Google production Two-Tower retrieval", "metric": "user satisfaction", "lift_percent": 3.77, "traffic": "live production A/B champion versus production baseline", "source_url": "https://arxiv.org/pdf/2609.01622v1", "source_location": "Section 5.7, Table 2", "retrieved_at": "2026-09-06" } ], "selection_exception": null, "evolve_operators": [ "controller:recevolve-knowledge-vcs-loop" ] } }, { "domain": "recommendation", "key": "recgpt-mobile", "title": "RecGPT-Mobile: On-Device Large Language Models for User Intent Understanding in Taobao Feed Recommendation", "paper_url": "https://arxiv.org/abs/2605.04726", "detail_path": "reproductions/2605.04726-recgpt-mobile/README.md", "topic": [ "llm-recommendation", "on-device", "intent-understanding", "serving-efficiency" ], "first_author": null, "first_author_affiliation": "Alibaba / Taobao", "published": "2026-05-06", "code": null, "adapter": { "adapter_key": "recgpt-mobile", "arxiv_id": "2605.04726", "title": "RecGPT-Mobile: On-Device Large Language Models for User Intent Understanding in Taobao Feed Recommendation", "paper_url": "https://arxiv.org/abs/2605.04726", "track": "recommendation", "organization": "Alibaba / Taobao", "published": "2026-05-06", "code_url": null, "topics": [ "llm-recommendation", "on-device", "intent-understanding", "serving-efficiency" ], "local_code_dir": "src/auto_research/reproductions/recgpt_mobile", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 1M next-intent proxy" ], "baseline": "同一 SmolLM2-135M-Instruct 的零样本 intent completion", "metrics": [ "baseline.examples", "baseline.mean_latency_ms", "baseline.primary_intent_accuracy", "baseline.semantic_intent_accuracy", "method.examples", "method.mean_latency_ms", "method.primary_intent_accuracy", "method.semantic_intent_accuracy", "paper_results.online_GMV_percent", "quantized_method.mean_latency_ms", "quantized_method.primary_intent_accuracy", "quantized_method.semantic_intent_accuracy", "relative.lora_semantic_accuracy_percent", "relative.quantized_vs_lora_semantic_percent", "relative.serialized_size_reduction_percent", "training.final_loss", "training.initial_loss" ], "default_seeds": [ 42 ], "budget": "steps=80", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Mobile Taobao feed", "metric": "CLICK", "lift_percent": 1.8, "traffic": "one-month test across four scenarios", "source_url": "https://arxiv.org/abs/2605.04726", "source_location": "original paper online-result disclosure: product=Mobile Taobao feed; metric=CLICK" }, { "product": "Mobile Taobao feed", "metric": "PAY", "lift_percent": 2.7, "traffic": "one-month test across four scenarios", "source_url": "https://arxiv.org/abs/2605.04726", "source_location": "original paper online-result disclosure: product=Mobile Taobao feed; metric=PAY" }, { "product": "Mobile Taobao feed", "metric": "GMV", "lift_percent": 2.5, "traffic": "one-month test across four scenarios", "source_url": "https://arxiv.org/abs/2605.04726", "source_location": "original paper online-result disclosure: product=Mobile Taobao feed; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "recgpt-v2", "title": "RecGPT-V2 Technical Report", "paper_url": "https://arxiv.org/abs/2512.14503", "detail_path": "reproductions/2512.14503-recgpt-v2/README.md", "topic": [ "llm-recommendation", "multi-agent", "reinforcement-learning", "ranking" ], "first_author": null, "first_author_affiliation": "Alibaba / Taobao", "published": "2025-12-16", "code": null, "adapter": { "adapter_key": "recgpt-v2", "arxiv_id": "2512.14503", "title": "RecGPT-V2 Technical Report", "paper_url": "https://arxiv.org/abs/2512.14503", "track": "recommendation", "organization": "Alibaba / Taobao", "published": "2025-12-16", "code_url": null, "topics": [ "llm-recommendation", "multi-agent", "reinforcement-learning", "ranking" ], "local_code_dir": "src/auto_research/reproductions/recgpt_v2", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "RecGPT-V1 single-route interest representation", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.mrr_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.mrr_at_10", "method.ndcg_at_10", "paper_results.ctr_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.mrr_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao", "metric": "CTR", "lift_percent": 2.98, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2512.14503", "source_location": "original paper online-result disclosure: product=Taobao; metric=CTR" }, { "product": "Taobao", "metric": "IPV", "lift_percent": 3.71, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2512.14503", "source_location": "original paper online-result disclosure: product=Taobao; metric=IPV" }, { "product": "Taobao", "metric": "NER", "lift_percent": 11.46, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2512.14503", "source_location": "original paper online-result disclosure: product=Taobao; metric=NER" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "recgpt-v3", "title": "RecGPT-V3 Technical Report", "paper_url": "https://arxiv.org/abs/2607.15591", "detail_path": "reproductions/2607.15591-recgpt-v3/README.md", "topic": [ "generative-recommendation", "semantic-id", "user-memory", "latent-reasoning" ], "first_author": null, "first_author_affiliation": "Alibaba / Taobao", "published": "2026-07-17", "code": null, "adapter": { "adapter_key": "recgpt-v3", "arxiv_id": "2607.15591", "title": "RecGPT-V3 Technical Report", "paper_url": "https://arxiv.org/abs/2607.15591", "track": "recommendation", "organization": "Alibaba / Taobao", "published": "2026-07-17", "code_url": null, "topics": [ "generative-recommendation", "semantic-id", "user-memory", "latent-reasoning" ], "local_code_dir": "src/auto_research/reproductions/recgpt_v3", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 1M" ], "baseline": "同一 MovieLens-1M split、item tower 和训练步数下，重算完整 40-event 历史的无状态 hybrid Transformer", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "diagnostics.memory_input_tokens", "diagnostics.memory_token_reduction_percent", "diagnostics.memory_traceable_fraction", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "paper_results.feed_CTR_percent", "paper_results.feed_GMV_percent", "paper_results.item_CTR_percent", "paper_results.item_GMV_percent", "paper_results.memory_compute_reduction_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "rqvae.final_loss", "rqvae.initial_loss", "setup.memory_slots" ], "default_seeds": [ 42 ], "budget": "steps=180", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao Guess What You Like (feed)", "metric": "IPV", "lift_percent": 1.28, "traffic": "1% treatment vs 1% RecGPT-V2 control", "source_url": "https://arxiv.org/abs/2607.15591", "source_location": "original paper online-result disclosure: product=Taobao Guess What You Like (feed); metric=IPV" }, { "product": "Taobao Guess What You Like (feed)", "metric": "CTR", "lift_percent": 1.0, "traffic": "1% treatment vs 1% RecGPT-V2 control", "source_url": "https://arxiv.org/abs/2607.15591", "source_location": "original paper online-result disclosure: product=Taobao Guess What You Like (feed); metric=CTR" }, { "product": "Taobao Guess What You Like (feed)", "metric": "GMV", "lift_percent": 3.97, "traffic": "1% treatment vs 1% RecGPT-V2 control", "source_url": "https://arxiv.org/abs/2607.15591", "source_location": "original paper online-result disclosure: product=Taobao Guess What You Like (feed); metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "recharness", "title": "RecHarness: A Bandit-Routed Agentic Harness for Self-Evolving Recommender Systems", "paper_url": "https://arxiv.org/abs/2607.29241", "detail_path": "reproductions/2607.29241-recharness/README.md", "topic": [ "auto-research", "bandit", "ranking" ], "first_author": null, "first_author_affiliation": "Huazhong Agricultural University (Kuaishou internship)", "published": "2026-07-31", "code": "https://github.com/6lyc/RecHarness", "adapter": { "adapter_key": "recharness", "arxiv_id": "2607.29241", "title": "RecHarness: A Bandit-Routed Agentic Harness for Self-Evolving Recommender Systems", "paper_url": "https://arxiv.org/abs/2607.29241", "track": "recommendation", "organization": "Huazhong Agricultural University (Kuaishou internship)", "published": "2026-07-31", "code_url": "https://github.com/6lyc/RecHarness", "topics": [ "auto-research", "bandit", "ranking" ], "local_code_dir": "src/auto_research/reproductions/recharness", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "short-video advertising", "metric": "ADVV", "lift_percent": 2.084, "traffic": "10% traffic, 7 days", "source_url": "https://arxiv.org/html/2607.29241v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "reco-reward", "title": "RecoReward: Recommender-Guided Multimodal Description Generation for Recommendation", "paper_url": "https://arxiv.org/abs/2607.25901", "detail_path": "reproductions/2607.25901-reco-reward/README.md", "topic": [ "llm-recommendation", "multimodal", "reinforcement-learning", "retrieval" ], "first_author": null, "first_author_affiliation": "Kuaishou / Nankai University / Chinese Academy of Sciences", "published": "2026-07-28", "code": null, "adapter": { "adapter_key": "reco-reward", "arxiv_id": "2607.25901", "title": "RecoReward: Recommender-Guided Multimodal Description Generation for Recommendation", "paper_url": "https://arxiv.org/abs/2607.25901", "track": "recommendation", "organization": "Kuaishou / Nankai University / Chinese Academy of Sciences", "published": "2026-07-28", "code_url": null, "topics": [ "llm-recommendation", "multimodal", "reinforcement-learning", "retrieval" ], "local_code_dir": "src/auto_research/reproductions/reco_reward", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "content-only semantic recall，实验组为 RAS 选择", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.validation_only_model_selection.fresh_hit_at_10", "stages.validation_only_model_selection.hit_at_10", "stages.validation_only_model_selection.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou live recommendation", "metric": "key-page effective-user penetration", "lift_percent": 0.265, "traffic": "one-week online A/B", "source_url": "https://arxiv.org/abs/2607.25901", "source_location": "original paper online-result disclosure: product=Kuaishou live recommendation; metric=key-page effective-user penetration" }, { "product": "Kuaishou live recommendation", "metric": "outflow exposure", "lift_percent": 0.791, "traffic": "one-week online A/B", "source_url": "https://arxiv.org/abs/2607.25901", "source_location": "original paper online-result disclosure: product=Kuaishou live recommendation; metric=outflow exposure" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "rest", "title": "From Language to Behavior: Scaling Sequence Transformers for Industrial Recommendation Ranking with Rec-Native Designs", "paper_url": "https://arxiv.org/abs/2609.01240", "detail_path": "reproductions/2609.01240-rest/README.md", "topic": [ "ranking-network", "long-sequence", "shared-prefix-serving" ], "first_author": null, "first_author_affiliation": "ByteDance", "published": "2026-09-01", "code": null, "adapter": { "adapter_key": "rest", "arxiv_id": "2609.01240", "title": "From Language to Behavior: Scaling Sequence Transformers for Industrial Recommendation Ranking with Rec-Native Designs", "paper_url": "https://arxiv.org/abs/2609.01240", "track": "recommendation", "organization": "ByteDance", "published": "2026-09-01", "code_url": null, "topics": [ "ranking-network", "long-sequence", "shared-prefix-serving" ], "local_code_dir": "src/auto_research/reproductions/rest", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "production advertising platform", "metric": "core revenue metric", "lift_percent": 11.93, "traffic": "one-week online A/B; fully deployed", "source_url": "https://arxiv.org/html/2609.01240v1", "source_location": "Abstract; Section 3.5", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "context:rest-dual-gate" ] } }, { "domain": "recommendation", "key": "reward-guided-decoding", "title": "Reward Guided Decoding for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2607.25344", "detail_path": "reproductions/2607.25344-reward-guided-decoding/README.md", "topic": [ "generative-recommendation", "decoding", "reward", "business-alignment" ], "first_author": null, "first_author_affiliation": "Institute of Information Engineering, Chinese Academy of Sciences", "published": "2026-07-28", "code": null, "adapter": { "adapter_key": "reward-guided-decoding", "arxiv_id": "2607.25344", "title": "Reward Guided Decoding for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2607.25344", "track": "recommendation", "organization": "Institute of Information Engineering, Chinese Academy of Sciences", "published": "2026-07-28", "code_url": null, "topics": [ "generative-recommendation", "decoding", "reward", "business-alignment" ], "local_code_dir": "src/auto_research/reproductions/reward_guided_decoding", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "generator likelihood decoding", "metrics": [ "hit_at_10", "ndcg_at_10", "guided_expected_reward", "kl_from_generator" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; beta=0.55", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou live-streaming recommendation", "metric": "watch time", "lift_percent": 0.689, "traffic": "two-week online A/B test", "source_url": "https://arxiv.org/html/2607.25344v1", "source_location": "Section 5.4 online A/B test", "retrieved_at": "2026-09-01" } ], "selection_exception": null, "evolve_operators": [ "reward:reward-guided" ] } }, { "domain": "recommendation", "key": "rgalign-rec", "title": "RGAlign-Rec: Ranking-Guided Alignment for Latent Query Reasoning in Recommendation Systems", "paper_url": "https://arxiv.org/abs/2602.12968", "detail_path": "reproductions/2602.12968-rgalign-rec/README.md", "topic": [ "llm-recommendation", "query-reasoning", "preference-alignment" ], "first_author": null, "first_author_affiliation": "Forth AI / Shopee / Singapore University of Technology and Design", "published": "2026-02-13", "code": null, "adapter": { "adapter_key": "rgalign-rec", "arxiv_id": "2602.12968", "title": "RGAlign-Rec: Ranking-Guided Alignment for Latent Query Reasoning in Recommendation Systems", "paper_url": "https://arxiv.org/abs/2602.12968", "track": "recommendation", "organization": "Forth AI / Shopee / Singapore University of Technology and Design", "published": "2026-02-13", "code_url": null, "topics": [ "llm-recommendation", "query-reasoning", "preference-alignment" ], "local_code_dir": "src/auto_research/reproductions/rgalign_rec", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "production zero-query chatbot", "metric": "CTR@3 incremental over QE-Rec", "lift_percent": 0.13, "traffic": "large-scale online A/B", "source_url": "https://arxiv.org/html/2602.12968v1", "source_location": "Section 5.4 / Table 5", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "rocs", "title": "ROCS: Request-Oriented Compute Sharing for Efficient Large-Scale Recommendation", "paper_url": "https://arxiv.org/abs/2607.27744", "detail_path": "reproductions/2607.27744-rocs/README.md", "topic": [ "retrieval", "ranking", "serving", "compute-sharing" ], "first_author": null, "first_author_affiliation": "Meta AI", "published": "2026-07-30", "code": "https://github.com/pytorch/FBGEMM/tree/main/fbgemm_gpu/experimental/ikbo", "adapter": { "adapter_key": "rocs", "arxiv_id": "2607.27744", "title": "ROCS: Request-Oriented Compute Sharing for Efficient Large-Scale Recommendation", "paper_url": "https://arxiv.org/abs/2607.27744", "track": "recommendation", "organization": "Meta AI", "published": "2026-07-30", "code_url": "https://github.com/pytorch/FBGEMM/tree/main/fbgemm_gpu/experimental/ikbo", "topics": [ "retrieval", "ranking", "serving", "compute-sharing" ], "local_code_dir": "src/auto_research/reproductions/rocs", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M (240 users / 400 items)" ], "baseline": "逐候选重复编码", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative_ndcg_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta retrieval systems", "metric": "QPS", "lift_percent": 200.0, "traffic": "deployed across ads and organic surfaces", "source_url": "https://arxiv.org/abs/2607.27744", "source_location": "original paper online-result disclosure: product=Meta retrieval systems; metric=QPS" }, { "product": "Meta short-form video ranking", "metric": "QPS", "lift_percent": 50.0, "traffic": "production deployment; LogLoss -0.5%", "source_url": "https://arxiv.org/abs/2607.27744", "source_location": "original paper online-result disclosure: product=Meta short-form video ranking; metric=QPS" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "rolegen", "title": "Awakening Dormant Users: Generative Recommendation with Counterfactual Functional Role Reasoning", "paper_url": "https://arxiv.org/abs/2602.13134", "detail_path": "reproductions/2602.13134-rolegen/README.md", "topic": [ "generative-recommendation", "counterfactual-reasoning", "dormant-users" ], "first_author": null, "first_author_affiliation": "Beihang University", "published": "2026-02-13", "code": null, "adapter": { "adapter_key": "rolegen", "arxiv_id": "2602.13134", "title": "Awakening Dormant Users: Generative Recommendation with Counterfactual Functional Role Reasoning", "paper_url": "https://arxiv.org/abs/2602.13134", "track": "recommendation", "organization": "Beihang University", "published": "2026-02-13", "code_url": null, "topics": [ "generative-recommendation", "counterfactual-reasoning", "dormant-users" ], "local_code_dir": "src/auto_research/reproductions/rolegen", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou e-commerce", "metric": "Order volume", "lift_percent": 7.3, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2602.13134v1", "source_location": "Abstract", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "reward:counterfactual-role" ] } }, { "domain": "recommendation", "key": "rq-gmm", "title": "RQ-GMM: Residual Quantized Gaussian Mixture Model for Multimodal Semantic Discretization in CTR Prediction", "paper_url": "https://arxiv.org/abs/2602.12593", "detail_path": "reproductions/2602.12593-rq-gmm/README.md", "topic": [ "multimodal-recommendation", "semantic-tokenization", "ctr-prediction" ], "first_author": null, "first_author_affiliation": "Tencent", "published": "2026-02-13", "code": null, "adapter": { "adapter_key": "rq-gmm", "arxiv_id": "2602.12593", "title": "RQ-GMM: Residual Quantized Gaussian Mixture Model for Multimodal Semantic Discretization in CTR Prediction", "paper_url": "https://arxiv.org/abs/2602.12593", "track": "recommendation", "organization": "Tencent", "published": "2026-02-13", "code_url": null, "topics": [ "multimodal-recommendation", "semantic-tokenization", "ctr-prediction" ], "local_code_dir": "src/auto_research/reproductions/rq_gmm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Tencent short-video advertising", "metric": "Advertiser value", "lift_percent": 1.502, "traffic": "large-scale production online A/B", "source_url": "https://arxiv.org/html/2602.12593v1", "source_location": "Abstract and online experiments", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "head:residual-gmm" ] } }, { "domain": "recommendation", "key": "s-grec", "title": "S-GRec: Personalized Semantic-Aware Generative Recommendation with Asymmetric Advantage", "paper_url": "https://arxiv.org/abs/2602.10606", "detail_path": "reproductions/2602.10606-s-grec/README.md", "topic": [ "generative-recommendation", "llm-as-judge", "grpo", "a2po" ], "first_author": null, "first_author_affiliation": "Tencent / WeChat Channels", "published": "2026-02", "code": null, "adapter": { "adapter_key": "s-grec", "arxiv_id": "2602.10606", "title": "S-GRec: Personalized Semantic-Aware Generative Recommendation with Asymmetric Advantage", "paper_url": "https://arxiv.org/abs/2602.10606", "track": "recommendation", "organization": "Tencent / WeChat Channels", "published": "2026-02", "code_url": null, "topics": [ "generative-recommendation", "llm-as-judge", "grpo", "a2po" ], "local_code_dir": "src/auto_research/reproductions/s_grec", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MiniOneRec Amazon Office_Products" ], "baseline": "相同公开 Office 数据、相同 SID Transformer 和 600-step 训练预算的 MiniOneRec-style business SFT checkpoint", "metrics": [ "psj.aligned_point_accuracy", "psj.pairwise_accuracy", "psj.sft_final_loss", "psj.sft_initial_loss", "psj.sft_point_accuracy", "test.a2po.ndcg_at_10", "test.adv_sum.ndcg_at_10", "test.business.ndcg_at_10", "test.reward_sum.hr_at_10", "test.reward_sum.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "WeChat Channels", "metric": "GMV", "lift_percent": 1.19, "traffic": "20% advertising traffic", "source_url": "https://arxiv.org/abs/2602.10606", "source_location": "original paper online-result disclosure: product=WeChat Channels; metric=GMV" }, { "product": "WeChat Channels", "metric": "CTR", "lift_percent": 1.16, "traffic": "20% advertising traffic", "source_url": "https://arxiv.org/abs/2602.10606", "source_location": "original paper online-result disclosure: product=WeChat Channels; metric=CTR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "s2gr", "title": "S2GR: Stepwise Semantic-Guided Reasoning in Latent Space for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2601.18664", "detail_path": "reproductions/2601.18664-s2gr/README.md", "topic": [ "generative-recommendation", "latent-reasoning", "semantic-id" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-01-26", "code": null, "adapter": { "adapter_key": "s2gr", "arxiv_id": "2601.18664", "title": "S2GR: Stepwise Semantic-Guided Reasoning in Latent Space for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2601.18664", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-01-26", "code_url": null, "topics": [ "generative-recommendation", "latent-reasoning", "semantic-id" ], "local_code_dir": "src/auto_research/reproductions/s2gr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou recommendation", "metric": "Online core metric", "lift_percent": 5.25, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2601.18664v1", "source_location": "Section 5.4", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "head:stepwise-semantic-reasoning" ] } }, { "domain": "recommendation", "key": "safro", "title": "SaFRO: Satisfaction-Aware Fusion via Dual-Relative Policy Optimization for Short-Video Search", "paper_url": "https://arxiv.org/abs/2603.19585", "detail_path": "reproductions/2603.19585-safro/README.md", "topic": [ "search-ranking", "reinforcement-learning", "multi-task-learning" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-03-20", "code": null, "adapter": { "adapter_key": "safro", "arxiv_id": "2603.19585", "title": "SaFRO: Satisfaction-Aware Fusion via Dual-Relative Policy Optimization for Short-Video Search", "paper_url": "https://arxiv.org/abs/2603.19585", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-03-20", "code_url": null, "topics": [ "search-ranking", "reinforcement-learning", "multi-task-learning" ], "local_code_dir": "src/auto_research/reproductions/safro", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou Search", "metric": "watch time", "lift_percent": 0.611, "traffic": "10% traffic, two months", "source_url": "https://arxiv.org/html/2603.19585v1", "source_location": "Section 5.4 / Table 3", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "sam", "title": "Learning to Forget: Satiation-Aware Long-Sequence Transducers for Mitigating Post-Purchase Redundancy", "paper_url": "https://arxiv.org/abs/2607.12714", "detail_path": "reproductions/2607.12714-sam/README.md", "topic": [ "ranking", "long-sequence", "satiation" ], "first_author": null, "first_author_affiliation": "Alibaba Group", "published": "2026-07-14", "code": null, "adapter": { "adapter_key": "sam", "arxiv_id": "2607.12714", "title": "Learning to Forget: Satiation-Aware Long-Sequence Transducers for Mitigating Post-Purchase Redundancy", "paper_url": "https://arxiv.org/abs/2607.12714", "track": "recommendation", "organization": "Alibaba Group", "published": "2026-07-14", "code_url": null, "topics": [ "ranking", "long-sequence", "satiation" ], "local_code_dir": "src/auto_research/reproductions/sam", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "movielens 100k" ], "baseline": "DIN-like interest score", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Alibaba recommendation", "metric": "CTR", "lift_percent": 1.1, "traffic": "1% traffic", "source_url": "https://arxiv.org/abs/2607.12714", "source_location": "original paper online-result disclosure: product=Alibaba recommendation; metric=CTR" }, { "product": "Alibaba recommendation", "metric": "GMV", "lift_percent": 0.9, "traffic": "1% traffic", "source_url": "https://arxiv.org/abs/2607.12714", "source_location": "original paper online-result disclosure: product=Alibaba recommendation; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "sarm", "title": "SARM: LLM-Augmented Semantic Anchor for End-to-End Live-Streaming Ranking", "paper_url": "https://arxiv.org/abs/2602.09401", "detail_path": "reproductions/2602.09401-sarm/README.md", "topic": [ "live-streaming", "semantic-anchor", "multimodal-recommendation" ], "first_author": null, "first_author_affiliation": "Institute of Information Engineering, CAS / Kuaishou", "published": "2026-02-10", "code": null, "adapter": { "adapter_key": "sarm", "arxiv_id": "2602.09401", "title": "SARM: LLM-Augmented Semantic Anchor for End-to-End Live-Streaming Ranking", "paper_url": "https://arxiv.org/abs/2602.09401", "track": "recommendation", "organization": "Institute of Information Engineering, CAS / Kuaishou", "published": "2026-02-10", "code_url": null, "topics": [ "live-streaming", "semantic-anchor", "multimodal-recommendation" ], "local_code_dir": "src/auto_research/reproductions/sarm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou Lite live-streaming", "metric": "watch count", "lift_percent": 1.19, "traffic": "two-week A/B; subsequently fully deployed", "source_url": "https://arxiv.org/html/2602.09401v1", "source_location": "Section 4.4 / Table 3", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "sasrec", "title": "Self-Attentive Sequential Recommendation", "paper_url": "https://arxiv.org/abs/1808.09781", "detail_path": "reproductions/1808.09781-sasrec/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": "https://github.com/kang205/SASRec", "adapter": { "adapter_key": "sasrec", "arxiv_id": "1808.09781", "title": "Self-Attentive Sequential Recommendation", "paper_url": "https://arxiv.org/abs/1808.09781", "track": "recommendation", "organization": null, "published": null, "code_url": "https://github.com/kang205/SASRec", "topics": [], "local_code_dir": "src/auto_research/reproductions/sasrec", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens-100K" ], "baseline": "popularity 排序", "metrics": [ "ndcg_gain_percent", "popularity.hit_at_10", "popularity.ndcg_at_10", "sasrec.hit_at_10", "sasrec.hit_at_10_std", "sasrec.ndcg_at_10", "sasrec.ndcg_at_10_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=320", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "saviorrec", "title": "SaviorRec: Semantic-Behavior Alignment for Cold-Start Recommendation", "paper_url": "https://arxiv.org/abs/2508.01375", "detail_path": "reproductions/2508.01375-saviorrec/README.md", "topic": [ "cold start", "semantic ID", "multimodal recommendation" ], "first_author": null, "first_author_affiliation": "Alibaba", "published": "2025-08", "code": null, "adapter": { "adapter_key": "saviorrec", "arxiv_id": "2508.01375", "title": "SaviorRec: Semantic-Behavior Alignment for Cold-Start Recommendation", "paper_url": "https://arxiv.org/abs/2508.01375", "track": "recommendation", "organization": "Alibaba", "published": "2025-08", "code_url": null, "topics": [ "cold start", "semantic ID", "multimodal recommendation" ], "local_code_dir": "src/auto_research/reproductions/saviorrec", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K genres and chronological co-interactions" ], "baseline": "等预算 Content Ranker", "metrics": [ "aggregate.auc_relative_percent", "aggregate.cold_auc_relative_percent", "aggregate.content_auc_mean", "aggregate.content_auc_std", "aggregate.content_cold_auc_mean", "aggregate.content_cold_auc_std", "aggregate.positive_auc_seeds", "aggregate.positive_cold_auc_seeds", "aggregate.saviorrec_auc_mean", "aggregate.saviorrec_auc_std", "aggregate.saviorrec_cold_auc_mean", "aggregate.saviorrec_cold_auc_std", "per_seed.42.content_auc", "per_seed.42.content_cold_auc", "per_seed.42.saviorrec_auc", "per_seed.42.saviorrec_cold_auc", "per_seed.43.content_auc", "per_seed.43.content_cold_auc", "per_seed.43.saviorrec_auc", "per_seed.43.saviorrec_cold_auc", "per_seed.44.content_auc", "per_seed.44.content_cold_auc", "per_seed.44.saviorrec_auc", "per_seed.44.saviorrec_cold_auc" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=180", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao Guess You Like cold-start", "metric": "Clicks", "lift_percent": 13.31, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2508.01375", "source_location": "original paper online-result disclosure: product=Taobao Guess You Like cold-start; metric=Clicks" }, { "product": "Taobao Guess You Like cold-start", "metric": "Orders", "lift_percent": 13.44, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2508.01375", "source_location": "original paper online-result disclosure: product=Taobao Guess You Like cold-start; metric=Orders" }, { "product": "Taobao Guess You Like cold-start", "metric": "CTR", "lift_percent": 12.8, "traffic": "production A/B", "source_url": "https://arxiv.org/abs/2508.01375", "source_location": "original paper online-result disclosure: product=Taobao Guess You Like cold-start; metric=CTR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "scalr", "title": "Synthetic Data from Cross-Domain Events for Large-Scale Recommendation Systems", "paper_url": "https://arxiv.org/abs/2606.00282", "detail_path": "reproductions/2606.00282-scalr/README.md", "topic": [ "cross-domain", "synthetic-data", "sampling", "conversion", "data-augmentation" ], "first_author": null, "first_author_affiliation": "Meta", "published": "2026-05-29", "code": null, "adapter": { "adapter_key": "scalr", "arxiv_id": "2606.00282", "title": "Synthetic Data from Cross-Domain Events for Large-Scale Recommendation Systems", "paper_url": "https://arxiv.org/abs/2606.00282", "track": "recommendation", "organization": "Meta", "published": "2026-05-29", "code_url": null, "topics": [ "cross-domain", "synthetic-data", "sampling", "conversion", "data-augmentation" ], "local_code_dir": "src/auto_research/reproductions/scalr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (220 users / 360 items; genre-derived domains)" ], "baseline": "deterministic top-k translation under the identical source events and candidate catalog", "metrics": [ "hit_at_10", "ndcg_at_10", "head_share_at_10", "synthetic_catalog_coverage" ], "default_seeds": [ 42, 43, 44 ], "budget": "two synthetic targets per source event", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta industrial recommendation platform", "metric": "conversion rate", "lift_percent": 0.14, "traffic": "live traffic over multiple weeks; paper reports a consistent +0.14% to +0.24% range", "source_url": "https://arxiv.org/html/2606.00282v1#S5.SS2", "source_location": "Section 5.2", "experiment_duration": "multiple weeks", "significance": "statistically significant", "retrieved_at": "2026-09-01" } ], "selection_exception": null, "evolve_operators": [ "data:scalr" ] } }, { "domain": "recommendation", "key": "self-evolving-rec", "title": "Self-Evolving Recommendation System: End-To-End Autonomous Model Optimization With LLM Agents", "paper_url": "https://arxiv.org/abs/2602.10226", "detail_path": "reproductions/2602.10226-self-evolving-rec/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "self-evolving-rec", "arxiv_id": "2602.10226", "title": "Self-Evolving Recommendation System: End-To-End Autonomous Model Optimization With LLM Agents", "paper_url": "https://arxiv.org/abs/2602.10226", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/self_evolving_rec", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "Human/Adagrad 配置", "metrics": [ "relative_ndcg_change_percent", "results.human_baseline.hit_at_10", "results.human_baseline.hit_at_10_std", "results.human_baseline.ndcg_at_10", "results.human_baseline.ndcg_at_10_std", "results.llm_agent_promoted.hit_at_10", "results.llm_agent_promoted.hit_at_10_std", "results.llm_agent_promoted.ndcg_at_10", "results.llm_agent_promoted.ndcg_at_10_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "semantic-native-longseq", "title": "Beyond Item IDs: Scaling Short-Form-Video Recommendation via Semantic-Native Long Sequence Modeling", "paper_url": "https://arxiv.org/abs/2606.07546", "detail_path": "reproductions/2606.07546-semantic-native-longseq/README.md", "topic": [ "long-sequence", "semantic-id", "rq-vae", "efficient-transformer", "cold-start" ], "first_author": null, "first_author_affiliation": "Google", "published": "2026-05-04", "code": null, "adapter": { "adapter_key": "semantic-native-longseq", "arxiv_id": "2606.07546", "title": "Beyond Item IDs: Scaling Short-Form-Video Recommendation via Semantic-Native Long Sequence Modeling", "paper_url": "https://arxiv.org/abs/2606.07546", "track": "recommendation", "organization": "Google", "published": "2026-05-04", "code_url": null, "topics": [ "long-sequence", "semantic-id", "rq-vae", "efficient-transformer", "cold-start" ], "local_code_dir": "src/auto_research/reproductions/semantic_native_longseq", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (220 users / 360 items)" ], "baseline": "12-event item-ID feature attention under the same public split and full catalog", "metrics": [ "hit_at_10", "ndcg_at_10", "head_share_at_10", "attention_pair_reduction" ], "default_seeds": [ 42, 43, 44 ], "budget": "48 events folded by 4 into 12 local tokens plus two global queries", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Google global short-video platform", "metric": "actively engaged users", "lift_percent": 0.52, "traffic": "large-scale online A/B; full SID + compressed-transformer system versus L=800 Video-ID baseline", "source_url": "https://arxiv.org/html/2606.07546v1#S3.SS5", "source_location": "Section 3.5, Table 4", "significance": "p < 0.05", "retrieved_at": "2026-09-01" }, { "product": "Google global short-video platform", "metric": "satisfied watch time", "lift_percent": 1.42, "traffic": "large-scale online A/B; full SID + compressed-transformer system", "source_url": "https://arxiv.org/html/2606.07546v1#S3.SS5", "source_location": "Section 3.5, Table 4", "significance": "p < 0.05", "retrieved_at": "2026-09-01" } ], "selection_exception": null, "evolve_operators": [ "rankmixer_semantic_native_longseq" ] } }, { "domain": "recommendation", "key": "sequenceo1", "title": "SequenceO1: End-to-End Ultra-Long (100K) Sequence Modeling in Recommendation with Low-Rank Caching", "paper_url": "https://arxiv.org/abs/2609.08443", "detail_path": "reproductions/2609.08443-sequenceo1/README.md", "topic": [ "long-sequence", "sketch-attention", "prototype-compression", "industrial-recommendation" ], "first_author": null, "first_author_affiliation": "ByteDance / Douyin", "published": "2026-09-08", "code": null, "adapter": { "adapter_key": "sequenceo1", "arxiv_id": "2609.08443", "title": "SequenceO1: End-to-End Ultra-Long (100K) Sequence Modeling in Recommendation with Low-Rank Caching", "paper_url": "https://arxiv.org/abs/2609.08443", "track": "recommendation", "organization": "ByteDance / Douyin", "published": "2026-09-08", "code_url": null, "topics": [ "long-sequence", "sketch-attention", "prototype-compression", "industrial-recommendation" ], "local_code_dir": "src/auto_research/reproductions/sequenceo1", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Douyin and Douyin Lite recommender", "metric": "30-day activeness", "lift_percent": 0.1968, "traffic": "one-month A/B followed by full-traffic deployment", "source_url": "https://arxiv.org/html/2609.08443", "source_location": "Section 4.3, Table 4", "experiment_duration": "one month", "significance": "all reported effects statistically significant", "retrieved_at": "2026-09-12" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "seral", "title": "Bursting Filter Bubble: Enhancing Serendipity Recommendations with Aligned Large Language Models", "paper_url": "https://arxiv.org/abs/2502.13539", "detail_path": "reproductions/2502.13539-seral/README.md", "topic": [ "llm-recommendation", "preference-alignment", "serendipity" ], "first_author": null, "first_author_affiliation": "Alibaba / Taobao", "published": "2025-02", "code": null, "adapter": { "adapter_key": "seral", "arxiv_id": "2502.13539", "title": "Bursting Filter Bubble: Enhancing Serendipity Recommendations with Aligned Large Language Models", "paper_url": "https://arxiv.org/abs/2502.13539", "track": "recommendation", "organization": "Alibaba / Taobao", "published": "2025-02", "code_url": null, "topics": [ "llm-recommendation", "preference-alignment", "serendipity" ], "local_code_dir": "src/auto_research/reproductions/seral", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens-100K" ], "baseline": "DIN，SERAL NDCG@10 相对 DIN +50.60%", "metrics": [ "results.din.hit_at_10", "results.din.ndcg_at_10", "results.ndcg_ablation_percent", "results.ndcg_vs_din_percent", "results.seral_ipo.hit_at_10", "results.seral_ipo.ndcg_at_10", "results.sft.hit_at_10", "results.sft.ndcg_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=100", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao Guess What You Like", "metric": "serendipity clicks", "lift_percent": 29.56, "traffic": "fully deployed online experiment", "source_url": "https://arxiv.org/abs/2502.13539", "source_location": "original paper online-result disclosure: product=Taobao Guess What You Like; metric=serendipity clicks" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "sessionrec", "title": "SessionRec: Next Session Prediction Paradigm For Generative Sequential Recommendation", "paper_url": "https://arxiv.org/abs/2502.10157", "detail_path": "reproductions/2502.10157-sessionrec/README.md", "topic": [ "session recommendation", "generative retrieval" ], "first_author": null, "first_author_affiliation": "Meituan", "published": "2025-02", "code": null, "adapter": { "adapter_key": "sessionrec", "arxiv_id": "2502.10157", "title": "SessionRec: Next Session Prediction Paradigm For Generative Sequential Recommendation", "paper_url": "https://arxiv.org/abs/2502.10157", "track": "recommendation", "organization": "Meituan", "published": "2025-02", "code_url": null, "topics": [ "session recommendation", "generative retrieval" ], "local_code_dir": "src/auto_research/reproductions/sessionrec", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "KuaiRand-Pure standard exposure log" ], "baseline": "等预算的扁平 Item Transformer", "metrics": [ "aggregate.item_transformer.ndcg_at_20_mean", "aggregate.item_transformer.ndcg_at_20_std", "aggregate.item_transformer.recall_at_20_mean", "aggregate.item_transformer.recall_at_20_std", "aggregate.sessionrec_ndcg_relative_percent", "aggregate.sessionrec_transformer.ndcg_at_20_mean", "aggregate.sessionrec_transformer.ndcg_at_20_std", "aggregate.sessionrec_transformer.recall_at_20_mean", "aggregate.sessionrec_transformer.recall_at_20_std", "per_seed.42.item_transformer_ndcg_at_20", "per_seed.42.sessionrec_ndcg_at_20", "per_seed.43.item_transformer_ndcg_at_20", "per_seed.43.sessionrec_ndcg_at_20", "per_seed.44.item_transformer_ndcg_at_20", "per_seed.44.sessionrec_ndcg_at_20" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=180", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meituan homepage", "metric": "Pay PV", "lift_percent": 0.603, "traffic": "7 days", "source_url": "https://arxiv.org/abs/2502.10157", "source_location": "original paper online-result disclosure: product=Meituan homepage; metric=Pay PV" }, { "product": "Meituan homepage", "metric": "PVCTCVR", "lift_percent": 0.564, "traffic": "7 days", "source_url": "https://arxiv.org/abs/2502.10157", "source_location": "original paper online-result disclosure: product=Meituan homepage; metric=PVCTCVR" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "setmir", "title": "SetMIR: Multi-Interest Retrieval as Set Prediction", "paper_url": "https://arxiv.org/abs/2608.30251", "detail_path": "reproductions/2608.30251-setmir/README.md", "topic": [ "multi-interest-retrieval", "set-prediction", "dynamic-dispatch" ], "first_author": null, "first_author_affiliation": "Snap Inc.", "published": "2026-08-31", "code": null, "adapter": { "adapter_key": "setmir", "arxiv_id": "2608.30251", "title": "SetMIR: Multi-Interest Retrieval as Set Prediction", "paper_url": "https://arxiv.org/abs/2608.30251", "track": "recommendation", "organization": "Snap Inc.", "published": "2026-08-31", "code_url": null, "topics": [ "multi-interest-retrieval", "set-prediction", "dynamic-dispatch" ], "local_code_dir": "src/auto_research/reproductions/setmir", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Snap Dynamic Product Ads", "metric": "overall CVR", "lift_percent": 3.1, "traffic": "deployed production retrieval source", "source_url": "https://arxiv.org/html/2608.30251v1", "source_location": "Abstract", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "context:setmir-query-set" ] } }, { "domain": "recommendation", "key": "sid-coord", "title": "SID-Coord: Coordinating Semantic IDs for ID-based Ranking in Short-Video Search", "paper_url": "https://arxiv.org/abs/2604.10471", "detail_path": "reproductions/2604.10471-sid-coord/README.md", "topic": [ "short-video-search", "semantic-id", "long-tail-ranking" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-04-12", "code": null, "adapter": { "adapter_key": "sid-coord", "arxiv_id": "2604.10471", "title": "SID-Coord: Coordinating Semantic IDs for ID-based Ranking in Short-Video Search", "paper_url": "https://arxiv.org/abs/2604.10471", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-04-12", "code_url": null, "topics": [ "short-video-search", "semantic-id", "long-tail-ranking" ], "local_code_dir": "src/auto_research/reproductions/sid_coord", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou short-video search", "metric": "Long-play rate", "lift_percent": 0.664, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2604.10471v1", "source_location": "Abstract", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "head:sid-coordination" ] } }, { "domain": "recommendation", "key": "sigma", "title": "SIGMA: A Semantic-Grounded Instruction-Driven Generative Multi-Task Recommender at AliExpress", "paper_url": "https://arxiv.org/abs/2602.22913", "detail_path": "reproductions/2602.22913-sigma/README.md", "topic": [ "generative-recommendation", "multi-task-sft", "hybrid-token", "semantic-grounding" ], "first_author": null, "first_author_affiliation": "Alibaba / AliExpress", "published": "2026-02", "code": null, "adapter": { "adapter_key": "sigma", "arxiv_id": "2602.22913", "title": "SIGMA: A Semantic-Grounded Instruction-Driven Generative Multi-Task Recommender at AliExpress", "paper_url": "https://arxiv.org/abs/2602.22913", "track": "recommendation", "organization": "Alibaba / AliExpress", "published": "2026-02", "code_url": null, "topics": [ "generative-recommendation", "multi-task-sft", "hybrid-token", "semantic-grounding" ], "local_code_dir": "src/auto_research/reproductions/sigma", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MiniOneRec Office" ], "baseline": "同一 grounded SmolLM、相同七任务 SFT 与 fused item vectors 下的 idonly item head", "metrics": [ "selected_relative_to_id_only_percent.ndcg_at_10", "test.apf.ndcg_at_10", "test.id_only.ndcg_at_10", "test.top1_prefix.ndcg_at_10", "training.grounding_final_loss", "training.grounding_initial_loss", "training.sft_final_loss", "training.sft_initial_loss" ], "default_seeds": [ 42 ], "budget": "steps=480", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "AliExpress", "metric": "Order Volume", "lift_percent": 2.8, "traffic": "5% traffic, 2 weeks", "source_url": "https://arxiv.org/abs/2602.22913", "source_location": "original paper online-result disclosure: product=AliExpress; metric=Order Volume" }, { "product": "AliExpress", "metric": "GMV", "lift_percent": 7.84, "traffic": "5% traffic, 2 weeks", "source_url": "https://arxiv.org/abs/2602.22913", "source_location": "original paper online-result disclosure: product=AliExpress; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "sim", "title": "Search-based User Interest Modeling with Lifelong Sequential Behavior Data for CTR Prediction", "paper_url": "https://arxiv.org/abs/2006.05639", "detail_path": "reproductions/2006.05639-sim/README.md", "topic": [ "ranking", "long-sequence", "training-serving" ], "first_author": null, "first_author_affiliation": "Alibaba", "published": "2020-06-10", "code": null, "adapter": { "adapter_key": "sim", "arxiv_id": "2006.05639", "title": "Search-based User Interest Modeling with Lifelong Sequential Behavior Data for CTR Prediction", "paper_url": "https://arxiv.org/abs/2006.05639", "track": "recommendation", "organization": "Alibaba", "published": "2020-06-10", "code_url": null, "topics": [ "ranking", "long-sequence", "training-serving" ], "local_code_dir": "src/auto_research/reproductions/sim", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "shared transition + content baseline，实验组为 SIM，只改变论文核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "paper_results.ctr_percent", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Alibaba display advertising", "metric": "CTR", "lift_percent": 7.1, "traffic": "main traffic", "source_url": "https://arxiv.org/abs/2006.05639", "source_location": "original paper online-result disclosure: product=Alibaba display advertising; metric=CTR" }, { "product": "Alibaba display advertising", "metric": "RPM", "lift_percent": 4.4, "traffic": "main traffic", "source_url": "https://arxiv.org/abs/2006.05639", "source_location": "original paper online-result disclosure: product=Alibaba display advertising; metric=RPM" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "sirf", "title": "SIRF: A Spec-Internalized Risk Foundation Model for Industrial Content Risk Control", "paper_url": "https://arxiv.org/abs/2609.11752", "detail_path": "reproductions/2609.11752-sirf/README.md", "topic": [ "content-understanding", "risk-control", "foundation-model", "industrial-deployment" ], "first_author": null, "first_author_affiliation": "Xiaohongshu", "published": "2026-09-10", "code": null, "adapter": { "adapter_key": "sirf", "arxiv_id": "2609.11752", "title": "SIRF: A Spec-Internalized Risk Foundation Model for Industrial Content Risk Control", "paper_url": "https://arxiv.org/abs/2609.11752", "track": "recommendation", "organization": "Xiaohongshu", "published": "2026-09-10", "code_url": null, "topics": [ "content-understanding", "risk-control", "foundation-model", "industrial-deployment" ], "local_code_dir": "src/auto_research/reproductions/sirf", "fidelity": "concept_demo", "evaluation_tier": "l1_mechanism", "datasets": [ "deterministic synthetic policy cases" ], "baseline": "same-source SFT linear features", "metrics": [ "precision", "black_recall", "accuracy" ], "default_seeds": [ 42, 43, 44 ], "budget": "4,000 train / 2,000 held-out policy cases", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Xiaohongshu freezing-risk adjudication", "metric": "relative mis-penalization reduction", "lift_percent": 70.0, "traffic": "random treatment/control split over the same released-user population; cumulative million-user scale", "source_url": "https://arxiv.org/html/2609.11752", "source_location": "Section 5.2", "significance": "weekly active penetration significantly better; paper reports high-single-digit to low-double-digit relative lift without an exact scalar", "retrieved_at": "2026-09-14" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "slimper", "title": "SlimPer: Make Personalization Model Slim and Smart", "paper_url": "https://arxiv.org/abs/2607.12281", "detail_path": "reproductions/2607.12281-slimper/README.md", "topic": [ "ranking", "long-sequence", "serving-efficiency" ], "first_author": null, "first_author_affiliation": "Meta Platforms, Inc.", "published": "2026-07-14", "code": null, "adapter": { "adapter_key": "slimper", "arxiv_id": "2607.12281", "title": "SlimPer: Make Personalization Model Slim and Smart", "paper_url": "https://arxiv.org/abs/2607.12281", "track": "recommendation", "organization": "Meta Platforms, Inc.", "published": "2026-07-14", "code_url": null, "topics": [ "ranking", "long-sequence", "serving-efficiency" ], "local_code_dir": "src/auto_research/reproductions/slimper", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "参数匹配的 6-layer full-sequence Transformer（138,337 参数）", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "baseline.parameters", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "method.parameters", "relative.attention_score_reduction_percent", "relative.hit_at_10_percent", "relative.intermediate_reduction_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": "User explicitly accepts the paper's statistically significant full-traffic Instagram launch as online evidence although exact lift percentages are undisclosed.", "evolve_operators": [] } }, { "domain": "recommendation", "key": "smes", "title": "SMES: Towards Scalable Multi-Task Recommendation via Expert Sparsity", "paper_url": "https://arxiv.org/abs/2602.09386", "detail_path": "reproductions/2602.09386-smes/README.md", "topic": [ "multi-task-learning", "sparse-moe", "expert-routing" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-02-10", "code": null, "adapter": { "adapter_key": "smes", "arxiv_id": "2602.09386", "title": "SMES: Towards Scalable Multi-Task Recommendation via Expert Sparsity", "paper_url": "https://arxiv.org/abs/2602.09386", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-02-10", "code_url": null, "topics": [ "multi-task-learning", "sparse-moe", "expert-routing" ], "local_code_dir": "src/auto_research/reproductions/smes", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou short-video recommendation", "metric": "User watch time", "lift_percent": 0.31, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2602.09386v1", "source_location": "Section 4.7", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "reward:expert-balance" ] } }, { "domain": "recommendation", "key": "snaplgr", "title": "LLM-Based Generative Retrieval for Snapchat Content Recommendation", "paper_url": "https://arxiv.org/abs/2607.28895", "detail_path": "reproductions/2607.28895-snaplgr/README.md", "topic": [ "retrieval", "generative-recommendation", "semantic-id", "llm", "multimodal" ], "first_author": null, "first_author_affiliation": "Snap Inc.", "published": "2026-07-30", "code": null, "adapter": { "adapter_key": "snaplgr", "arxiv_id": "2607.28895", "title": "LLM-Based Generative Retrieval for Snapchat Content Recommendation", "paper_url": "https://arxiv.org/abs/2607.28895", "track": "recommendation", "organization": "Snap Inc.", "published": "2026-07-30", "code_url": null, "topics": [ "retrieval", "generative-recommendation", "semantic-id", "llm", "multimodal" ], "local_code_dir": "src/auto_research/reproductions/snaplgr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K content/co-engagement proxy" ], "baseline": "vanilla residual-quantized semantic IDs", "metrics": [ "hit_at_10", "ndcg_at_10", "sid_utilization", "sid_collision_rate" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; 3-level width-8 SID", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Snapchat content recommendation", "metric": "View Time", "lift_percent": 0.37, "traffic": "7-day production A/B test", "source_url": "https://arxiv.org/html/2607.28895v1", "source_location": "Section 4.5, Table 8", "significance": "p = 0.007", "retrieved_at": "2026-09-01" } ], "selection_exception": null, "evolve_operators": [ "head:snaplgr-sid" ] } }, { "domain": "recommendation", "key": "solaris", "title": "SOLARIS: Speculative Offloading of Latent-bAsed Representation for Inference Scaling", "paper_url": "https://arxiv.org/abs/2604.12110", "detail_path": "reproductions/2604.12110-solaris/README.md", "topic": [ "foundation-model", "serving", "latent-cache" ], "first_author": null, "first_author_affiliation": "Meta", "published": "2026-04-13", "code": null, "adapter": { "adapter_key": "solaris", "arxiv_id": "2604.12110", "title": "SOLARIS: Speculative Offloading of Latent-bAsed Representation for Inference Scaling", "paper_url": "https://arxiv.org/abs/2604.12110", "track": "recommendation", "organization": "Meta", "published": "2026-04-13", "code_url": null, "topics": [ "foundation-model", "serving", "latent-cache" ], "local_code_dir": "src/auto_research/reproductions/solaris", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (260 users / 420 items)" ], "baseline": "request-time ranker，实验组为 predictive latent cache", "metrics": [ "baseline.fresh_hit_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta recommendation", "metric": "Top-line revenue", "lift_percent": 0.67, "traffic": "fully deployed", "source_url": "https://arxiv.org/abs/2604.12110", "source_location": "original paper online-result disclosure: product=Meta recommendation; metric=Top-line revenue" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "sona", "title": "Sona Technical Report", "paper_url": "https://arxiv.org/abs/2608.11015", "detail_path": "reproductions/2608.11015-sona/README.md", "topic": [ "generative-recommendation", "semantic-id", "ranking", "history-compression" ], "first_author": null, "first_author_affiliation": "Yandex", "published": "2026-08-11", "code": null, "adapter": { "adapter_key": "sona", "arxiv_id": "2608.11015", "title": "Sona Technical Report", "paper_url": "https://arxiv.org/abs/2608.11015", "track": "recommendation", "organization": "Yandex", "published": "2026-08-11", "code_url": null, "topics": [ "generative-recommendation", "semantic-id", "ranking", "history-compression" ], "local_code_dir": "src/auto_research/reproductions/sona", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "同预算共享 encoder 排序器，实验组为 history compression + SID decoder + item ranker", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "sona.hit_at_10", "sona.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Yandex Music My Vibe smart speakers", "metric": "active users", "lift_percent": 4.53, "traffic": "live-traffic online A/B; cascade replaced by one model", "source_url": "https://arxiv.org/pdf/2608.11015", "source_location": "Online evaluation section", "significance": "reported statistically significant", "retrieved_at": "2026-08-13" }, { "product": "Yandex Music My Vibe smart speakers", "metric": "total listening time", "lift_percent": 6.3, "traffic": "live-traffic online A/B; cascade replaced by one model", "source_url": "https://arxiv.org/pdf/2608.11015", "source_location": "Online evaluation section", "significance": "reported statistically significant", "retrieved_at": "2026-08-13" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "sort-gen", "title": "A Generative Re-ranking Model for List-level Multi-objective Optimization at Taobao", "paper_url": "https://arxiv.org/abs/2505.07197", "detail_path": "reproductions/2505.07197-sort-gen/README.md", "topic": [ "reranking", "multi-objective", "ordered-regression", "diversity" ], "first_author": null, "first_author_affiliation": "Alibaba / Taobao & Tmall", "published": "2025-05-12", "code": null, "adapter": { "adapter_key": "sort-gen", "arxiv_id": "2505.07197", "title": "A Generative Re-ranking Model for List-level Multi-objective Optimization at Taobao", "paper_url": "https://arxiv.org/abs/2505.07197", "track": "recommendation", "organization": "Alibaba / Taobao & Tmall", "published": "2025-05-12", "code_url": null, "topics": [ "reranking", "multi-objective", "ordered-regression", "diversity" ], "local_code_dir": "src/auto_research/reproductions/sort_gen", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 1M exposure-slate proxy" ], "baseline": "item-level greedy formula", "metrics": [ "baseline.click_per_slate", "baseline.gmv_proxy_per_slate", "baseline.ilad", "baseline.model_calls_per_slate", "baseline.pay_per_slate", "method.click_per_slate", "method.gmv_proxy_per_slate", "method.ilad", "method.model_calls_per_slate", "method.pay_per_slate", "paper_results.latency_ms", "paper_results.vs_deployed_FFT_fastDPP_GMV_percent", "paper_results.vs_formula_GMV_percent", "relative.click_per_slate_percent", "relative.gmv_proxy_per_slate_percent", "relative.ilad_percent", "relative.pay_per_slate_percent", "training.final_loss", "training.initial_loss" ], "default_seeds": [ 42 ], "budget": "steps=180", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao Baiyibutie", "metric": "CLICK", "lift_percent": 4.13, "traffic": "two-week A/B vs deployed FFT context model + fastDPP", "source_url": "https://arxiv.org/abs/2505.07197", "source_location": "original paper online-result disclosure: product=Taobao Baiyibutie; metric=CLICK" }, { "product": "Taobao Baiyibutie", "metric": "GMV", "lift_percent": 8.1, "traffic": "two-week A/B vs deployed FFT context model + fastDPP", "source_url": "https://arxiv.org/abs/2505.07197", "source_location": "original paper online-result disclosure: product=Taobao Baiyibutie; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "sort-ranking", "title": "SORT: A Systematically Optimized Ranking Transformer for Industrial-scale Recommenders", "paper_url": "https://arxiv.org/abs/2603.03988", "detail_path": "reproductions/2603.03988-sort-ranking/README.md", "topic": [ "ranking", "transformer", "serving-efficiency" ], "first_author": null, "first_author_affiliation": "Alibaba International Digital Commerce", "published": "2026-03-04", "code": null, "adapter": { "adapter_key": "sort-ranking", "arxiv_id": "2603.03988", "title": "SORT: A Systematically Optimized Ranking Transformer for Industrial-scale Recommenders", "paper_url": "https://arxiv.org/abs/2603.03988", "track": "recommendation", "organization": "Alibaba International Digital Commerce", "published": "2026-03-04", "code_url": null, "topics": [ "ranking", "transformer", "serving-efficiency" ], "local_code_dir": "src/auto_research/reproductions/sort_ranking", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "AliExpress recommendation", "metric": "orders", "lift_percent": 6.35, "traffic": "three production scenarios", "source_url": "https://arxiv.org/html/2603.03988v1", "source_location": "Section 4.7 / Table 5", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "sparsectr", "title": "Unleashing the Potential of Sparse Attention on Long-term Behaviors for CTR Prediction", "paper_url": "https://arxiv.org/abs/2601.17836", "detail_path": "reproductions/2601.17836-sparsectr/README.md", "topic": [ "ctr-prediction", "long-sequence", "sparse-attention" ], "first_author": null, "first_author_affiliation": "Institute of Software, Chinese Academy of Sciences", "published": "2026-01-25", "code": null, "adapter": { "adapter_key": "sparsectr", "arxiv_id": "2601.17836", "title": "Unleashing the Potential of Sparse Attention on Long-term Behaviors for CTR Prediction", "paper_url": "https://arxiv.org/abs/2601.17836", "track": "recommendation", "organization": "Institute of Software, Chinese Academy of Sciences", "published": "2026-01-25", "code_url": null, "topics": [ "ctr-prediction", "long-sequence", "sparse-attention" ], "local_code_dir": "src/auto_research/reproductions/sparsectr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meituan list advertising", "metric": "CTR", "lift_percent": 1.72, "traffic": "1% traffic production online A/B for 7 days", "source_url": "https://arxiv.org/html/2601.17836v1", "source_location": "Section 4.7", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "context:evolutionary-sparse-attention" ] } }, { "domain": "recommendation", "key": "spear", "title": "SPEAR: Selection-aware Personalized End-to-end Adaptive Rewriting and Retrieval for Community Search", "paper_url": "https://arxiv.org/abs/2608.01738", "detail_path": "reproductions/2608.01738-spear/README.md", "topic": [ "search", "query-rewriting", "retrieval", "personalization" ], "first_author": null, "first_author_affiliation": "Dewu", "published": "2026-08-03", "code": "https://github.com/mallocagi1-cell/spear", "adapter": { "adapter_key": "spear", "arxiv_id": "2608.01738", "title": "SPEAR: Selection-aware Personalized End-to-end Adaptive Rewriting and Retrieval for Community Search", "paper_url": "https://arxiv.org/abs/2608.01738", "track": "recommendation", "organization": "Dewu", "published": "2026-08-03", "code_url": "https://github.com/mallocagi1-cell/spear", "topics": [ "search", "query-rewriting", "retrieval", "personalization" ], "local_code_dir": "src/auto_research/reproductions/spear", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享 transition + content scorer，实验组只加入 SPEAR 核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "paper_results.click_recall_at_10_percent", "paper_results.query_view_ctr_percent", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Dewu community search", "metric": "Query-view CTR", "lift_percent": 0.259, "traffic": "online A/B; fully deployed", "source_url": "https://arxiv.org/abs/2608.01738", "source_location": "original paper online-result disclosure: product=Dewu community search; metric=Query-view CTR" }, { "product": "Dewu community search", "metric": "Average reading depth", "lift_percent": 0.733, "traffic": "online A/B; fully deployed", "source_url": "https://arxiv.org/abs/2608.01738", "source_location": "original paper online-result disclosure: product=Dewu community search; metric=Average reading depth" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "specformer", "title": "SpecFormer: Mitigating Embedding and Attention Collapse via Spectral-Aware Transformer for Recommendation", "paper_url": "https://arxiv.org/abs/2607.24025", "detail_path": "reproductions/2607.24025-specformer/README.md", "topic": [ "ranking", "transformer", "spectral-softening" ], "first_author": null, "first_author_affiliation": "Zhejiang University", "published": "2026-07-27", "code": null, "adapter": { "adapter_key": "specformer", "arxiv_id": "2607.24025", "title": "SpecFormer: Mitigating Embedding and Attention Collapse via Spectral-Aware Transformer for Recommendation", "paper_url": "https://arxiv.org/abs/2607.24025", "track": "recommendation", "organization": "Zhejiang University", "published": "2026-07-27", "code_url": null, "topics": [ "ranking", "transformer", "spectral-softening" ], "local_code_dir": "src/auto_research/reproductions/specformer", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "standard Transformer/DLRM-style interaction", "metrics": [ "hit_at_10", "ndcg_at_10", "spectral_condition_before", "softened_rank" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Alibaba e-commerce advertising", "metric": "CTR", "lift_percent": 1.34, "traffic": "10-day test; 10% production traffic", "source_url": "https://arxiv.org/html/2607.24025v1", "source_location": "Section V-F, Table V", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "context:spectral-soften" ] } }, { "domain": "recommendation", "key": "ssrlive", "title": "SSRLive: Live Streaming Recommendation with Dynamic Semantic ID", "paper_url": "https://arxiv.org/abs/2606.06970", "detail_path": "reproductions/2606.06970-ssrlive/README.md", "topic": [ "live-streaming", "semantic-id", "multitask-ranking" ], "first_author": null, "first_author_affiliation": "Taobao & Tmall Group, Alibaba", "published": "2026-06-05", "code": null, "adapter": { "adapter_key": "ssrlive", "arxiv_id": "2606.06970", "title": "SSRLive: Live Streaming Recommendation with Dynamic Semantic ID", "paper_url": "https://arxiv.org/abs/2606.06970", "track": "recommendation", "organization": "Taobao & Tmall Group, Alibaba", "published": "2026-06-05", "code_url": null, "topics": [ "live-streaming", "semantic-id", "multitask-ranking" ], "local_code_dir": "src/auto_research/reproductions/ssrlive", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Alibaba live streaming", "metric": "Watch time", "lift_percent": 3.38, "traffic": "production A/B; statistically significant", "source_url": "https://arxiv.org/html/2606.06970v1", "source_location": "Section 6, Table 4", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "head:dynamic-sid" ] } }, { "domain": "recommendation", "key": "steps", "title": "A Self-Triggered Agentic Push Recommendation System", "paper_url": "https://arxiv.org/abs/2608.01949", "detail_path": "reproductions/2608.01949-steps/README.md", "topic": [ "push-recommendation", "agentic-recommendation", "decision-transformer", "serving" ], "first_author": null, "first_author_affiliation": "ByteDance / Douyin", "published": "2026-08-03", "code": null, "adapter": { "adapter_key": "steps", "arxiv_id": "2608.01949", "title": "A Self-Triggered Agentic Push Recommendation System", "paper_url": "https://arxiv.org/abs/2608.01949", "track": "recommendation", "organization": "ByteDance / Douyin", "published": "2026-08-03", "code_url": null, "topics": [ "push-recommendation", "agentic-recommendation", "decision-transformer", "serving" ], "local_code_dir": "src/auto_research/reproductions/steps", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "共享 transition + content scorer，实验组只加入 STEPS 核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.execution_agent_trajectory_reward", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Douyin Push", "metric": "Active days", "lift_percent": 0.2843, "traffic": "fully deployed; online A/B", "source_url": "https://arxiv.org/abs/2608.01949", "source_location": "original paper online-result disclosure: product=Douyin Push; metric=Active days" }, { "product": "Douyin Push", "metric": "Push permission disablement", "lift_percent": -1.9089, "traffic": "fully deployed; online A/B", "source_url": "https://arxiv.org/abs/2608.01949", "source_location": "original paper online-result disclosure: product=Douyin Push; metric=Push permission disablement" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "swag-bid", "title": "Beyond Single-Episode Optimization: Sliding-Window Aware Generative Auto-Bidding for Long-Term Advertising Effectiveness", "paper_url": "https://arxiv.org/abs/2607.25233", "detail_path": "reproductions/2607.25233-swag-bid/README.md", "topic": [ "advertising", "generative-model", "decision-transformer", "long-horizon" ], "first_author": null, "first_author_affiliation": "Alibaba International Digital Commerce / Dalian University of Technology", "published": "2026-07-28", "code": null, "adapter": { "adapter_key": "swag-bid", "arxiv_id": "2607.25233", "title": "Beyond Single-Episode Optimization: Sliding-Window Aware Generative Auto-Bidding for Long-Term Advertising Effectiveness", "paper_url": "https://arxiv.org/abs/2607.25233", "track": "recommendation", "organization": "Alibaba International Digital Commerce / Dalian University of Technology", "published": "2026-07-28", "code_url": null, "topics": [ "advertising", "generative-model", "decision-transformer", "long-horizon" ], "local_code_dir": "src/auto_research/reproductions/swag_bid", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "single-episode Decision Transformer proxy，实验组为 sliding-window planner", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "paper_results.cost_percent", "paper_results.gmv_percent", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.validation_only_model_selection.fresh_hit_at_10", "stages.validation_only_model_selection.hit_at_10", "stages.validation_only_model_selection.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "steps=True", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "AliExpress advertising", "metric": "GMV", "lift_percent": 3.42, "traffic": "21-day campaign-randomized online A/B", "source_url": "https://arxiv.org/abs/2607.25233", "source_location": "original paper online-result disclosure: product=AliExpress advertising; metric=GMV" }, { "product": "AliExpress advertising", "metric": "ROAS", "lift_percent": 5.65, "traffic": "21-day campaign-randomized online A/B", "source_url": "https://arxiv.org/abs/2607.25233", "source_location": "original paper online-result disclosure: product=AliExpress advertising; metric=ROAS" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "tagllm", "title": "TagLLM: A Fine-Grained Tag Generation Approach for Note Recommendation", "paper_url": "https://arxiv.org/abs/2603.21481", "detail_path": "reproductions/2603.21481-tagllm/README.md", "topic": [ "content-understanding", "tag-generation", "knowledge-distillation" ], "first_author": null, "first_author_affiliation": "Tongji University", "published": "2026-03-23", "code": null, "adapter": { "adapter_key": "tagllm", "arxiv_id": "2603.21481", "title": "TagLLM: A Fine-Grained Tag Generation Approach for Note Recommendation", "paper_url": "https://arxiv.org/abs/2603.21481", "track": "recommendation", "organization": "Tongji University", "published": "2026-03-23", "code_url": null, "topics": [ "content-understanding", "tag-generation", "knowledge-distillation" ], "local_code_dir": "src/auto_research/reproductions/tagllm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Note recommendation", "metric": "Cold-start page-view CTR", "lift_percent": 32.37, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2603.21481v1", "source_location": "Abstract", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "head:fine-grained-tags" ] } }, { "domain": "recommendation", "key": "tagr", "title": "TAGR: Temporally Adaptive Generative Recommendation for Industrial Live-Streaming Advertising", "paper_url": "https://arxiv.org/abs/2608.24034", "detail_path": "reproductions/2608.24034-tagr/README.md", "topic": [ "generative-recommendation", "live-streaming-advertising", "semantic-id", "online-rl" ], "first_author": "Wencai Ye", "first_author_affiliation": "Kuaishou Technology", "published": "2026-08-25", "code": null, "adapter": { "adapter_key": "tagr", "arxiv_id": "2608.24034", "title": "TAGR: Temporally Adaptive Generative Recommendation for Industrial Live-Streaming Advertising", "paper_url": "https://arxiv.org/abs/2608.24034", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-08-25", "code_url": null, "topics": [ "generative-recommendation", "live-streaming-advertising", "semantic-id", "online-rl" ], "local_code_dir": "src/auto_research/reproductions/tagr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "production-style first-order transition recommender", "metrics": [ "Hit@10", "NDCG@10", "head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou live-stream advertising", "metric": "live-room entry rate", "lift_percent": 8.5, "traffic": "production A/B under the same downstream stack", "source_url": "https://arxiv.org/html/2608.24034v1#S4.SS3", "source_location": "Section 4.3 / Table 1", "retrieved_at": "2026-08-26" }, { "product": "Kuaishou live-stream advertising", "metric": "shopping-cart click rate", "lift_percent": 7.4, "traffic": "production A/B under the same downstream stack", "source_url": "https://arxiv.org/html/2608.24034v1#S4.SS3", "source_location": "Section 4.3 / Table 1", "retrieved_at": "2026-08-26" }, { "product": "Kuaishou live-stream advertising", "metric": "revenue", "lift_percent": 16.1, "traffic": "multi-week production experiment", "source_url": "https://arxiv.org/html/2608.24034v1#S4.SS3", "source_location": "Section 4.3 / Table 1", "retrieved_at": "2026-08-26" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "taiji", "title": "Taiji: Pareto Optimal Policy Optimization with Semantics-IDs Trade-off for Industrial LLM-Enhanced Recommendation", "paper_url": "https://arxiv.org/abs/2606.03866", "detail_path": "reproductions/2606.03866-taiji/README.md", "topic": [ "advertising", "reinforcement-alignment", "pareto-optimization" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-06-02", "code": null, "adapter": { "adapter_key": "taiji", "arxiv_id": "2606.03866", "title": "Taiji: Pareto Optimal Policy Optimization with Semantics-IDs Trade-off for Industrial LLM-Enhanced Recommendation", "paper_url": "https://arxiv.org/abs/2606.03866", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-06-02", "code_url": null, "topics": [ "advertising", "reinforcement-alignment", "pareto-optimization" ], "local_code_dir": "src/auto_research/reproductions/taiji", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou advertising", "metric": "Revenue", "lift_percent": 3.3, "traffic": "7 days; 10% control and 10% treatment", "source_url": "https://arxiv.org/html/2606.03866v1", "source_location": "Section 3.4, Table 2", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "reward:pareto-semantic-id" ] } }, { "domain": "recommendation", "key": "tgr", "title": "TGR: Advancing Industrial Recommendation from Generative-Paradigm Ranking toward Unified Generation and Reasoning", "paper_url": "https://arxiv.org/abs/2609.00986", "detail_path": "reproductions/2609.00986-tgr/README.md", "topic": [ "generative-recommendation", "ranking-network", "reason-token" ], "first_author": null, "first_author_affiliation": "Tencent", "published": "2026-09-01", "code": null, "adapter": { "adapter_key": "tgr", "arxiv_id": "2609.00986", "title": "TGR: Advancing Industrial Recommendation from Generative-Paradigm Ranking toward Unified Generation and Reasoning", "paper_url": "https://arxiv.org/abs/2609.00986", "track": "recommendation", "organization": "Tencent", "published": "2026-09-01", "code_url": null, "topics": [ "generative-recommendation", "ranking-network", "reason-token" ], "local_code_dir": "src/auto_research/reproductions/tgr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Tencent production surfaces", "metric": "advertising revenue", "lift_percent": 1.71, "traffic": "five A/B-tested scenarios; full launches", "source_url": "https://arxiv.org/html/2609.00986v1", "source_location": "Abstract and online results", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "head:tgr-generation-reasoning" ] } }, { "domain": "recommendation", "key": "tiger", "title": "Recommender Systems with Generative Retrieval", "paper_url": "https://arxiv.org/abs/2305.05065", "detail_path": "reproductions/2305.05065-tiger/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "tiger", "arxiv_id": "2305.05065", "title": "Recommender Systems with Generative Retrieval", "paper_url": "https://arxiv.org/abs/2305.05065", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/tiger", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K genre content" ], "baseline": "词表容量和参数量匹配的随机分层 ID 生成器", "metrics": [ "ndcg_gain_percent", "random_id.hit_at_10", "random_id.hit_at_10_std", "random_id.ndcg_at_10", "random_id.ndcg_at_10_std", "rqvae.final_loss", "rqvae.initial_loss", "tiger.hit_at_10", "tiger.hit_at_10_std", "tiger.ndcg_at_10", "tiger.ndcg_at_10_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=240", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "tm20k", "title": "Teacher Retains Full Tokens, Student Merges Efficiently: TM20K for E-Commerce Sequence Modeling in Ad Recommendation", "paper_url": "https://arxiv.org/abs/2608.07055", "detail_path": "reproductions/2608.07055-tm20k/README.md", "topic": [ "advertising", "long-sequence", "token-merging", "distillation" ], "first_author": null, "first_author_affiliation": "ByteDance", "published": "2026-08-07", "code": null, "adapter": { "adapter_key": "tm20k", "arxiv_id": "2608.07055", "title": "Teacher Retains Full Tokens, Student Merges Efficiently: TM20K for E-Commerce Sequence Modeling in Ad Recommendation", "paper_url": "https://arxiv.org/abs/2608.07055", "track": "recommendation", "organization": "ByteDance", "published": "2026-08-07", "code_url": null, "topics": [ "advertising", "long-sequence", "token-merging", "distillation" ], "local_code_dir": "src/auto_research/reproductions/tm20k", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (full public histories)" ], "baseline": "same merged-token student without teacher distillation", "metrics": [ "hit_at_10", "ndcg_at_10", "compression_ratio", "distilled_teacher_mse" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; at most 8 merged tokens", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "ByteDance e-commerce advertising", "metric": "ADSS", "lift_percent": 1.036, "traffic": "5-day experiment serving hundreds of millions of users", "source_url": "https://arxiv.org/html/2608.07055v1", "source_location": "Section 5.4, Table 7", "significance": "reported p-value 0% after rounding", "retrieved_at": "2026-09-01" } ], "selection_exception": null, "evolve_operators": [ "context:tm20k-merge" ] } }, { "domain": "recommendation", "key": "tmallgs", "title": "TMallGS: Scaling Unified Feature and Sequence Modeling for Generative E-commerce Search", "paper_url": "https://arxiv.org/abs/2607.13398", "detail_path": "reproductions/2607.13398-tmallgs/README.md", "topic": [ "search-ranking", "transformer", "scaling", "feature-interaction" ], "first_author": null, "first_author_affiliation": "Taobao & Tmall Group of Alibaba", "published": "2026-07-15", "code": null, "adapter": { "adapter_key": "tmallgs", "arxiv_id": "2607.13398", "title": "TMallGS: Scaling Unified Feature and Sequence Modeling for Generative E-commerce Search", "paper_url": "https://arxiv.org/abs/2607.13398", "track": "recommendation", "organization": "Taobao & Tmall Group of Alibaba", "published": "2026-07-15", "code_url": null, "topics": [ "search-ranking", "transformer", "scaling", "feature-interaction" ], "local_code_dir": "src/auto_research/reproductions/tmallgs", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "DIN/RankMixer 风格 late-fusion Transformer，实验组为 TMallGS field-adaptive gated Transformer", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "steps=60", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Tmall Search", "metric": "UCTCVR", "lift_percent": 1.38, "traffic": "30 days; p<0.05", "source_url": "https://arxiv.org/abs/2607.13398", "source_location": "original paper online-result disclosure: product=Tmall Search; metric=UCTCVR" }, { "product": "Tmall Search", "metric": "GMV", "lift_percent": 1.52, "traffic": "30 days; p<0.05", "source_url": "https://arxiv.org/abs/2607.13398", "source_location": "original paper online-result disclosure: product=Tmall Search; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "tokenminds", "title": "TokenMinds: Pretrained User Tokens and Embeddings for User Understanding in Large Recommender Systems", "paper_url": "https://arxiv.org/abs/2606.25147", "detail_path": "reproductions/2606.25147-tokenminds/README.md", "topic": [ "llm-recommendation", "user-modeling", "semantic-id", "cross-scenario", "ranking", "serving" ], "first_author": null, "first_author_affiliation": "Google DeepMind / YouTube", "published": "2026-06-23", "code": null, "adapter": { "adapter_key": "tokenminds", "arxiv_id": "2606.25147", "title": "TokenMinds: Pretrained User Tokens and Embeddings for User Understanding in Large Recommender Systems", "paper_url": "https://arxiv.org/abs/2606.25147", "track": "recommendation", "organization": "Google DeepMind / YouTube", "published": "2026-06-23", "code_url": null, "topics": [ "llm-recommendation", "user-modeling", "semantic-id", "cross-scenario", "ranking", "serving" ], "local_code_dir": "src/auto_research/reproductions/tokenminds", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "同数据、同 seed、同 120 steps 的 dense-only GRU 用户编码器", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "training.baseline.final_ranking_loss", "training.baseline.initial_ranking_loss", "training.tokenminds.final_ranking_loss", "training.tokenminds.final_sid_loss", "training.tokenminds.initial_ranking_loss" ], "default_seeds": [ 42 ], "budget": "steps=120", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "YouTube SFV", "metric": "Engaged Users", "lift_percent": 0.11, "traffic": "seven-day production A/B; subsequently full user traffic", "source_url": "https://arxiv.org/pdf/2606.25147", "source_location": "Table 4 and Section 4.3", "experiment_duration": "7 days", "significance": "95% confidence", "retrieved_at": "2026-08-09" }, { "product": "YouTube SFV", "metric": "Satisfied Engagement", "lift_percent": 0.62, "traffic": "seven-day production A/B; subsequently full user traffic", "source_url": "https://arxiv.org/pdf/2606.25147", "source_location": "Table 4 and Section 4.3", "experiment_duration": "7 days", "significance": "95% confidence", "retrieved_at": "2026-08-09" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "tokenmixer-large", "title": "TokenMixer-Large: Scaling Up Large Ranking Models in Industrial Recommenders", "paper_url": "https://arxiv.org/abs/2602.06563", "detail_path": "reproductions/2602.06563-tokenmixer-large/README.md", "topic": [ "ranking", "token-mixing", "scaling" ], "first_author": null, "first_author_affiliation": "ByteDance", "published": "2026-02-06", "code": null, "adapter": { "adapter_key": "tokenmixer-large", "arxiv_id": "2602.06563", "title": "TokenMixer-Large: Scaling Up Large Ranking Models in Industrial Recommenders", "paper_url": "https://arxiv.org/abs/2602.06563", "track": "recommendation", "organization": "ByteDance", "published": "2026-02-06", "code_url": null, "topics": [ "ranking", "token-mixing", "scaling" ], "local_code_dir": "src/auto_research/reproductions/tokenmixer_large", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MovieLens 100K (932 users / 1682 items)" ], "baseline": "同协议 RankMixer dense，实验组为 TokenMixer-Large", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "baseline.parameters", "method.hit_at_10", "method.ndcg_at_10", "method.parameters", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "steps=80", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "ByteDance e-commerce", "metric": "Orders", "lift_percent": 1.66, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2602.06563", "source_location": "original paper online-result disclosure: product=ByteDance e-commerce; metric=Orders" }, { "product": "ByteDance e-commerce", "metric": "Per-capita payment GMV", "lift_percent": 2.98, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2602.06563", "source_location": "original paper online-result disclosure: product=ByteDance e-commerce; metric=Per-capita payment GMV" }, { "product": "ByteDance ads", "metric": "ADSS", "lift_percent": 2.0, "traffic": "online A/B", "source_url": "https://arxiv.org/abs/2602.06563", "source_location": "original paper online-result disclosure: product=ByteDance ads; metric=ADSS" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "toolrec", "title": "ToolRec: Calibrated Preference Alignment for Query Recommendation in On-Device Assistants", "paper_url": "https://arxiv.org/abs/2606.08466", "detail_path": "reproductions/2606.08466-toolrec/README.md", "topic": [ "query-recommendation", "preference-alignment", "tool-use" ], "first_author": null, "first_author_affiliation": "Huazhong University of Science and Technology", "published": "2026-06-07", "code": null, "adapter": { "adapter_key": "toolrec", "arxiv_id": "2606.08466", "title": "ToolRec: Calibrated Preference Alignment for Query Recommendation in On-Device Assistants", "paper_url": "https://arxiv.org/abs/2606.08466", "track": "recommendation", "organization": "Huazhong University of Science and Technology", "published": "2026-06-07", "code_url": null, "topics": [ "query-recommendation", "preference-alignment", "tool-use" ], "local_code_dir": "src/auto_research/reproductions/toolrec", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "OPPO Xiaobu", "metric": "CTR", "lift_percent": 3.32, "traffic": "7 days; 5% traffic per model", "source_url": "https://arxiv.org/html/2606.08466v1", "source_location": "Section 5.2, Table 1", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "reward:tool-calibration" ] } }, { "domain": "recommendation", "key": "transact-v2", "title": "TransAct V2: Lifelong User Action Sequence Modeling on Pinterest Recommendation", "paper_url": "https://arxiv.org/abs/2506.02267", "detail_path": "reproductions/2506.02267-transact-v2/README.md", "topic": [], "first_author": null, "first_author_affiliation": null, "published": null, "code": null, "adapter": { "adapter_key": "transact-v2", "arxiv_id": "2506.02267", "title": "TransAct V2: Lifelong User Action Sequence Modeling on Pinterest Recommendation", "paper_url": "https://arxiv.org/abs/2506.02267", "track": "recommendation", "organization": null, "published": null, "code_url": null, "topics": [], "local_code_dir": "src/auto_research/reproductions/transact_v2", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "参数量匹配的 TransAct", "metrics": [ "ndcg_gain_percent", "transact.hit_at_10", "transact.hit_at_10_std", "transact.ndcg_at_10", "transact.ndcg_at_10_std", "transact_v2.hit_at_10", "transact_v2.hit_at_10_std", "transact_v2.ndcg_at_10", "transact_v2.ndcg_at_10_std" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=160", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "transretrieval", "title": "TransRetrieval: Scaling Up Transformer-Based Retrieval for Industrial Recommendation", "paper_url": "https://arxiv.org/abs/2608.25528", "detail_path": "reproductions/2608.25528-transretrieval/README.md", "topic": [ "retrieval", "transformer-scaling", "target-token-compression", "multi-domain" ], "first_author": null, "first_author_affiliation": "Renmin University of China / Taobao & Tmall Group, Alibaba", "published": "2026-08-26", "code": null, "adapter": { "adapter_key": "transretrieval", "arxiv_id": "2608.25528", "title": "TransRetrieval: Scaling Up Transformer-Based Retrieval for Industrial Recommendation", "paper_url": "https://arxiv.org/abs/2608.25528", "track": "recommendation", "organization": "Renmin University of China / Taobao & Tmall Group, Alibaba", "published": "2026-08-26", "code_url": null, "topics": [ "retrieval", "transformer-scaling", "target-token-compression", "multi-domain" ], "local_code_dir": "src/auto_research/reproductions/transretrieval", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "same-feature Transformer-style two-tower mean aggregation", "metrics": [ "Hit@10", "NDCG@10", "head share@10", "token norm dispersion" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Alibaba display advertising", "metric": "platform revenue", "lift_percent": 2.53, "traffic": "5% traffic; month-long A/B; p<0.0001", "source_url": "https://arxiv.org/html/2608.25528v1#S4.SS5", "source_location": "Section 4.5 / Table 7", "experiment_duration": "one month", "significance": "p<0.0001", "retrieved_at": "2026-08-28" }, { "product": "Alibaba display advertising", "metric": "RPM", "lift_percent": 1.28, "traffic": "5% traffic; month-long A/B; p<0.0001", "source_url": "https://arxiv.org/html/2608.25528v1#S4.SS5", "source_location": "Section 4.5 / Table 7", "experiment_duration": "one month", "significance": "p<0.0001", "retrieved_at": "2026-08-28" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "transx", "title": "TransX: Scaling Transformer-based Recommendation via Behavioral and Serving Stream Crossings", "paper_url": "https://arxiv.org/abs/2607.28940", "detail_path": "reproductions/2607.28940-transx/README.md", "topic": [ "ranking", "long-sequence", "cross-attention", "serving" ], "first_author": null, "first_author_affiliation": "LinkedIn", "published": "2026-07-31", "code": null, "adapter": { "adapter_key": "transx", "arxiv_id": "2607.28940", "title": "TransX: Scaling Transformer-based Recommendation via Behavioral and Serving Stream Crossings", "paper_url": "https://arxiv.org/abs/2607.28940", "track": "recommendation", "organization": "LinkedIn", "published": "2026-07-31", "code_url": null, "topics": [ "ranking", "long-sequence", "cross-attention", "serving" ], "local_code_dir": "src/auto_research/reproductions/transx", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "monolithic recent-sequence scorer", "metrics": [ "hit_at_10", "ndcg_at_10", "attention_pair_reduction" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; global token plus 8 cached local tokens", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "LinkedIn recommender systems", "metric": "CTR", "lift_percent": 6.0, "traffic": "large-scale online A/B test", "source_url": "https://arxiv.org/html/2607.28940v1", "source_location": "Abstract and Section 6", "significance": "statistically significant for users with eligible history", "retrieved_at": "2026-09-01" } ], "selection_exception": null, "evolve_operators": [ "context:transx-cross-stream" ] } }, { "domain": "recommendation", "key": "tsgr", "title": "TSGR: Taobao Search Generative Retrieval", "paper_url": "https://arxiv.org/abs/2607.18796", "detail_path": "reproductions/2607.18796-tsgr/README.md", "topic": [ "generative-retrieval", "semantic-id", "search", "pre-ranking" ], "first_author": null, "first_author_affiliation": "Taobao & Tmall Group of Alibaba / Zhejiang University", "published": "2026-07-21", "code": null, "adapter": { "adapter_key": "tsgr", "arxiv_id": "2607.18796", "title": "TSGR: Taobao Search Generative Retrieval", "paper_url": "https://arxiv.org/abs/2607.18796", "track": "recommendation", "organization": "Taobao & Tmall Group of Alibaba / Zhejiang University", "published": "2026-07-21", "code_url": null, "topics": [ "generative-retrieval", "semantic-id", "search", "pre-ranking" ], "local_code_dir": "src/auto_research/reproductions/tsgr", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "语义召回加传统预排代理，实验组为 QP-SID 加联合 VRM", "metrics": [ "baseline.fresh_hit_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao Search", "metric": "IPV", "lift_percent": 0.43, "traffic": "1% traffic; 38 days; statistically significant", "source_url": "https://arxiv.org/abs/2607.18796", "source_location": "original paper online-result disclosure: product=Taobao Search; metric=IPV" }, { "product": "Taobao Search", "metric": "Transaction Count", "lift_percent": 1.12, "traffic": "1% traffic; 38 days; statistically significant", "source_url": "https://arxiv.org/abs/2607.18796", "source_location": "original paper online-result disclosure: product=Taobao Search; metric=Transaction Count" }, { "product": "Taobao Search", "metric": "GMV", "lift_percent": 1.64, "traffic": "1% traffic; 38 days; statistically significant", "source_url": "https://arxiv.org/abs/2607.18796", "source_location": "original paper online-result disclosure: product=Taobao Search; metric=GMV" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "tubifm", "title": "TubiFM: Unified Item, Carousel, and Search Ranking for Streaming Discovery", "paper_url": "https://arxiv.org/abs/2605.23702", "detail_path": "reproductions/2605.23702-tubifm/README.md", "topic": [ "foundation-model", "multi-task-learning", "search-ranking" ], "first_author": null, "first_author_affiliation": "Tubi", "published": "2026-05-22", "code": null, "adapter": { "adapter_key": "tubifm", "arxiv_id": "2605.23702", "title": "TubiFM: Unified Item, Carousel, and Search Ranking for Streaming Discovery", "paper_url": "https://arxiv.org/abs/2605.23702", "track": "recommendation", "organization": "Tubi", "published": "2026-05-22", "code_url": null, "topics": [ "foundation-model", "multi-task-learning", "search-ranking" ], "local_code_dir": "src/auto_research/reproductions/tubifm", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Tubi Search", "metric": "total viewing time", "lift_percent": 3.9, "traffic": "production A/B; p<0.05", "source_url": "https://arxiv.org/html/2605.23702v1", "source_location": "paper online evaluation section", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "twice", "title": "TWICE: Two-Clock, Two-Window Learning for Long-Horizon Conversion Prediction in Online Advertising", "paper_url": "https://arxiv.org/abs/2607.25404", "detail_path": "reproductions/2607.25404-twice/README.md", "topic": [ "advertising", "cvr", "delayed-feedback", "full-traffic" ], "first_author": null, "first_author_affiliation": "Kuaishou", "published": "2026-07-28", "code": null, "adapter": { "adapter_key": "twice", "arxiv_id": "2607.25404", "title": "TWICE: Two-Clock, Two-Window Learning for Long-Horizon Conversion Prediction in Online Advertising", "paper_url": "https://arxiv.org/abs/2607.25404", "track": "recommendation", "organization": "Kuaishou", "published": "2026-07-28", "code_url": null, "topics": [ "advertising", "cvr", "delayed-feedback", "full-traffic" ], "local_code_dir": "src/auto_research/reproductions/twice", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "mature-label next-item CVR，实验组为双时钟 current-status", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "paper_results.expected_revenue_percent", "paper_results.revenue_percent", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.validation_only_model_selection.fresh_hit_at_10", "stages.validation_only_model_selection.hit_at_10", "stages.validation_only_model_selection.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kwai advertising", "metric": "expected revenue", "lift_percent": 2.486, "traffic": "online A/B; later full traffic", "source_url": "https://arxiv.org/abs/2607.25404", "source_location": "original paper online-result disclosure: product=Kwai advertising; metric=expected revenue" }, { "product": "Kwai advertising", "metric": "conversions", "lift_percent": 2.061, "traffic": "online A/B; later full traffic", "source_url": "https://arxiv.org/abs/2607.25404", "source_location": "original paper online-result disclosure: product=Kwai advertising; metric=conversions" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "twin-v2", "title": "TWIN V2: Scaling Ultra-Long User Behavior Sequence Modeling for Enhanced CTR Prediction at Kuaishou", "paper_url": "https://arxiv.org/abs/2407.16357", "detail_path": "reproductions/2407.16357-twin-v2/README.md", "topic": [ "ranking", "long-sequence", "training-serving" ], "first_author": null, "first_author_affiliation": "Kuaishou", "published": "2024-07-23", "code": null, "adapter": { "adapter_key": "twin-v2", "arxiv_id": "2407.16357", "title": "TWIN V2: Scaling Ultra-Long User Behavior Sequence Modeling for Enhanced CTR Prediction at Kuaishou", "paper_url": "https://arxiv.org/abs/2407.16357", "track": "recommendation", "organization": "Kuaishou", "published": "2024-07-23", "code_url": null, "topics": [ "ranking", "long-sequence", "training-serving" ], "local_code_dir": "src/auto_research/reproductions/twin_v2", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "shared transition + content baseline，实验组为 TWIN-V2，只改变论文核心机制", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.validation.fresh_hit_at_10", "stages.validation.hit_at_10", "stages.validation.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou Featured-Video", "metric": "watch time", "lift_percent": 0.672, "traffic": "main traffic", "source_url": "https://arxiv.org/abs/2407.16357", "source_location": "original paper online-result disclosure: product=Kuaishou Featured-Video; metric=watch time" }, { "product": "Kuaishou Discovery", "metric": "watch time", "lift_percent": 0.8, "traffic": "main traffic", "source_url": "https://arxiv.org/abs/2407.16357", "source_location": "original paper online-result disclosure: product=Kuaishou Discovery; metric=watch time" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "twitch-mor", "title": "Multi-Objective Ranking for Live-Streaming: Balancing Fresh and Delayed Signals with Segment-Aware Targeting", "paper_url": "https://arxiv.org/abs/2608.04455", "detail_path": "reproductions/2608.04455-twitch-mor/README.md", "topic": [ "ranking", "multi-task-learning", "mmoe", "live-streaming" ], "first_author": null, "first_author_affiliation": "Twitch", "published": "2026-08-05", "code": null, "adapter": { "adapter_key": "twitch-mor", "arxiv_id": "2608.04455", "title": "Multi-Objective Ranking for Live-Streaming: Balancing Fresh and Delayed Signals with Segment-Aware Targeting", "paper_url": "https://arxiv.org/abs/2608.04455", "track": "recommendation", "organization": "Twitch", "published": "2026-08-05", "code_url": null, "topics": [ "ranking", "multi-task-learning", "mmoe", "live-streaming" ], "local_code_dir": "src/auto_research/reproductions/twitch_mor", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "去掉论文特有机制、其余数据切分与预算相同的 matched control", "metrics": [ "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "relative.objective_mse_percent", "variants.fresh-delayed lifecycle mmoe.hit_at_10", "variants.fresh-delayed lifecycle mmoe.ndcg_at_10", "variants.single-objective dnn.hit_at_10", "variants.single-objective dnn.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Twitch live recommendation", "metric": "daily active viewers", "lift_percent": 0.09, "traffic": "14-day A/B", "significance": "p<0.01", "source_url": "https://arxiv.org/abs/2608.04455", "source_location": "original paper online-result disclosure: product=Twitch live recommendation; metric=daily active viewers" }, { "product": "Twitch engaged viewers", "metric": "capped ARPU", "lift_percent": 0.56, "traffic": "14-day A/B", "significance": "p<0.05", "source_url": "https://arxiv.org/abs/2608.04455", "source_location": "original paper online-result disclosure: product=Twitch engaged viewers; metric=capped ARPU" }, { "product": "Twitch mobile livefeed", "metric": "positive interactions", "lift_percent": 1.12, "traffic": "14-day A/B", "significance": "p<0.001", "source_url": "https://arxiv.org/abs/2608.04455", "source_location": "original paper online-result disclosure: product=Twitch mobile livefeed; metric=positive interactions" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "uame", "title": "Uncertainty as Remedy: Mitigating Satisfaction Label Bias in Short Video Multi-Objective Ensemble Ranking", "paper_url": "https://arxiv.org/abs/2607.17092", "detail_path": "reproductions/2607.17092-uame/README.md", "topic": [ "multi-objective-ranking", "uncertainty", "loss-design" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-07-19", "code": null, "adapter": { "adapter_key": "uame", "arxiv_id": "2607.17092", "title": "Uncertainty as Remedy: Mitigating Satisfaction Label Bias in Short Video Multi-Objective Ensemble Ranking", "paper_url": "https://arxiv.org/abs/2607.17092", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-07-19", "code_url": null, "topics": [ "multi-objective-ranking", "uncertainty", "loss-design" ], "local_code_dir": "src/auto_research/reproductions/uame", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "同 backbone 的 deterministic multi-objective pairwise ranker", "metrics": [ "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "ndcg_at_10_relative_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou short-video feed vs EMER", "metric": "LongView", "lift_percent": 1.614, "traffic": "7 days; 5% traffic; p<0.005", "source_url": "https://arxiv.org/abs/2607.17092", "source_location": "original paper online-result disclosure: product=Kuaishou short-video feed vs EMER; metric=LongView" }, { "product": "Kuaishou short-video feed vs EASQ", "metric": "Forward", "lift_percent": 1.598, "traffic": "7 days; 5% traffic; p<0.005", "source_url": "https://arxiv.org/abs/2607.17092", "source_location": "original paper online-result disclosure: product=Kuaishou short-video feed vs EASQ; metric=Forward" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "ug-sep", "title": "Compute Only Once: UG-Separation for Efficient Large Recommendation Models", "paper_url": "https://arxiv.org/abs/2602.10455", "detail_path": "reproductions/2602.10455-ug-sep/README.md", "topic": [ "ranking-network", "inference-efficiency", "token-mixing" ], "first_author": null, "first_author_affiliation": "ByteDance AML", "published": "2026-02-11", "code": null, "adapter": { "adapter_key": "ug-sep", "arxiv_id": "2602.10455", "title": "Compute Only Once: UG-Separation for Efficient Large Recommendation Models", "paper_url": "https://arxiv.org/abs/2602.10455", "track": "recommendation", "organization": "ByteDance AML", "published": "2026-02-11", "code_url": null, "topics": [ "ranking-network", "inference-efficiency", "token-mixing" ], "local_code_dir": "src/auto_research/reproductions/ug_sep", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "ByteDance recommendation", "metric": "Inference latency reduction", "lift_percent": 20.0, "traffic": "production online A/B and deployment", "source_url": "https://arxiv.org/html/2602.10455v1", "source_location": "Abstract and online experiments", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "head:ug-separation" ] } }, { "domain": "recommendation", "key": "ultra-hstu", "title": "Bending the Scaling Law Curve in Large-Scale Recommendation Systems", "paper_url": "https://arxiv.org/abs/2602.16986", "detail_path": "reproductions/2602.16986-ultra-hstu/README.md", "topic": [ "ranking", "long-sequence", "scaling-law", "attention", "serving" ], "first_author": null, "first_author_affiliation": "Meta", "published": "2026-02-19", "code": null, "adapter": { "adapter_key": "ultra-hstu", "arxiv_id": "2602.16986", "title": "Bending the Scaling Law Curve in Large-Scale Recommendation Systems", "paper_url": "https://arxiv.org/abs/2602.16986", "track": "recommendation", "organization": "Meta", "published": "2026-02-19", "code_url": null, "topics": [ "ranking", "long-sequence", "scaling-law", "attention", "serving" ], "local_code_dir": "src/auto_research/reproductions/ultra_hstu", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (220 users / 360 items)" ], "baseline": "short-history transition-content ranker，实验组执行 ULTRA-HSTU compact core", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta recommendation", "metric": "consumption", "lift_percent": 4.11, "traffic": "multiple rigorous 30-day A/B tests; fully deployed", "source_url": "https://arxiv.org/html/2602.16986v2#S5.SS4", "source_location": "Section 5.4", "experiment_duration": "30 days", "retrieved_at": "2026-08-09" }, { "product": "Meta recommendation", "metric": "engagement", "lift_percent": 8.0, "traffic": "reported range +2% to +8%; fully deployed", "source_url": "https://arxiv.org/html/2602.16986v2#S5.SS4", "source_location": "Section 5.4", "experiment_duration": "30 days", "retrieved_at": "2026-08-09" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "uniformer", "title": "UniFormer: Efficient and Unified Model-Centric Scaling for Industrial Recommendation", "paper_url": "https://arxiv.org/abs/2606.27058", "detail_path": "reproductions/2606.27058-uniformer/README.md", "topic": [ "ranking", "feature-tokenization", "multi-task" ], "first_author": null, "first_author_affiliation": "Kuaishou", "published": "2026-06-25", "code": null, "adapter": { "adapter_key": "uniformer", "arxiv_id": "2606.27058", "title": "UniFormer: Efficient and Unified Model-Centric Scaling for Industrial Recommendation", "paper_url": "https://arxiv.org/abs/2606.27058", "track": "recommendation", "organization": "Kuaishou", "published": "2026-06-25", "code_url": null, "topics": [ "ranking", "feature-tokenization", "multi-task" ], "local_code_dir": "src/auto_research/reproductions/uniformer", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "single-space feature interaction", "metrics": [ "hit_at_10", "ndcg_at_10", "feature_spaces", "task_tokens" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou Lite", "metric": "Watch Time", "lift_percent": 1.113, "traffic": "seven-day test; 5% production traffic", "source_url": "https://arxiv.org/html/2606.27058v1", "source_location": "Section 5.6, Table 2", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "context:unified-token" ] } }, { "domain": "recommendation", "key": "unimvt", "title": "Jointly Optimizing Debiased CTR and Uplift for Coupons Marketing: A Unified Causal Framework", "paper_url": "https://arxiv.org/abs/2602.12972", "detail_path": "reproductions/2602.12972-unimvt/README.md", "topic": [ "causal-recommendation", "uplift-modeling", "coupon-marketing" ], "first_author": null, "first_author_affiliation": "Kuaishou Technology", "published": "2026-02-13", "code": null, "adapter": { "adapter_key": "unimvt", "arxiv_id": "2602.12972", "title": "Jointly Optimizing Debiased CTR and Uplift for Coupons Marketing: A Unified Causal Framework", "paper_url": "https://arxiv.org/abs/2602.12972", "track": "recommendation", "organization": "Kuaishou Technology", "published": "2026-02-13", "code_url": null, "topics": [ "causal-recommendation", "uplift-modeling", "coupon-marketing" ], "local_code_dir": "src/auto_research/reproductions/unimvt", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou coupon marketing", "metric": "Coupon revenue", "lift_percent": 18.14, "traffic": "10% production traffic online A/B", "source_url": "https://arxiv.org/html/2602.12972v1", "source_location": "Section 5.4", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "reward:causal-uplift" ] } }, { "domain": "recommendation", "key": "unir2", "title": "Unifying Generative Recall and Multi-Objective Ranking in a Single Decoder-Only Sequence", "paper_url": "https://arxiv.org/abs/2607.24439", "detail_path": "reproductions/2607.24439-unir2/README.md", "topic": [ "generative-recommendation", "unified-recall-ranking", "multi-objective-ranking", "lora" ], "first_author": null, "first_author_affiliation": "Kuaishou / IIE, CAS / UCAS", "published": "2026-07-27", "code": null, "adapter": { "adapter_key": "unir2", "arxiv_id": "2607.24439", "title": "Unifying Generative Recall and Multi-Objective Ranking in a Single Decoder-Only Sequence", "paper_url": "https://arxiv.org/abs/2607.24439", "track": "recommendation", "organization": "Kuaishou / IIE, CAS / UCAS", "published": "2026-07-27", "code_url": null, "topics": [ "generative-recommendation", "unified-recall-ranking", "multi-objective-ranking", "lora" ], "local_code_dir": "src/auto_research/reproductions/unir2", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-1M" ], "baseline": "同一 MovieLens-1M 数据、SID、训练 step 和 full-catalog 候选上的独立生成 recall + ranking cascade", "metrics": [ "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "relative.sid_code_accuracy_percent", "variants.UniR2.final_loss", "variants.UniR2.hit_at_10", "variants.UniR2.initial_loss", "variants.UniR2.ndcg_at_10", "variants.UniR2.sid_code_accuracy", "variants.separate cascade.final_loss", "variants.separate cascade.hit_at_10", "variants.separate cascade.initial_loss", "variants.separate cascade.ndcg_at_10", "variants.separate cascade.sid_code_accuracy" ], "default_seeds": [ 42 ], "budget": "steps=130", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou App", "metric": "play volume", "lift_percent": 1.177, "traffic": "5% traffic, two weeks", "source_url": "https://arxiv.org/abs/2607.24439", "source_location": "original paper online-result disclosure: product=Kuaishou App; metric=play volume" }, { "product": "Kuaishou App", "metric": "like rate", "lift_percent": 2.56, "traffic": "5% traffic, two weeks", "source_url": "https://arxiv.org/abs/2607.24439", "source_location": "original paper online-result disclosure: product=Kuaishou App; metric=like rate" }, { "product": "Kuaishou Lite", "metric": "total gifting amount", "lift_percent": 2.569, "traffic": "5% traffic, two weeks", "source_url": "https://arxiv.org/abs/2607.24439", "source_location": "original paper online-result disclosure: product=Kuaishou Lite; metric=total gifting amount" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "unirec", "title": "UniRec: Cross-stage Multi-Task Fusion with Preference Alignment for Cascaded Recommender Systems", "paper_url": "https://arxiv.org/abs/2609.11052", "detail_path": "reproductions/2609.11052-unirec/README.md", "topic": [ "pre-ranking", "ranking", "cross-stage-alignment", "industrial-recommendation" ], "first_author": null, "first_author_affiliation": "Kuaishou", "published": "2026-09-10", "code": null, "adapter": { "adapter_key": "unirec", "arxiv_id": "2609.11052", "title": "UniRec: Cross-stage Multi-Task Fusion with Preference Alignment for Cascaded Recommender Systems", "paper_url": "https://arxiv.org/abs/2609.11052", "track": "recommendation", "organization": "Kuaishou", "published": "2026-09-10", "code_url": null, "topics": [ "pre-ranking", "ranking", "cross-stage-alignment", "industrial-recommendation" ], "local_code_dir": "src/auto_research/reproductions/unirec", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Kuaishou live recommender", "metric": "app usage", "lift_percent": 0.616, "traffic": "20% live traffic followed by full deployment", "source_url": "https://arxiv.org/html/2609.11052", "source_location": "Section 5.4, Table 5", "retrieved_at": "2026-09-12" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "unirec-coa", "title": "UniRec: Bridging the Expressive Gap between Generative and Discriminative Recommendation via Chain-of-Attribute", "paper_url": "https://arxiv.org/abs/2604.12234", "detail_path": "reproductions/2604.12234-unirec-coa/README.md", "topic": [ "generative-recommendation", "semantic-id", "preference-alignment" ], "first_author": null, "first_author_affiliation": "Authors did not disclose affiliation / large-scale e-commerce platform", "published": "2026-04-14", "code": null, "adapter": { "adapter_key": "unirec-coa", "arxiv_id": "2604.12234", "title": "UniRec: Bridging the Expressive Gap between Generative and Discriminative Recommendation via Chain-of-Attribute", "paper_url": "https://arxiv.org/abs/2604.12234", "track": "recommendation", "organization": "Authors did not disclose affiliation / large-scale e-commerce platform", "published": "2026-04-14", "code_url": null, "topics": [ "generative-recommendation", "semantic-id", "preference-alignment" ], "local_code_dir": "src/auto_research/reproductions/unirec_coa", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "e-commerce feed and landing page", "metric": "GMV", "lift_percent": 5.6, "traffic": "20% user traffic per bucket", "source_url": "https://arxiv.org/html/2604.12234v1", "source_location": "Section 4.2 / Table 5", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "uniscale", "title": "UniScale: Synergistic Entire Space Data and Model Scaling for Search Ranking", "paper_url": "https://arxiv.org/abs/2603.24226", "detail_path": "reproductions/2603.24226-uniscale/README.md", "topic": [ "search-ranking", "model-scaling", "entire-space-learning" ], "first_author": null, "first_author_affiliation": "Taobao & Tmall Group / Alibaba", "published": "2026-03-25", "code": null, "adapter": { "adapter_key": "uniscale", "arxiv_id": "2603.24226", "title": "UniScale: Synergistic Entire Space Data and Model Scaling for Search Ranking", "paper_url": "https://arxiv.org/abs/2603.24226", "track": "recommendation", "organization": "Taobao & Tmall Group / Alibaba", "published": "2026-03-25", "code_url": null, "topics": [ "search-ranking", "model-scaling", "entire-space-learning" ], "local_code_dir": "src/auto_research/reproductions/uniscale", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K" ], "baseline": "shared transition + content scorer", "metrics": [ "Hit@10", "NDCG@10", "Fresh Hit@10", "Head share@10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Taobao Search", "metric": "GMV", "lift_percent": 2.04, "traffic": "5% traffic, ten days", "source_url": "https://arxiv.org/html/2603.24226v1", "source_location": "Section 4.5", "retrieved_at": "2026-08-24" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "univa", "title": "Unified Value Alignment for Generative Recommendation in Industrial Advertising", "paper_url": "https://arxiv.org/abs/2605.05803", "detail_path": "reproductions/2605.05803-univa/README.md", "topic": [ "generative-advertising", "commercial-sid", "ppo", "value-aware-serving" ], "first_author": null, "first_author_affiliation": "Tencent / WeChat Channels", "published": "2026-05", "code": null, "adapter": { "adapter_key": "univa", "arxiv_id": "2605.05803", "title": "Unified Value Alignment for Generative Recommendation in Industrial Advertising", "paper_url": "https://arxiv.org/abs/2605.05803", "track": "recommendation", "organization": "Tencent / WeChat Channels", "published": "2026-05", "code_url": null, "topics": [ "generative-advertising", "commercial-sid", "ppo", "value-aware-serving" ], "local_code_dir": "src/auto_research/reproductions/univa", "fidelity": "full_pipeline", "evaluation_tier": "l3_paper_pipeline", "datasets": [ "MiniOneRec Amazon Office_Products" ], "baseline": "同一 MiniOneRec Office 数据、相同 HSTU/decoder 容量、步数、随机种子和全商品目录下的 semanticsidsl（原始三级 Semantic SID + 监督学习）", "metrics": [ "selected_relative_to_semantic_sid_percent.wndcg_at_100", "test.commercial_sid_sl.wndcg_at_100", "test.semantic_sid_sl.wndcg_at_100", "test.univa_full.wndcg_at_100" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=80", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "WeChat Channels Ads", "metric": "GMV", "lift_percent": 1.5, "traffic": "5% traffic, 2026-03-07 to 2026-03-11", "source_url": "https://arxiv.org/abs/2605.05803", "source_location": "original paper online-result disclosure: product=WeChat Channels Ads; metric=GMV" }, { "product": "WeChat Channels Ads", "metric": "GMV(normal)", "lift_percent": 1.42, "traffic": "5% traffic, 2026-03-07 to 2026-03-11", "source_url": "https://arxiv.org/abs/2605.05803", "source_location": "original paper online-result disclosure: product=WeChat Channels Ads; metric=GMV(normal)" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "whale", "title": "WHALE: A Scalable Unified Model for Recommendation with Wukong-HSTU Architecture", "paper_url": "https://arxiv.org/abs/2607.17017", "detail_path": "reproductions/2607.17017-whale/README.md", "topic": [ "ranking", "hstu", "feature-interaction", "scaling" ], "first_author": null, "first_author_affiliation": "Meta Platforms, Inc.", "published": "2026-07-19", "code": null, "adapter": { "adapter_key": "whale", "arxiv_id": "2607.17017", "title": "WHALE: A Scalable Unified Model for Recommendation with Wukong-HSTU Architecture", "paper_url": "https://arxiv.org/abs/2607.17017", "track": "recommendation", "organization": "Meta Platforms, Inc.", "published": "2026-07-19", "code_url": null, "topics": [ "ranking", "hstu", "feature-interaction", "scaling" ], "local_code_dir": "src/auto_research/reproductions/whale", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "双分支 late fusion，实验组为 WHALE 渐进交换", "metrics": [ "baseline.hit_at_10", "baseline.ndcg_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42 ], "budget": "steps=60", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta social recommendation surface", "metric": "Primary metric", "lift_percent": 0.113, "traffic": "14 days; production baseline", "source_url": "https://arxiv.org/abs/2607.17017", "source_location": "original paper online-result disclosure: product=Meta social recommendation surface; metric=Primary metric" }, { "product": "Meta social recommendation surface", "metric": "Metric 1", "lift_percent": 0.824, "traffic": "14 days; production baseline", "source_url": "https://arxiv.org/abs/2607.17017", "source_location": "original paper online-result disclosure: product=Meta social recommendation surface; metric=Metric 1" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "wide-deep", "title": "Wide & Deep Learning for Recommender Systems", "paper_url": "https://arxiv.org/abs/1606.07792", "detail_path": "reproductions/1606.07792-wide-deep/README.md", "topic": [ "ranking", "classic", "feature-interaction" ], "first_author": null, "first_author_affiliation": "Google", "published": "2016-06-24", "code": "https://github.com/tensorflow/models/tree/master/official/r1/wide_deep", "adapter": { "adapter_key": "wide-deep", "arxiv_id": "1606.07792", "title": "Wide & Deep Learning for Recommender Systems", "paper_url": "https://arxiv.org/abs/1606.07792", "track": "recommendation", "organization": "Google", "published": "2016-06-24", "code_url": "https://github.com/tensorflow/models/tree/master/official/r1/wide_deep", "topics": [ "ranking", "classic", "feature-interaction" ], "local_code_dir": "src/auto_research/reproductions/wide_deep", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K compact" ], "baseline": "参数匹配的 deep-only 模型", "metrics": [ "baseline.ndcg_at_10", "method.ndcg_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Google Play", "metric": "app acquisition", "lift_percent": 3.9, "traffic": "1% treatment vs 1% control, three weeks", "source_url": "https://arxiv.org/abs/1606.07792", "source_location": "original paper online-result disclosure: product=Google Play; metric=app acquisition" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "youtube-dnn", "title": "Deep Neural Networks for YouTube Recommendations", "paper_url": "https://research.google/pubs/deep-neural-networks-for-youtube-recommendations/", "detail_path": "reproductions/recsys2016-youtube-dnn-youtube-dnn/README.md", "topic": [ "retrieval", "ranking", "classic", "two-stage" ], "first_author": null, "first_author_affiliation": "Google / YouTube", "published": "2016-09-15", "code": null, "adapter": { "adapter_key": "youtube-dnn", "arxiv_id": "recsys2016-youtube-dnn", "title": "Deep Neural Networks for YouTube Recommendations", "paper_url": "https://research.google/pubs/deep-neural-networks-for-youtube-recommendations/", "track": "recommendation", "organization": "Google / YouTube", "published": "2016-09-15", "code_url": null, "topics": [ "retrieval", "ranking", "classic", "two-stage" ], "local_code_dir": "src/auto_research/reproductions/youtube_dnn", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens-100K compact" ], "baseline": "历史均值 two-tower，实验组增加 YouTube DNN 非线性用户塔", "metrics": [ "baseline.head_share_at_10", "baseline.head_share_at_10_std", "baseline.hit_at_10", "baseline.hit_at_10_std", "baseline.ndcg_at_10", "baseline.ndcg_at_10_std", "method.head_share_at_10", "method.head_share_at_10_std", "method.hit_at_10", "method.hit_at_10_std", "method.ndcg_at_10", "method.ndcg_at_10_std", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent" ], "default_seeds": [ 42, 43, 44 ], "budget": "steps=80", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [], "selection_exception": "用户明确批准 YouTube DNN 为经典例外；论文描述线上服务系统但未披露可核验的量化 lift。", "evolve_operators": [] } }, { "domain": "recommendation", "key": "youtube-freshness", "title": "Breaking the Loop: An Empirical Comparison of Strategies for Novelty and Freshness in YouTube Music", "paper_url": "https://arxiv.org/abs/2607.23749", "detail_path": "reproductions/2607.23749-youtube-freshness/README.md", "topic": [ "ranking", "feedback-loop", "freshness", "exploration" ], "first_author": null, "first_author_affiliation": "YouTube Music / Google", "published": "2026-07-26", "code": null, "adapter": { "adapter_key": "youtube-freshness", "arxiv_id": "2607.23749", "title": "Breaking the Loop: An Empirical Comparison of Strategies for Novelty and Freshness in YouTube Music", "paper_url": "https://arxiv.org/abs/2607.23749", "track": "recommendation", "organization": "YouTube Music / Google", "published": "2026-07-26", "code_url": null, "topics": [ "ranking", "feedback-loop", "freshness", "exploration" ], "local_code_dir": "src/auto_research/reproductions/youtube_freshness", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "popularity-biased continuous ranker，实验组为四机制组合", "metrics": [ "baseline.fresh_hit_at_10", "baseline.head_share_at_10", "baseline.hit_at_10", "baseline.ndcg_at_10", "method.fresh_hit_at_10", "method.head_share_at_10", "method.hit_at_10", "method.ndcg_at_10", "relative.fresh_hit_at_10_percent", "relative.head_share_at_10_percent", "relative.hit_at_10_percent", "relative.ndcg_at_10_percent", "stages.architecture_bias_tower_removed_at_serving", "stages.validation_only_model_selection.fresh_hit_at_10", "stages.validation_only_model_selection.hit_at_10", "stages.validation_only_model_selection.ndcg_at_10" ], "default_seeds": [ 42 ], "budget": "adapter-defined fixed budget", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "YouTube Music homepage", "metric": "1-day new-release engagement", "lift_percent": 4.33, "traffic": "two-week A/B; millions of daily users per arm", "source_url": "https://arxiv.org/abs/2607.23749", "source_location": "original paper online-result disclosure: product=YouTube Music homepage; metric=1-day new-release engagement" } ], "selection_exception": null, "evolve_operators": [] } }, { "domain": "recommendation", "key": "zenith", "title": "Zenith: Scaling up Ranking Models for Billion-scale Livestreaming Recommendation", "paper_url": "https://arxiv.org/abs/2601.21285", "detail_path": "reproductions/2601.21285-zenith/README.md", "topic": [ "ranking-network", "scaling-law", "feature-interaction" ], "first_author": null, "first_author_affiliation": "North Carolina State University", "published": "2026-01-29", "code": null, "adapter": { "adapter_key": "zenith", "arxiv_id": "2601.21285", "title": "Zenith: Scaling up Ranking Models for Billion-scale Livestreaming Recommendation", "paper_url": "https://arxiv.org/abs/2601.21285", "track": "recommendation", "organization": "North Carolina State University", "published": "2026-01-29", "code_url": null, "topics": [ "ranking-network", "scaling-law", "feature-interaction" ], "local_code_dir": "src/auto_research/reproductions/zenith", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "transition + content + popularity", "metrics": [ "hit_at_10", "ndcg_at_10", "fresh_hit_at_10", "head_share_at_10" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "TikTok Live", "metric": "Quality watch sessions per user", "lift_percent": 9.93, "traffic": "production online A/B", "source_url": "https://arxiv.org/html/2601.21285v1", "source_location": "Table 4", "retrieved_at": "2026-09-05" } ], "selection_exception": null, "evolve_operators": [ "zenith" ] } }, { "domain": "recommendation", "key": "zorro", "title": "ZoRRO: A Zero-Weight Personalized Recommender System for Scalable News Recommendation", "paper_url": "https://arxiv.org/abs/2607.10910", "detail_path": "reproductions/2607.10910-zorro/README.md", "topic": [ "news-recommendation", "training-free", "recency" ], "first_author": null, "first_author_affiliation": "Technical University of Denmark", "published": "2026-07-12", "code": null, "adapter": { "adapter_key": "zorro", "arxiv_id": "2607.10910", "title": "ZoRRO: A Zero-Weight Personalized Recommender System for Scalable News Recommendation", "paper_url": "https://arxiv.org/abs/2607.10910", "track": "recommendation", "organization": "Technical University of Denmark", "published": "2026-07-12", "code_url": null, "topics": [ "news-recommendation", "training-free", "recency" ], "local_code_dir": "src/auto_research/reproductions/zorro", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "most-popular/content-transition ranker", "metrics": [ "hit_at_10", "ndcg_at_10", "trainable_parameters", "history_items" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; validation-only blend selection", "device_capabilities": [ "cpu" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Ekstra Bladet", "metric": "CTR", "lift_percent": 4.19, "traffic": "six-day live test; 45% treatment traffic", "source_url": "https://arxiv.org/html/2607.10910v1", "source_location": "Section 4.3, Table 7", "retrieved_at": "2026-09-02" } ], "selection_exception": null, "evolve_operators": [ "context:zero-weight" ] } } ] }