{ "schema_version": 1, "batch": "2026-10-23-priority-components", "delivery_status": "bounded_scope_complete", "all_sources_exhaustive": false, "review_watermark_advanced": false, "papers": [ { "id": "2609.30716", "key": "gemma-priority-protocol", "domain": "foundation-models", "title": "Words Speak Louder Than Order: A Behavioral Evaluation of Gemma 4", "paper_url": "https://arxiv.org/abs/2609.30716v1", "detail_path": "foundation-models/2609.30716-gemma-priority-protocol/README.md", "priority": "P0", "first_author": "Amanda Fitch", "first_author_affiliation": "Google", "verified_priority_organizations": [ "Google" ], "original_published_at": "2026-09-25T02:43:40Z", "intake_class": "historical-backfill", "first_publication_after_baseline_capture": false, "observed_versions": [], "bulletin_categories": [], "bulletin_entry_kinds": [], "status": "implemented", "implementation_status": "component_only", "delivery_status": "bounded_component_accepted", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_run_scope_limited_protocol", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "requires_gpu_validation": false, "cost_assessment": { "source_url": "https://arxiv.org/html/2609.30716v1", "basis": "named method/resource passages and official author release; not measured complete-run time", "scale": "双文档冲突的四路正交对照、十条件共同保留集、逐项目contrast及BCa/TOST。CPU协议与固定作者log复算，不是新模型推理；退化和BCa失败不返回通过。784次FP32模型实跑未进入本轮scope，未下载checkpoint，也不无依据标为高成本。", "available_resource": "One NVIDIA A100 or A30, or local Mac" }, "component_evidence": "docs/foundation-models/2609.30716-gemma-priority-protocol/metrics/component-cpu-seeds42-44.json", "resolution": "bounded_l1_component_accepted", "reason": "冻结源码三种子L1组件验收；CUDA组件真实A100执行。完整实验状态逐篇保留，不宣称模型能力或完整论文复现。" }, { "id": "2610.30951", "key": "openattention", "domain": "foundation-models", "title": "OpenAttention: One Percent Is Enough with Tail Proxies", "paper_url": "https://arxiv.org/abs/2610.30951v1", "detail_path": "foundation-models/2610.30951-openattention/README.md", "priority": "P1", "first_author": "Hyunwoo Yu", "first_author_affiliation": "Sogang University", "verified_priority_organizations": [], "original_published_at": "2026-10-22T09:03:08Z", "intake_class": "new-to-repository-recent-first-publication", "first_publication_after_baseline_capture": true, "observed_versions": [ "2610.30951v1" ], "bulletin_categories": [ "cs.AI", "cs.LG" ], "bulletin_entry_kinds": [ "cross-list", "new" ], "status": "implemented", "implementation_status": "component_only", "delivery_status": "bounded_component_accepted", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "requires_gpu_validation": true, "cost_assessment": { "source_url": "https://arxiv.org/html/2610.30951v1", "basis": "named method/resource passages and official author release; not measured complete-run time", "scale": "单头参考保留QGC查询二阶矩几何聚类、int8 QGS、ErrorGate与water-filling精确token预算，以及剔除已读token后重算剩余计数/均值并加log(count)的softmax代理。仅生成张量机制诊断；不声称1%读取准确率或Triton加速。7B/8B多模型、16K至96K长上下文全基准成本过高，不复现（审查DONE）。", "available_resource": "One NVIDIA A100 or A30, or local Mac" }, "component_evidence": "docs/foundation-models/2610.30951-openattention/metrics/component-cuda-seeds42-44.json", "resolution": "bounded_l1_component_accepted", "reason": "冻结源码三种子L1组件验收；CUDA组件真实A100执行。完整实验状态逐篇保留，不宣称模型能力或完整论文复现。", "gpu_validation_artifact": "docs/gpu-validations/openattention-a100-20261023.json" }, { "id": "2610.30206", "key": "moezip", "domain": "foundation-models", "title": "MoEZip: Routing-Aware KV Cache Compression for Sparse Mixture-of-Experts LLMs", "paper_url": "https://arxiv.org/abs/2610.30206v1", "detail_path": "foundation-models/2610.30206-moezip/README.md", "priority": "P1", "first_author": "Minsang Kim", "first_author_affiliation": "Korea University", "verified_priority_organizations": [], "original_published_at": "2026-10-21T13:06:30Z", "intake_class": "new-to-repository-recent-first-publication", "first_publication_after_baseline_capture": false, "observed_versions": [ "2610.30206v1" ], "bulletin_categories": [ "cs.AI", "cs.LG" ], "bulletin_entry_kinds": [ "cross-list", "new" ], "status": "implemented", "implementation_status": "component_only", "delivery_status": "bounded_component_accepted", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "requires_gpu_validation": true, "cost_assessment": { "source_url": "https://arxiv.org/html/2610.30206v1", "basis": "named method/resource passages and official author release; not measured complete-run time", "scale": "保留softmax/sigmoid的logit坐标Fisher trace、instruction-to-context跨头max注意力传播、fragility自适应全局anchor预算及每层每头base无重叠补齐，并真实gather KV。只接收prefill统计，不接收答案或gold plan，不加载模型。论文每run两张A100-80GB、235B用八卡、总15000至16000 A100小时；完整实验成本过高，不复现（审查DONE）。", "available_resource": "One NVIDIA A100 or A30, or local Mac" }, "component_evidence": "docs/foundation-models/2610.30206-moezip/metrics/component-cuda-seeds42-44.json", "resolution": "bounded_l1_component_accepted", "reason": "冻结源码三种子L1组件验收；CUDA组件真实A100执行。完整实验状态逐篇保留，不宣称模型能力或完整论文复现。", "gpu_validation_artifact": "docs/gpu-validations/moezip-a100-20261023.json" }, { "id": "2610.31563", "key": "aopd", "domain": "post-training", "title": "When to Follow the Teacher: Advantage-Based On-Policy Distillation", "paper_url": "https://arxiv.org/abs/2610.31563v1", "detail_path": "post-training/2610.31563-aopd/README.md", "priority": "P1", "first_author": "Mahdi Sabbaghi", "first_author_affiliation": "University of Pennsylvania", "verified_priority_organizations": [], "original_published_at": "2026-10-22T17:05:21Z", "intake_class": "new-to-repository-recent-first-publication", "first_publication_after_baseline_capture": true, "observed_versions": [ "2610.31563v1" ], "bulletin_categories": [ "cs.AI", "cs.LG" ], "bulletin_entry_kinds": [ "cross-list", "new" ], "status": "implemented", "implementation_status": "component_only", "delivery_status": "bounded_component_accepted", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "requires_gpu_validation": true, "cost_assessment": { "source_url": "https://arxiv.org/html/2610.31563v1", "basis": "named method/resource passages and official author release; not measured complete-run time", "scale": "按学生优势加权教师动作，以原有student rollout、终局奖励、LOO baseline和教师重要性权重估计双项目标。实现Eq3.6/3.8全词表期望、双项同截断且不归一化、冻结rollout/teacher；穷举两步MDP独立rollout pairs检验期望与梯度，仅L1数值诊断，无LLM能力评测。", "available_resource": "One NVIDIA A100 or A30, or local Mac" }, "component_evidence": "docs/post-training/2610.31563-aopd/metrics/component-cuda-seeds42-44.json", "resolution": "bounded_l1_component_accepted", "reason": "冻结源码三种子L1组件验收；CUDA组件真实A100执行。完整实验状态逐篇保留，不宣称模型能力或完整论文复现。", "gpu_validation_artifact": "docs/gpu-validations/aopd-a100-20261023.json" }, { "id": "2610.30796", "key": "gift", "domain": "post-training", "title": "GIFT: Gradient Informed Filtering for Distillation Augmented Reasoning RL", "paper_url": "https://arxiv.org/abs/2610.30796v1", "detail_path": "post-training/2610.30796-gift/README.md", "priority": "P1", "first_author": "Shijie Li", "first_author_affiliation": "Shanghai Jiao Tong University", "verified_priority_organizations": [], "original_published_at": "2026-10-22T06:22:45Z", "intake_class": "new-to-repository-recent-first-publication", "first_publication_after_baseline_capture": true, "observed_versions": [ "2610.30796v1" ], "bulletin_categories": [ "cs.LG" ], "bulletin_entry_kinds": [ "new" ], "status": "implemented", "implementation_status": "component_only", "delivery_status": "bounded_component_accepted", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "requires_gpu_validation": true, "cost_assessment": { "source_url": "https://arxiv.org/html/2610.30796v1", "basis": "named method/resource passages and official author release; not measured complete-run time", "scale": "在原词表logit空间分解教师KL与GRPO梯度，冲突按a排序、非冲突按正交幅度b排序，exact quota按全valid token预算退火，保留KL不投影且分母不变。实现sparse top-k+tail真实梯度与autograd核验，仅L1诊断；不是永久conflict-free或完整推理训练复现。", "available_resource": "One NVIDIA A100 or A30, or local Mac" }, "component_evidence": "docs/post-training/2610.30796-gift/metrics/component-cuda-seeds42-44.json", "resolution": "bounded_l1_component_accepted", "reason": "冻结源码三种子L1组件验收；CUDA组件真实A100执行。完整实验状态逐篇保留，不宣称模型能力或完整论文复现。", "gpu_validation_artifact": "docs/gpu-validations/gift-a100-20261023.json" }, { "id": "2610.30941", "key": "pair", "domain": "recommendation", "title": "PAIR: Preference-Aligned Intent Retrieval for Premium People Recommendation at LinkedIn", "paper_url": "https://arxiv.org/abs/2610.30941v1", "detail_path": "recommendation-components/2610.30941-pair/README.md", "priority": "P1", "first_author": "Zhongkai Shangguan", "first_author_affiliation": "LinkedIn", "verified_priority_organizations": [], "original_published_at": "2026-10-22T08:57:58Z", "intake_class": "new-to-repository-recent-first-publication", "first_publication_after_baseline_capture": true, "observed_versions": [ "2610.30941v1" ], "bulletin_categories": [ "cs.IR", "cs.LG" ], "bulletin_entry_kinds": [ "cross-list", "new" ], "status": "implemented", "implementation_status": "component_only", "delivery_status": "bounded_component_accepted", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "requires_gpu_validation": true, "cost_assessment": { "source_url": "https://arxiv.org/html/2610.30941v1", "basis": "named method/resource passages and official author release; not measured complete-run time", "scale": "通过曝光条件偏好对比目标、可学习位置历史注意力和归一化加性融合建模职业社交互补意图。本地仅实现可独立验证的向量/损失组件；原8GPU私有数据完整实验成本过高，不复现（审查 DONE）。", "available_resource": "One NVIDIA A100 or A30, or local Mac" }, "component_evidence": "docs/recommendation-components/2610.30941-pair/metrics/component-cuda-seeds42-44.json", "resolution": "bounded_l1_component_accepted", "reason": "冻结源码三种子L1组件验收；CUDA组件真实A100执行。完整实验状态逐篇保留，不宣称模型能力或完整论文复现。", "gpu_validation_artifact": "docs/gpu-validations/pair-a100-20261023.json" }, { "id": "2610.31521", "key": "jev-dlc", "domain": "foundation-models", "title": "Jev-DLC: Decision-Level Continual Learning from Typed Feedback", "paper_url": "https://arxiv.org/abs/2610.31521v1", "detail_path": "system-one/2610.31521-jev-dlc/README.md", "priority": "P1", "first_author": "Yangyang Feng", "first_author_affiliation": "The Hong Kong University of Science and Technology (Guangzhou)", "verified_priority_organizations": [], "original_published_at": "2026-10-22T16:33:30Z", "intake_class": "new-to-repository-recent-first-publication", "first_publication_after_baseline_capture": true, "observed_versions": [ "2610.31521v1" ], "bulletin_categories": [ "cs.AI" ], "bulletin_entry_kinds": [ "new" ], "status": "implemented", "implementation_status": "component_only", "delivery_status": "bounded_component_accepted", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "requires_gpu_validation": false, "cost_assessment": { "source_url": "https://arxiv.org/html/2610.31521v1", "basis": "named method/resource passages and official author release; not measured complete-run time", "scale": "在冻结scorer上结合语义计数、两路PA-I修正和fixed-share可靠性，以先提交决策后反馈的时序持续适配。本地为外部logits/表示的CPU状态组件，不加载JEV；27B完整实验成本过高，不复现（审查 DONE）。", "available_resource": "One NVIDIA A100 or A30, or local Mac" }, "component_evidence": "docs/system-one/2610.31521-jev-dlc/metrics/component-cpu-seeds42-44.json", "resolution": "bounded_l1_component_accepted", "reason": "冻结源码三种子L1组件验收；CUDA组件真实A100执行。完整实验状态逐篇保留，不宣称模型能力或完整论文复现。" }, { "id": "2610.30675", "key": "onelatent", "domain": "recommendation", "title": "OneLatent: Latent Reasoning for Efficient Foundation Recommendation Models", "paper_url": "https://arxiv.org/abs/2610.30675v1", "detail_path": "recommendation-components/2610.30675-onelatent/README.md", "priority": "P1", "first_author": "Hao Jiang", "first_author_affiliation": "Kuaishou Technology", "verified_priority_organizations": [], "original_published_at": "2026-10-22T03:39:38Z", "intake_class": "new-to-repository-recent-first-publication", "first_publication_after_baseline_capture": true, "observed_versions": [ "2610.30675v1" ], "bulletin_categories": [ "cs.IR" ], "bulletin_entry_kinds": [ "new" ], "status": "deferred", "implementation_status": "deferred_high_cost", "delivery_status": "review_done_not_reproduced", "component_implemented": false, "component_accepted": false, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "requires_gpu_validation": false, "cost_assessment": { "source_url": "https://arxiv.org/html/2610.30675v1", "basis": "named method/resource passages and official author release; not measured complete-run time", "scale": "原文Appendix A.1为56GPU全参数训练、235B教师与工业语料，超单卡A100/A30/Mac低成本预算；不启动完整实验。", "available_resource": "One NVIDIA A100 or A30, or local Mac" } }, { "id": "2610.30844", "key": "rkl-policy-distillation", "domain": "foundation-models", "title": "Reverse-KL Policy Distillation: Statistical Guarantees and Limits", "paper_url": "https://arxiv.org/abs/2610.30844v1", "detail_path": "foundation-models/2610.30844-rkl-policy-distillation/README.md", "priority": "P0", "first_author": "Haolin Liu", "first_author_affiliation": "University of Virginia", "verified_priority_organizations": [ "Google" ], "original_published_at": "2026-10-22T07:19:49Z", "intake_class": "new-to-repository-recent-first-publication", "first_publication_after_baseline_capture": true, "observed_versions": [ "2610.30844v1" ], "bulletin_categories": [ "cs.LG" ], "bulletin_entry_kinds": [ "new" ], "status": "implemented", "implementation_status": "component_only", "delivery_status": "bounded_component_accepted", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_run_scope_limited_protocol", "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "requires_gpu_validation": false, "cost_assessment": { "source_url": "https://arxiv.org/html/2610.30844v1", "basis": "named method/resource passages and official author release; not measured complete-run time", "scale": "Google Research/NTU合作者理论稿；完整保留有限策略class中Algorithm1的整episode采样、仅到访state full teacher反馈、累计KL clip增量、同一clipped prefix概率比作用于candidate/rollout差值、固定eta=1/256指数权更新与最小index破同分。精确有限MDP风险/relative-second-moment审计只在训练后进行；CPU L1诊断，不是LLM蒸馏优化器、无能力或样本复杂度实证宣称。", "available_resource": "One NVIDIA A100 or A30, or local Mac" }, "component_evidence": "docs/foundation-models/2610.30844-rkl-policy-distillation/metrics/component-cpu-seeds42-44.json", "resolution": "bounded_l1_component_accepted", "reason": "冻结源码三种子L1组件验收；CUDA组件真实A100执行。完整实验状态逐篇保留，不宣称模型能力或完整论文复现。" } ], "fixed_component_keys": [ "aopd", "gemma-priority-protocol", "gift", "jev-dlc", "moezip", "openattention", "pair", "rkl-policy-distillation" ], "reference_only_keys": [ "onelatent" ], "selected_scope_complete": true, "official_new_directory_dispositions": [ { "title": "Regional Policies for Memory Tiering Systems", "url": "https://research.google/pubs/regional-policies-for-memory-tiering-systems/", "directory_date": null, "decision": "outside_topic_scope", "review_done": true, "reason": "REGENT is virtual-address-space memory tiering, not an accepted LLM/recommendation/agent algorithm. Official DIMES program lists first author Hayden Coffey at UW-Madison. Exact first public timestamp not established; no Oct23-first-publication claim.", "primary_source": "https://dimes.ws/program/", "first_publication_date": null, "directory_new_url_not_first_publication": true }, { "title": "π-Flux Tuning of Quantum Paramagnetism in the Triangular-Lattice Hubbard Model", "url": "https://research.google/pubs/%CF%80-flux-tuning-of-quantum-paramagnetism-in-the-triangular-lattice-hubbard-model/", "directory_date": null, "decision": "outside_topic_scope", "review_done": true, "reason": "Quantum condensed-matter physics; not in the accepted research tracks.", "first_publication_date": null, "directory_new_url_not_first_publication": true }, { "title": "Stakes-Adaptive Verification and Origin-Traced Rollback in Autonomous Agent Workflows", "url": "https://research.google/pubs/stakes-adaptive-verification-and-origin-traced-rollback-in-autonomous-agent-workflows/", "directory_date": null, "decision": "archived_source_exception", "review_done": true, "full_text_review_done": false, "blocking": false, "automatic_retry": false, "reason": "Official TDC12058 metadata dates the defensive publication to Oct8. Linked full text returned403/HTML, not a valid PDF. Method and experimental claims are not declared fully reviewed and no implementation is invented.", "primary_source": "https://www.tdcommons.org/dpubs_series/12058/", "unavailable_full_text": "https://www.tdcommons.org/cgi/viewcontent.cgi?article=13422&context=dpubs_series", "first_publication_date": "2026-10-08", "exact_first_publication_timestamp": null, "directory_new_url_not_first_publication": true }, { "title": "Words Speak Louder Than Order: A Behavioral Evaluation of Gemma 4", "url": "https://research.google/pubs/words-speak-louder-than-order-a-behavioral-evaluation-of-gemma-4/", "directory_date": null, "decision": "selected_historical_backfill", "key": "gemma-priority-protocol", "review_done": true, "first_publication_date": "2026-09-25", "exact_first_publication_timestamp": "2026-09-25T02:43:40Z", "primary_source": "https://arxiv.org/abs/2609.30716v1", "reason": "Google author affiliation verified in full-text author block; bounded external-prediction protocol, not new model inference.", "directory_new_url_not_first_publication": true }, { "title": "Nature of the bosonic edge phase in a double quantum spin Hall insulator", "url": "https://research.google/pubs/nature-of-the-bosonic-edge-phase-in-a-double-quantum-spin-hall-insulator/", "directory_date": null, "decision": "outside_topic_scope", "review_done": true, "reason": "Quantum condensed-matter physics; not in the accepted research tracks.", "first_publication_date": null, "directory_new_url_not_first_publication": true } ], "additional_dispositions": [ { "id": "2610.30423", "title": "When Agentic Retrieval Fails Silently: Framework and Production Evidence from a Creative AI Assistant", "source_url": "https://arxiv.org/html/2610.30423v1", "decision": "not_selected_under_industrial_evidence_gate", "review_done": true, "automatic_retry": false, "automatic_implementation_queue": false, "reason": "Adobe LARA is an observational production stratification/attribution framework. A controlled quantified online A/B result or qualifying full-traffic comparison was not established; production observation alone does not waive the industrial implementation gate." }, { "id": "2610.30154", "title": "baseVultra: Compact Nonlinear Recurrence with Reallocated Capacity", "source_url": "https://arxiv.org/abs/2610.30154v1", "decision": "archived_identity_date_exception", "review_done": true, "full_text_review_done": false, "blocking": false, "automatic_retry": false, "automatic_implementation_queue": false, "reason": "Observed Oct23 bulletin ID conflicts with the linked v1 page reporting Aug23. Identity/date conflict is archived, not forced into today's first-publication count or a new implementation." }, { "id": "2610.30920", "title": "DConstrain: Compiling Context-Parameterized Transitions for Diffusion Language Model Constrained Decoding", "source_url": "https://arxiv.org/html/2610.30920v1", "decision": "affiliation_unconfirmed", "verified_priority_organizations": null, "full_text_review_done": false, "blocking": false, "automatic_retry": false, "automatic_implementation_queue": false, "reason": "Targeted HTML author block unavailable. Google/Meta collaboration remains unknown, not a negative affiliation finding." }, { "id": "2610.30934", "title": "Verbalizing Activation Differences with Natural Language Autoencoders", "source_url": "https://arxiv.org/html/2610.30934v1", "decision": "affiliation_unconfirmed", "verified_priority_organizations": null, "full_text_review_done": false, "blocking": false, "automatic_retry": false, "automatic_implementation_queue": false, "reason": "Targeted HTML returned author names but no explicit institution. Google/Meta collaboration remains unknown, not a negative affiliation finding." }, { "id": "2610.30880", "title": "One Teacher, Many Views: Reducing Reference Sensitivity in On-Policy Self-Distillation", "source_url": "https://arxiv.org/html/2610.30880v1", "decision": "not_reproduced_high_cost", "status": "deferred", "review_done": true, "original_published_at": "2026-10-22T07:59:48Z", "first_author": "Khouloud Saadi", "first_author_affiliation": "KAUST", "code": null, "code_status": "not_found_in_reviewed_primary_sources", "code_license": null, "automatic_retry": false, "automatic_implementation_queue": false, "cost_assessment": { "source_url": "https://arxiv.org/html/2610.30880v1", "section": "Computational Cost / Table6", "available_resource": "One NVIDIA A100 or A30, or local Mac", "scale": "K=4 partial views; authors report4xA10080GB and1h36 for100steps", "measurement_owner": "authors, not local run" }, "reason": "MCPD probability-mean multiview distillation reviewed; full training exceeds the single-device affordable scope. Review DONE, not reproduced, no automatic retry." }, { "id": "2610.30351", "title": "Route-and-Generate: KL-Regularized Router Improvement for Reasoning Routing in LLMs", "source_url": "https://arxiv.org/html/2610.30351v1", "decision": "not_selected_method_ambiguity", "review_done": false, "blocking": false, "original_published_at": "2026-10-21T19:09:04Z", "first_author_affiliation": "Hanoi University of Science and Technology", "code": null, "code_status": "not_found_in_reviewed_primary_sources", "code_license": null, "automatic_retry": false, "automatic_implementation_queue": false, "reason": "Router closed-form KL mechanism reviewed, but the format-conditioned mean is undefined for n_f=0 and author code was not found. Method ambiguity remains unresolved; no completed faithful reproduction claimed." }, { "id": "2610.30167", "title": "IRAH: Hierarchical Agentic Tree Search for Retrieval Pipeline Optimization", "source_url": "https://arxiv.org/html/2610.30167v1", "decision": "deferred_not_selected", "review_done": false, "blocking": false, "original_published_at": "2026-10-16T17:03:17Z", "first_author_affiliation": "University of Virginia; work during Capital One internship", "code": null, "code_status": "not_verified", "code_license": null, "automatic_retry": false, "automatic_implementation_queue": false, "reason": "Defining method requires LLM reflection mutation, two-stage UCT and offline meta-learned warm start. No heuristic stand-in or unmeasured high-cost claim; not selected in this fixed batch." }, { "id": "2610.31401", "title": "HERD: Harness-Evolving Recursive Distillation for Continual Learning", "source_url": "https://arxiv.org/html/2610.31401v1", "decision": "source_unresolved_not_selected", "review_done": false, "blocking": false, "original_published_at": "2026-10-22T15:03:20Z", "first_author_affiliation": "Monash University", "code": "https://github.com/a1exsun/HERD", "observed_main_revision_not_paper_revision": "3ff1cf03d7a34565bcdbd510fae09aed7971286a", "paper_code_branch": "arr", "paper_code_revision": null, "code_license": null, "automatic_retry": false, "automatic_implementation_queue": false, "cost_assessment": { "source_url": "https://arxiv.org/html/2610.31401v1", "scale": "Qwen3.5-4B,25stages,32x32records per stage; authors report10723training seconds", "measurement_owner": "authors, not local run", "available_resource": "One NVIDIA A100 or A30, or local Mac" }, "reason": "Main README says paper code is on arr branch, which was not retrieved. Main commit is not relabelled the paper revision; no implementation or full-source closure claimed." } ], "component_count": 8, "gpu_validated_component_count": 5, "capability_reproductions_claimed": 0 }