{ "schema_version": 1, "batch": "2026-10-29-priority-components", "counts": { "selected_papers": 9, "components_accepted": 8, "gpu_validated_components": 4, "cpu_only_components": 4, "reference_only_high_cost_done": 1, "components_pending_verification": 0, "capability_reproductions_claimed": 0, "latest_category_entries": 1478, "latest_unique_identities": 1059, "selected_first_publications_after_previous_capture": 6, "selected_historical_official_source_backfills": 2, "section_entry_counts": { "new": 546, "cross-list": 355, "replacement": 577 } }, "selected_scope_complete": true, "delivery_status": "bounded_scope_complete", "all_sources_exhaustive": false, "review_watermark_advanced": false, "review_watermark": { "before": "2026-10-15", "after": "2026-10-15" }, "fixed_component_keys": [ "discernmem", "gui-jepa", "hope-hilbert", "kl-mopd", "rescom", "ruharness", "select-ttt", "sleep-lti" ], "reference_only_keys": [ "pailitao-mmsearch2" ], "papers": [ { "domain": "agent-research", "key": "discernmem", "title": "Remembering Is Not Enough: Refining Memories and Learning When to Use Them in Multi-Agent Systems", "paper_url": "https://arxiv.org/abs/2610.36652v1", "detail_path": "agent-research/2610.36652-discernmem/README.md", "priority": "P1", "first_author": "Yipeng Hao", "first_author_affiliation": "East China Normal University", "published_at": "2026-10-27T05:12:46Z", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.36652v1#A3", "model_scale": "Qwen3-4B-Instruct-2507与Qwen3-Embedding-0.6B；AutoGen/CAMEL/MacNet三架构、四benchmark，另需QA检索与代码执行GPU", "paper_configuration": "原论文Appendix C明确报告主实验、消融和超参扫描合计约900 GPU小时；每配置只跑一次，不能冒称多seed能力结论", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "decision": "成本过高，不复现（审查 DONE）；不下载模型/数据、不启动LLM replay矩阵；仅独立纯CPU测量合同组件，不自动重试" }, "implementation_status": "component_only", "id": "2610.36652", "intake_class": "new_announcement", "status": "implemented", "first_publication_after_previous_capture": true, "component_accepted": true, "component_implemented": true, "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "resolution": "bounded_l1_component_accepted", "requires_gpu_validation": false, "gpu_validation_passed": false, "component_evidence": "docs/agent-research/2610.36652-discernmem/metrics/component-cpu-seeds42-44.json", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false }, { "domain": "foundation-models", "key": "gui-jepa", "title": "GUI-JEPA: Candidate-Conditioned JEPA Latents as a Fast Action Policy for Mobile GUI Agents", "paper_url": "https://arxiv.org/abs/2610.36877v1", "detail_path": "foundation-models/2610.36877-gui-jepa/README.md", "priority": "P1", "first_author": "Bingchao Wang", "first_author_affiliation": "Shanghai Jiao Tong University / Nanyang Technological University", "published_at": "2026-10-27T09:17:51Z", "cost_assessment": { "source": "固定v1 Section3/4和AppendixH：https://arxiv.org/html/2610.36877v1", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "model_scale": "303M ViT-L+63M CLIP及多个ranker，总0.50B；另有Qwen2.5-VL-7B LoRA同协议参考", "data_scale": "189262截图预训，AndroidControl13594episodes/74714steps，6epochs/3seeds和多个testsubsplit", "paper_configuration": "作者H20编码器预训练2739秒约0.76GPUhour，尚不含策略/3种子/ranker/7B完整参考矩阵；不能据此声称完整流程低成本", "estimated_cost_basis": "规模、训练阶段和多臂矩阵只读审查，不是本地完整实测；缩维latent核无需截图/编码器权重", "decision": "完整实验成本过高，不复现（审查 DONE）；不下载、不执行GUI、不试训练、不自动重试。仅残差predictor与结构头L1，其CUDA验收单独绑定真实回执，不代表GUI能力评测。" }, "implementation_status": "component_only", "id": "2610.36877", "intake_class": "newly_announced", "status": "implemented", "first_publication_after_previous_capture": true, "component_accepted": true, "component_implemented": true, "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "resolution": "bounded_l1_component_accepted", "requires_gpu_validation": true, "gpu_validation_passed": true, "component_evidence": "docs/foundation-models/2610.36877-gui-jepa/metrics/component-cuda-seeds42-44.json", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false }, { "domain": "foundation-models", "key": "hope-hilbert", "title": "Hilbert Operator for Progressive Encoding (HOPE): A Mathematical Framework for Deconstructing Learned Representations in Deep Networks", "paper_url": "https://arxiv.org/abs/2607.21366v1", "detail_path": "foundation-models/2607.21366-hope-hilbert/README.md", "priority": "P0", "first_author": "Hossein Mobahi", "first_author_affiliation": "Google DeepMind", "published_at": "2026-07-23T14:31:16Z", "cost_assessment": { "source": "论文Sections4–6/9、AppendixE；固定官方README的Run time and memory与Evaluation protocol", "model_scale": "Keras ImageNet ResNet-50；另有DEFT的8层BN-VGG迁移评测", "data_scale": "默认10144张ImageNet图像，每10编码步评估，共308次；另有7条压缩基线、DEFT预搜索及20迁移场景", "available_resource": "单张A100或A30，或本地Mac", "known_cost": "作者报告共享64核CPU编码7–10分钟，完整默认评估13–24小时；GPU总耗时未给出，不臆造", "decision": "完整实验成本过高，不复现（审查DONE）；不下载ImageNet/权重，不运行DEFT或完整压缩。只实现纯CPU的小规模核、剪枝与单步可逆合同" }, "implementation_status": "component_only", "id": "2607.21366", "intake_class": "official_code_release_backfill", "status": "implemented", "first_publication_after_previous_capture": false, "component_accepted": true, "component_implemented": true, "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "resolution": "bounded_l1_component_accepted", "requires_gpu_validation": false, "gpu_validation_passed": false, "component_evidence": "docs/foundation-models/2607.21366-hope-hilbert/metrics/component-cpu-seeds42-44.json", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false }, { "domain": "recommendation", "key": "kl-mopd", "title": "KL-MOPD: Discrepancy-Aware Multi-Objective On-Policy Distillation for Generative Recommendation", "paper_url": "https://arxiv.org/abs/2610.36973v1", "detail_path": "recommendation-components/2610.36973-kl-mopd/README.md", "priority": "P1", "first_author": "Weiqin Yang", "first_author_affiliation": "Zhejiang University", "published_at": "2026-10-27T10:28:05Z", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.36973v1", "source_location": "Section4.1", "model_scale": "TIGER/REG4Rec; two encoder/decoder layers dim64 on public datasets;5decoder layers dim1024 on Industry", "data_scale": "three public datasets plus billion-scale private logs;last-two-day fine-tune/eval each>11M requests", "available_resource": "One NVIDIA A100 or A30, or local Mac", "estimated_cost_basis": "initial training, two teacher fine-tunes, distillation and multi-baseline sweeps, not a single tensor test;GPU hours undisclosed", "decision": "成本过高，不复现（审查 DONE）；只执行公式组件，不下载大数据，不自动重试" }, "implementation_status": "component_only", "id": "2610.36973", "intake_class": "new_announcement", "status": "implemented", "first_publication_after_previous_capture": true, "component_accepted": true, "component_implemented": true, "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "resolution": "bounded_l1_component_accepted", "requires_gpu_validation": true, "gpu_validation_passed": true, "component_evidence": "docs/recommendation-components/2610.36973-kl-mopd/metrics/component-cuda-seeds42-44.json", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false }, { "domain": "recommendation", "key": "pailitao-mmsearch2", "title": "PailitaoMMSearch2.0 Technical Report: Scaling Generative Retrieval to Unified End-to-End Multimodal Search", "paper_url": "https://arxiv.org/abs/2610.36958v1", "detail_path": "recommendation-components/2610.36958-pailitao-mmsearch2/README.md", "priority": "P1", "first_author": "Xiaohan Ye", "first_author_affiliation": "Taobao & Tmall Group of Alibaba", "published_at": "2026-10-27T10:17:33Z", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.36958v1", "source_location": "Sections4–7", "model_scale": "0.8B/2B/4B VLM, CPT/SFT/intent-balancedRL andSample–Beam", "data_scale": "large production image-text/catalog data andmulti-scenario14dayA/B;3SIDlevels×8192codes", "available_resource": "One NVIDIA A100 or A30, or local Mac", "estimated_cost_basis": "private multimodal data and multi-stage/multi-model rollout training;no totalGPUhourdisclosure", "decision": "成本过高，不复现（审查 DONE）；资料收录，无可执行组件/指标/GPU回执，不自动重试" }, "implementation_status": "deferred_high_cost", "id": "2610.36958", "intake_class": "new_announcement", "status": "deferred", "reason": "成本过高，不复现（审查 DONE）：0.8B/2B/4B多阶段VLM训练依赖私有生产数据和真实A/B，单卡预算不执行完整实验；仅资料收录，不自动重试。", "first_publication_after_previous_capture": true, "component_accepted": false, "component_implemented": false, "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "resolution": "not_reproduced_high_cost", "requires_gpu_validation": false, "gpu_validation_passed": false, "component_evidence": null, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false }, { "domain": "foundation-models", "key": "rescom", "title": "ResCom: Activation Residual Compensation for Large Language Model Quantization", "paper_url": "https://arxiv.org/abs/2610.35955v1", "detail_path": "foundation-models/2610.35955-rescom/README.md", "priority": "P1", "first_author": "Ning Yang", "first_author_affiliation": "Institute of Automation, Chinese Academy of Sciences", "published_at": "2026-10-25T03:02:21Z", "cost_assessment": { "source": "固定v1 Sections3/4和AppendicesG/H：https://arxiv.org/html/2610.35955v1", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "model_scale": "完整比较包含Llama/Mistral/Gemma 1B至70B、多量化校准方法和原生MXFP4", "data_scale": "128个2048token WikiText-2校准窗口，另有C4、MMLU、GSM8K、LongBench及下游矩阵", "paper_configuration": "B200上闭式校准、模型质量评测、compact-storage与原生低比特内核测量；作者1B/3B校准耗时不代表A100/A30完整矩阵耗时", "estimated_cost_basis": "模型/基准/内核矩阵与下载量的只读审查，不是本地完整运行测量；仅缩维线性残差核无资源下载", "decision": "完整实验成本过高，不复现（审查 DONE）；不下载、不试跑、不自动重试。只接受独立L1闭式残差组件，其GPU验收单独绑定真实回执，不代表完整模型实验。" }, "implementation_status": "component_only", "id": "2610.35955", "intake_class": "newly_announced", "status": "implemented", "first_publication_after_previous_capture": false, "component_accepted": true, "component_implemented": true, "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "resolution": "bounded_l1_component_accepted", "requires_gpu_validation": true, "gpu_validation_passed": true, "component_evidence": "docs/foundation-models/2610.35955-rescom/metrics/component-cuda-seeds42-44.json", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false }, { "domain": "agent-research", "key": "ruharness", "title": "When the Critic Falls Behind: Co-Evolving Rubrics and Harnesses for Continual Agent Self-Improvement", "paper_url": "https://arxiv.org/abs/2610.36826v1", "detail_path": "agent-research/2610.36826-ruharness/README.md", "priority": "P1", "first_author": "Hao Chen", "first_author_affiliation": "Harbin Institute of Technology", "published_at": "2026-10-27T08:24:44Z", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.36826v1#S4", "model_scale": "Qwen3.5-35B-A3B和GLM-4.7-Flash，由同一大模型执行Agent、诊断、rubric更新和harness提案", "paper_configuration": "EvoAgentBench五域；713条evolution、268条validation与337条held-out；多轮每轮至多3个协调修改候选，每个至多3组件，重复真实工具/代码/检索任务；原文未报告总GPU小时，不编造耗时", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "decision": "大模型多域长程harness进化矩阵成本过高，不复现（审查 DONE）；仅纯CPU统计/接受门槛，不下载权重或试跑API，不自动重试" }, "implementation_status": "component_only", "id": "2610.36826", "intake_class": "new_announcement", "status": "implemented", "first_publication_after_previous_capture": true, "component_accepted": true, "component_implemented": true, "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "resolution": "bounded_l1_component_accepted", "requires_gpu_validation": false, "gpu_validation_passed": false, "component_evidence": "docs/agent-research/2610.36826-ruharness/metrics/component-cpu-seeds42-44.json", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false }, { "domain": "foundation-models", "key": "select-ttt", "title": "Select-TTT: Query-Conditioned Selective Fast Weight Updates for In-Place Test-Time Training", "paper_url": "https://arxiv.org/abs/2610.36794v1", "detail_path": "foundation-models/2610.36794-select-ttt/README.md", "priority": "P1", "first_author": "Yue Wu", "first_author_affiliation": "University of Science and Technology of China", "published_at": "2026-10-27T07:58:39Z", "cost_assessment": { "source": "固定v1 Sections4/5和AppendixB：https://arxiv.org/html/2610.36794v1", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "model_scale": "Llama-3.1-8B-Instruct及Qwen3-4B，真实In-PlaceTTT插桩和CPT", "data_scale": "LongMemEval-S约115Ktoken历史、LoCoMo，以及LongCite-45K过滤子集", "paper_configuration": "H200上batch8、最长65536token、最多5000训练steps，另有双backbone多基线推理/消融矩阵", "estimated_cost_basis": "长上下文训练/评测和模型规模只读审查，非本地完整成本测量；张量公式核无需checkpoint或数据", "decision": "完整实验成本过高，不复现（审查 DONE）；不下载、不启动探针、不自动重试。仅L1评分/写入规则，其GPU验收单独绑定真实回执，不代表完整模型实验。" }, "implementation_status": "component_only", "id": "2610.36794", "intake_class": "newly_announced", "status": "implemented", "first_publication_after_previous_capture": true, "component_accepted": true, "component_implemented": true, "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "resolution": "bounded_l1_component_accepted", "requires_gpu_validation": true, "gpu_validation_passed": true, "component_evidence": "docs/foundation-models/2610.36794-select-ttt/metrics/component-cuda-seeds42-44.json", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false }, { "domain": "foundation-models", "key": "sleep-lti", "title": "Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories", "paper_url": "https://arxiv.org/abs/2606.03979v2", "detail_path": "foundation-models/2606.03979-sleep-lti/README.md", "priority": "P0", "first_author": "Ali Behrouz", "first_author_affiliation": "Google Research", "published_at": "2026-06-02T17:56:55Z", "cost_assessment": { "source_url": "https://arxiv.org/html/2606.03979v2", "source_location": "Sections3–4 and AppendixB", "model_scale": "3B/8B Hope/CMS, low-rank parameter expansion and dream-specific training/selection", "data_scale": "continual tasks and BABILong up to10M context", "available_resource": "One NVIDIA A100 or A30, or local Mac", "estimated_cost_basis": "long-context multi-phase training, sampling and repeated LoRA fits; fullGPUhours not supplied", "decision": "成本过高，不复现（审查 DONE）；只实现CPU LTI奖励，不下载/试训/自动重试" }, "implementation_status": "component_only", "id": "2606.03979", "intake_class": "official_blog_historical_backfill", "status": "implemented", "first_publication_after_previous_capture": false, "component_accepted": true, "component_implemented": true, "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "resolution": "bounded_l1_component_accepted", "requires_gpu_validation": false, "gpu_validation_passed": false, "component_evidence": "docs/foundation-models/2606.03979-sleep-lti/metrics/component-cpu-seeds42-44.json", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false } ], "updated_existing_keys": [ "embedding-gemma2", "rpl" ], "additional_dispositions": [ { "id": "2610.37041", "version": "v1", "key": "guided-rlvr", "domain": "post-training", "title": "Guided-RLVR: Guided Exploration in Reinforcement Learning with Verifiable Rewards", "paper_url": "https://arxiv.org/abs/2610.37041v1", "reviewed_version_url": "https://arxiv.org/html/2610.37041v1", "first_author": "Bo Yang", "first_author_affiliation": "Beihang University", "published_at": "2026-10-27T11:13:13Z", "announced_on": "2026-10-28", "code": null, "code_status": "截至2026-10-29，固定原文/摘要未给公开作者仓库，精确题名查找也未确认原作者代码；不宣称已开源", "upstream_revision": null, "priority": "P1", "intake_class": "new_announcement", "implementation_status": "deferred_high_cost", "full_experiment_status": "not_reproduced_high_cost", "review_resolution": "not_reproduced_high_cost", "review_done": true, "automatic_retry": false, "automatic_implementation_queue": false, "cost_assessment": { "source_url": "https://arxiv.org/html/2610.37041v1#S4.SS5", "source_sections": "Section4.5 Table4, AppendixB/C", "available_resource": "one NVIDIA A100 or A30, or local Mac", "model_scale": "Qwen2.5-7B-Instruct / Qwen3-4B + jointly trained MLP controller; optional DeepSeek-R1 trace supervised cold-start", "data_scale": "MATH/Code-R1各2000训练、500测试；竞赛3任务以2000 DAPO样本训练；G=10，其中8 guided，最多8192tokens", "measured_paper_cost": "作者独立MATH timing: one RTX PRO6000 Blackwell,5000steps,17.92GPUh for Guided-RLVR(no cold-start),vs15.92GRPO; this is author measurement,not local", "decision": "成本过高，不复现（审查 DONE）；不下载大模型/trace，不启动训练探针或多臂benchmark，不自动重试" }, "method": "句界、高entropy且满足间隔时，轻量控制器选语义steering token，采样侧加非负logit偏置；correct且实际采到promotion的math rollout才加log-perplexity差bonus。训练policy以同前缀未加boost的old/current比率更新，controller以成功intervention事件和共享group advantage更新。", "fidelity_caveat": "AppendixA Eq5的exact delta与Eq6/final使用近似delta不同；ptarget=.5时近似delta=-log(pk)>0，即pk>.5仍增加，而正文‘已超过target就clip0’文字并不随近似成立。不能无说明改用exact公式并称完整忠实原实现。AppendixB.2阈值需明确当前entropy加入窗口时序，先不以自行假定填补controller。", "source_evidence": [ "https://arxiv.org/abs/2610.37041v1", "https://arxiv.org/html/2610.37041v1#S3.SS2", "https://arxiv.org/html/2610.37041v1#A1", "https://arxiv.org/html/2610.37041v1#S4.SS5" ] } ] }