{ "schema_version": 1, "reviewed_at": "2026-10-12", "assignment_path": ".cache/oct12-scan/fulltext/post-training-assignment.json", "review_scope": "逐条摘要 + 官方原文限定方法/评测片段；非逐篇全段落阅读全文，非新的穷尽检索声明", "frozen_implementation_scope_preserved": true, "summary": { "assignment_records": 105, "abstracts_reviewed": 105, "bounded_primary_fulltext_reviews": 105, "full_articles_read": 0, "verified_html_sha256": 103, "html_sha256_mismatches": 0, "accepted_existing": 9, "reviewed_deferred": 90, "rejected_or_out_of_scope": 6, "next_batch_candidates_after_bounded_review": 90, "fulltext_unresolved": 0, "new_implementation_claims": 0, "priority_affiliation_reviews": [ "2610.13617", "2610.13058", "2610.09384" ], "confirmed_google_research_candidate": "2610.13617", "author_now_at_google_is_not_institutional_attribution": "2610.09384", "unfinished": [], "verified_pdf_sha256": 2 }, "records": [ { "arxiv_id": "2610.13755", "title": "Matched-Context Group Policy Optimization for LLM Agent Training", "published": "2026-10-09T15:41:37Z", "status": "reviewed_deferred", "decision_rationale": "MCGPO 用固定 sibling folds、state/rubric 匹配和反折 terminal evidence 构造 pairwise preference，3-seed ALFWorld/WebShop 控制支持后训练机制；冻结16篇之外，下一批需同状态/进度和label隔离实现。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.13755v1", "sha256": "d68cc32d561fc07bb10a04b38a1fa7c443b1a50663f5dd3fdc8d200cba91d188", "local_html_sha256": "d68cc32d561fc07bb10a04b38a1fa7c443b1a50663f5dd3fdc8d200cba91d188", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.13755v1", "result": "retrieved", "sha256": "d68cc32d561fc07bb10a04b38a1fa7c443b1a50663f5dd3fdc8d200cba91d188" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-agent" ], "tracks": [ "agent", "foundation-model", "post-training" ] }, { "arxiv_id": "2610.13330", "title": "Masked Diffusion Language Agents for Tool-Integrated Chemical Reasoning", "published": "2026-10-09T10:07:37Z", "status": "reviewed_deferred", "decision_rationale": "ChemDiffAgent 对交织tool trajectory做masked denoising，agentic SFT+variance-reduced preference optimization；6类化学任务有匹配预算评测，核心需要实际diffusion backbone和化学tool graph，不能普通AR proxy。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.13330v1", "sha256": "2a9cbaa7ce1108b2cf91a4a5fa3c4a7bf90af533a17fa9637a86950d224dd05b", "local_html_sha256": "2a9cbaa7ce1108b2cf91a4a5fa3c4a7bf90af533a17fa9637a86950d224dd05b", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.13330v1", "result": "retrieved", "sha256": "2a9cbaa7ce1108b2cf91a4a5fa3c4a7bf90af533a17fa9637a86950d224dd05b" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "4 ChemDiffAgent", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "5 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/bilemond/ChemDiffAgent" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "agent-planning", "tool-agent" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.12601", "title": "Reflective Prompt Optimization of a Deep-Research Agent: A Topic-Disjoint, Multi-Seed Protocol for Measuring Transferable Gains and Prompt-Level Overfitting", "published": "2026-10-08T18:23:23Z", "status": "reviewed_deferred", "decision_rationale": "GEPA提示进化的topic-disjoint 500-task评测和verifier ablation可用于评测完整性；属于冻结模型prompt/agent优化，下一批转agent/evaluation，并需隔离winning prompt内answer-key风险。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12601v1", "sha256": "b639966ea3e1b51667cf8817044471892cfb515025d18323caa410f22d74cd69", "local_html_sha256": "b639966ea3e1b51667cf8817044471892cfb515025d18323caa410f22d74cd69", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12601v1", "result": "retrieved", "sha256": "b639966ea3e1b51667cf8817044471892cfb515025d18323caa410f22d74cd69" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/gepa-ai/gepa" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "multi-agent" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.12341", "title": "Can AI Agents Learn Their Way to the Top? Evaluating Heuristic Learning in a Long-Running Game Agent Competition", "published": "2026-10-08T17:15:48Z", "status": "reviewed_deferred", "decision_rationale": "AHL在固定模型权重下修改可执行game policies，AAArena人类ladder预算评测；主要是agent程序优化而非参数后训练，转agent track下一批，不能登记LLM权重训练复现。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12341v1", "sha256": "b41615e1af7afaeb2188eca093e6b300a43e3c83f393407abbd5d3b82f7c438e", "local_html_sha256": "b41615e1af7afaeb2188eca093e6b300a43e3c83f393407abbd5d3b82f7c438e", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12341v1", "result": "retrieved", "sha256": "b41615e1af7afaeb2188eca093e6b300a43e3c83f393407abbd5d3b82f7c438e" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 Adversarial Heuristic Learning", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/THU-CST-SAST/AAArena" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "agentic-rl", "software-agent" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.12061", "title": "When Should Agents Think? Adaptive Reasoning via Cross-Turn Estimation", "published": "2026-10-08T14:43:09Z", "status": "accepted", "decision_rationale": "RACE/LoGiC通过移除reasoning后的reference-action likelihood检测跨turn支持，结合SFT/RL；已有implemented条目，本次保留并核对原文，不重复宣称新实现。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12061v1", "sha256": "077e07c12d1ac7fc5d819c34ca359357e34ba2e74887cc097ac40206119b7ba6", "local_html_sha256": "077e07c12d1ac7fc5d819c34ca359357e34ba2e74887cc097ac40206119b7ba6", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12061v1", "result": "retrieved", "sha256": "077e07c12d1ac7fc5d819c34ca359357e34ba2e74887cc097ac40206119b7ba6" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/HututuAI/RACE" ], "repository_status_at_assignment": "implemented", "new_implementation_claimed": false, "next_action": "保留已有登记；不重复实施或补称本次 GPU 通过", "matched_queries": [ "agentic-rl", "llm-agent" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.11959", "title": "MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement", "published": "2026-10-08T13:41:53Z", "status": "reviewed_deferred", "decision_rationale": "MiMo-V2.6真实大规模agentic RL：1568 prompts×16 rollouts，asynchronous GRPO/staleness4、frozen MoE router及groupwise graders；公开训练日志/环境链路，资源规模远超冻结批次，下一批需要分解可执行机制。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11959v1", "sha256": "b07bc0b1e19ddc8bd6d8753fadbf9b2c5c3eb7899f7925650f1ff1e81df963aa", "local_html_sha256": "b07bc0b1e19ddc8bd6d8753fadbf9b2c5c3eb7899f7925650f1ff1e81df963aa", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11959v1", "result": "retrieved", "sha256": "b07bc0b1e19ddc8bd6d8753fadbf9b2c5c3eb7899f7925650f1ff1e81df963aa" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "4 Scaling Reinforcement Learning", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "5 Experiments: You Only RL Once", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/datacurve-ai/deep-swe", "https://github.com/harbor-framework/terminal-bench/releases/tag/v4.0.0", "https://github.com/vllm-project/humming/releases/tag/v0.1.15", "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B", "https://huggingface.co/collections/PrimeIntellect/swe-rl" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "multi-agent" ], "tracks": [ "agent", "foundation-model", "post-training" ] }, { "arxiv_id": "2610.11678", "title": "TRACE: Diagnosing Verifier Brittleness in Agentic Evaluation", "published": "2026-10-08T10:52:26Z", "status": "reviewed_deferred", "decision_rationale": "TRACE区分行为变化与verifier变化，paired mutation和same-trajectory rescoring在tau2-bench有重复/positive controls；synthetic agents可读intended fix只属diagnostic，下一批转evaluation integrity，不当capability implementation。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11678v1", "sha256": "b642aed3ab93ab210b67c2973285657af44253eb668eda4e4fb39f47d99aed86", "local_html_sha256": "b642aed3ab93ab210b67c2973285657af44253eb668eda4e4fb39f47d99aed86", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11678v1", "result": "retrieved", "sha256": "b642aed3ab93ab210b67c2973285657af44253eb668eda4e4fb39f47d99aed86" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 TRACE: From Score Differences to Explanations", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Study 2: LLM Agents on τ 2 \\tau^{2} -bench", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/RGaonkar/trace-verifier-stress-tests", "https://github.com/sierra-research/tau2-bench" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-agent" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.11464", "title": "Who Verifies the Verifier? Co-Evolving Inspectable Graders with Self-Improving Agents", "published": "2026-10-08T08:13:53Z", "status": "reviewed_deferred", "decision_rationale": "演化inspectable pass/fail/abstain grader以10-item dev anchors选择，test locked，MBPP+/Spider2/report workloads；适合verifier/skill co-evolution下一批，必须验证grader validity，downstream task lift不能替代。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11464v1", "sha256": "db45d964d77020219f78316da71149f8a3bf31157e5fb69f7f96dc75679e2302", "local_html_sha256": "db45d964d77020219f78316da71149f8a3bf31157e5fb69f7f96dc75679e2302", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11464v1", "result": "retrieved", "sha256": "db45d964d77020219f78316da71149f8a3bf31157e5fb69f7f96dc75679e2302" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 The verifier and its evolution", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experimental setup", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-agent" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.11451", "title": "SAIL: Scientific Agentic Intelligence via a Science-Aware Loop", "published": "2026-10-08T08:04:22Z", "status": "reviewed_deferred", "decision_rationale": "SAIL science-aware failure diagnosis/data construction循环，SFT→specialists→MOPD→agentic RL；AstaBench/SciCode/DRB2评测，science benchmarks与development隔离。需真实任务环境及35B backbone，冻结批次外待下一批。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11451v1", "sha256": "d9e33a3b9f1ad0b2d1d9271aef70de82ca6a4c50d74db115034d8d434230538b", "local_html_sha256": "d9e33a3b9f1ad0b2d1d9271aef70de82ca6a4c50d74db115034d8d434230538b", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11451v1", "result": "retrieved", "sha256": "d9e33a3b9f1ad0b2d1d9271aef70de82ca6a4c50d74db115034d8d434230538b" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 Science-Aware Improvement Loop", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "3 Training Recipe", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "5 Evaluation and Analysis", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://huggingface.co/Qwen/Qwen3.6-35B-A3B", "https://huggingface.co/XiaomiMiMo/MiMo-V2.5", "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro", "https://huggingface.co/apodex/Apodex-1.0-mini", "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731", "https://huggingface.co/endless-frontier/BigBang-v1", "https://huggingface.co/inclusionAI/Ling-3.0-flash", "https://huggingface.co/inclusionAI/Ring-2.6-1T", "https://huggingface.co/meituan-longcat/LongCat-2.0-FP8", "https://huggingface.co/nex-agi/Nex-N2-mini", "https://huggingface.co/nex-agi/Nex-N2.5-mini", "https://huggingface.co/tencent/Hy3", "https://huggingface.co/zai-org/GLM-5.2" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "agent-planning" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.11373", "title": "From a Prompt to Repertoires: Evolving Functional REpertoires Enable LLM Continual Learning", "published": "2026-10-08T07:06:52Z", "status": "reviewed_deferred", "decision_rationale": "EFRE用历史一致性检测兼容更新或新增函数，在固定模型下external repertoire适应ToolUse→FinQA→SciKE-Bio；可作为agent continual learning候选，非权重后训练，下一批需要实际函数执行与遗忘评测。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11373v1", "sha256": "e4dd3319b80efaf58e3a7e9422c3495991171f875e201de48e9122965aafa5ef", "local_html_sha256": "e4dd3319b80efaf58e3a7e9422c3495991171f875e201de48e9122965aafa5ef", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11373v1", "result": "retrieved", "sha256": "e4dd3319b80efaf58e3a7e9422c3495991171f875e201de48e9122965aafa5ef" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://huggingface.co/Qwen/Qwen3-8B", "https://huggingface.co/Qwen/Qwen3-Embedding-0.6B" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "agentic-rl", "llm-agent" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.11183", "title": "RAG-Stress: Probing the Limits of Evidence Reliance in Retrieval-Augmented Generation", "published": "2026-10-08T03:40:48Z", "status": "reviewed_deferred", "decision_rationale": "RAG-Stress对同question/reference的单断言干预，baseline-correct subset测misleading rate；这是证据依赖诊断，下一批转evaluation，不作后训练改进算法或能力实施证据。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11183v1", "sha256": "c707584dc1f93b78923e66dbdb5e8e69e0b01159473db66447a2b581ca902cb2", "local_html_sha256": "c707584dc1f93b78923e66dbdb5e8e69e0b01159473db66447a2b581ca902cb2", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11183v1", "result": "retrieved", "sha256": "c707584dc1f93b78923e66dbdb5e8e69e0b01159473db66447a2b581ca902cb2" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 RAG-Stress Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "agentic-rl" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.11152", "title": "Do LLMs Learn from Rewards in Context? : Rethinking the role of reward in In-Context Reinforcement Learning", "published": "2026-10-08T03:15:00Z", "status": "reviewed_deferred", "decision_rationale": "direct ICRL通过reward flip/random/remove和trajectory腐化作对照，ScienceWorld/MBPP+/HumanEval+显示提升不由reward语义驱动；适合agent memory评测下一批，固定模型推理时学习不能称参数RL复现。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11152v1", "sha256": "9db3a6e41dd61d84d96ef53760602b4673b9318057a9eeeff20b7263cbba3bf5", "local_html_sha256": "9db3a6e41dd61d84d96ef53760602b4673b9318057a9eeeff20b7263cbba3bf5", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11152v1", "result": "retrieved", "sha256": "9db3a6e41dd61d84d96ef53760602b4673b9318057a9eeeff20b7263cbba3bf5" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Preliminaries", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Reward Semantics Do Not Drive Improvement", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "agent-memory", "agentic-rl", "llm-agent" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.10942", "title": "StoreBench: A Live-Commerce Environment for Evaluating and Training Autonomous Operator Agents", "published": "2026-10-07T21:46:10Z", "status": "reviewed_deferred", "decision_rationale": "StoreBench Medusa live-commerce动态环境、action-driven时间、anchor-calibrated reward和GRPO训练有heldout split；完整环境/eval被保留未公开，不能承诺完整复现，下一批需确认公开sample与scorer可执行范围。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10942v1", "sha256": "ff478d4dbdf2da211ee9d046ef0091a4ddd7933e7200be38d65cd84fc7382ea7", "local_html_sha256": "ff478d4dbdf2da211ee9d046ef0091a4ddd7933e7200be38d65cd84fc7382ea7", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10942v1", "result": "retrieved", "sha256": "ff478d4dbdf2da211ee9d046ef0091a4ddd7933e7200be38d65cd84fc7382ea7" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 The StoreBench environment", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "5 Experimental setup", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/medusajs/medusa", "https://github.com/surge-ai/dayjob" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "agent-planning", "agentic-rl", "llm-agent", "tool-agent" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.10878", "title": "Stochastic Teacher Intervention for Agentic On-Policy Distillation", "published": "2026-10-07T20:29:29Z", "status": "reviewed_deferred", "decision_rationale": "STI-OPD用k3师生KL映射随机teacher intervention，mixed policy token须importance-weight reverse KL校正；工具推理/ALFWorld/ScienceWorld多benchmark，下一批必须实际环境执行teacher替换，不能仅loss proxy。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10878v1", "sha256": "a7bd7272519466f5a2080b196019a07433b2bcc9d482616930f25be4105a88fd", "local_html_sha256": "a7bd7272519466f5a2080b196019a07433b2bcc9d482616930f25be4105a88fd", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10878v1", "result": "retrieved", "sha256": "a7bd7272519466f5a2080b196019a07433b2bcc9d482616930f25be4105a88fd" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "tool-agent" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.10787", "title": "NavGPT-3: Harnessing Context in a Hierarchical Navigation Runtime", "published": "2026-10-07T18:45:56Z", "status": "reviewed_deferred", "decision_rationale": "NavGPT-3分离reasoning/acting/monitoring线程，VLA codec视觉token分配及embodied harness有R2R-CE/RxR-CE及真robot实验；主要embodied foundation/agent系统，下一批需模型工具权限与动作runtime，非本批posttraining operator。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10787v1", "sha256": "14eb5d50fb5286992716de3db7a06b598df7c7c00e98f3cb9e28f829a51dc404", "local_html_sha256": "14eb5d50fb5286992716de3db7a06b598df7c7c00e98f3cb9e28f829a51dc404", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10787v1", "result": "retrieved", "sha256": "14eb5d50fb5286992716de3db7a06b598df7c7c00e98f3cb9e28f829a51dc404" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method: Harness, Runtime, and VLA Co-Design", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/OpenDriveLab/OpenScene", "https://github.com/deepseek-ai/deepseek-harness" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "agentic-rl", "tool-agent" ], "tracks": [ "agent", "foundation-model", "post-training" ] }, { "arxiv_id": "2610.10407", "title": "SOTA: Stock Options Trading Agents Guided by Option-Implied Return Distributions", "published": "2026-10-07T16:55:05Z", "status": "reviewed_deferred", "decision_rationale": "SOTA SFT→RL训练strategy-level期权选择，确定resolver执行合约/仓位，2024训练与2025 out-of-sample时间隔离明确；交易数据部分为收费源，财务结果不可泛化为投资建议，冻结批次外下一批先解决数据可得性。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10407v1", "sha256": "2053c3a903955b1962456713e90cc5b9d270b92f0b70a2d70684b11e38251fe2", "local_html_sha256": "2053c3a903955b1962456713e90cc5b9d270b92f0b70a2d70684b11e38251fe2", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10407v1", "result": "retrieved", "sha256": "2053c3a903955b1962456713e90cc5b9d270b92f0b70a2d70684b11e38251fe2" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 SOTA: Structured Strategy Selection for Stock Option Trading Agents", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "agentic-rl" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.10179", "title": "Beyond Outcome Rewards: Constructing and Assigning Retrieval Credit for Search Agents", "published": "2026-10-07T14:48:47Z", "status": "reviewed_deferred", "decision_rationale": "Search-agent intermediate retrieval progress credit比较scalar/local归属，固定E5/KILT retriever、14k prompts、5 rollouts同预算GRPO控制；核心检索credit需真实query/evidence事件，下一批研究。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10179v1", "sha256": "9a693b49ea1ec48fc6772e5e4e153fa034ec2f4d09f19d2adcc455026f378ad2", "local_html_sha256": "9a693b49ea1ec48fc6772e5e4e153fa034ec2f4d09f19d2adcc455026f378ad2", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10179v1", "result": "retrieved", "sha256": "9a693b49ea1ec48fc6772e5e4e153fa034ec2f4d09f19d2adcc455026f378ad2" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Assigning retrieval credit to policy updates", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experimental design and results", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "agentic-rl", "llm-agent", "multi-agent" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.09858", "title": "Training Advisors for LLM Agents from Task Outcomes", "published": "2026-10-07T11:14:26Z", "status": "reviewed_deferred", "decision_rationale": "Caddie冻结base agent、用最终outcome RL训练自然语言critic，MuSiQue/ALFWorld训练并跨base和DeepDive/tau3迁移；下一批需实际critic intervention和continuation outcomes，不能reference critique fixture。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09858v1", "sha256": "e2c933d0175bd4636ad8e2512000b7f0a217990953aee9a125b2523471d7ec42", "local_html_sha256": "e2c933d0175bd4636ad8e2512000b7f0a217990953aee9a125b2523471d7ec42", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09858v1", "result": "retrieved", "sha256": "e2c933d0175bd4636ad8e2512000b7f0a217990953aee9a125b2523471d7ec42" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "3 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/sierra-research/tau2-bench", "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507", "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507", "https://huggingface.co/Qwen/Qwen3.8-27B", "https://huggingface.co/moonshotai/Kimi-K3" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "agentic-rl", "llm-agent", "multi-agent", "tool-agent" ], "tracks": [ "agent", "foundation-model", "post-training" ] }, { "arxiv_id": "2610.09074", "title": "TAP: Efficient Long-Horizon Agent Pruning via Trajectory-Anchored Recovery", "published": "2026-10-06T20:18:17Z", "status": "reviewed_deferred", "decision_rationale": "TAP iterative FFN structural pruning结合teacher-trajectory anchored student on-policy responses，恢复loss梯度重评channel；ALFWorld/WebShop pruning/latency评测，需真实剪除并恢复train/infer，下一批。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09074v1", "sha256": "b639a1033aef4c92200424599dcb57965ed8d85701a2170cb612540caacad1a5", "local_html_sha256": "b639a1033aef4c92200424599dcb57965ed8d85701a2170cb612540caacad1a5", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09074v1", "result": "retrieved", "sha256": "b639a1033aef4c92200424599dcb57965ed8d85701a2170cb612540caacad1a5" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "agentic-rl" ], "tracks": [ "agent", "foundation-model", "post-training" ] }, { "arxiv_id": "2610.08963", "title": "On KL-Regularized Policy Optimization", "published": "2026-10-06T18:30:42Z", "status": "reviewed_deferred", "decision_rationale": "官方PDF已取得并限定阅读方法/结论。KLPO以sampler为KL锚点，将Gibbs最优性转为log-ratio最小二乘；profile intercept需要真实sampler-to-trainer KL，独立MC-KL/交叉拟合防止偏差，不能以同一动作或截断KL冒充无偏更新。结论明确大规模实证为后续工作，保留理论/后训练机制下一批，非现成已验证LLM能力收益。", "priority_reviewed": false, "verified_organization": "Princeton University", "source": { "url": "https://arxiv.org/html/2610.08963v1", "sha256": null, "local_html_sha256": null, "local_html_sha256_matches": null, "attempts": [ { "url": "https://arxiv.org/html/2610.08963v1", "result": "failed", "error": "404 Client Error: Not Found for url: https://arxiv.org/html/2610.08963v1" }, { "url": "https://arxiv.org/html/2610.08963v1", "result": "failed", "error": "404 Client Error: Not Found for url: https://arxiv.org/html/2610.08963v1" }, { "url": "https://arxiv.org/pdf/2610.08963v1", "result": "retrieved", "retrieved_on": "2026-10-12", "command": "curl -sSL --fail --max-time 45 https://arxiv.org/pdf/2610.08963v1 -o tmp/pdfs/2610.08963.pdf", "sha256": "de3e299d401af2715a49090328f86a8e4ba284983e189d8119912f4a573f4590" } ], "alternate_evidence": { "url": "https://arxiv.org/pdf/2610.08963v1", "format": "official_arxiv_pdf", "sha256": "de3e299d401af2715a49090328f86a8e4ba284983e189d8119912f4a573f4590", "local_path": "tmp/pdfs/2610.08963.pdf", "local_sha256_verified": true, "total_pdf_pages": 53 } }, "read_scope": "abstract_plus_bounded_official_pdf_method_and_evaluation_pages", "full_article_read": false, "evidence_locations": [ { "heading": "Title/author and arXiv v1 margin", "pdf_page": 1, "scope": "full page text read; manuscript cover date is not substituted for arXiv v1 date" }, { "heading": "Method and evaluation/conclusion", "pdf_pages": [ 5, 6, 8, 9, 11, 13, 14, 15 ], "scope": "listed page text read; not all article pages or appendices", "method_evidence": "sampler version-specific probabilities/conditionals；Gibbs log-ratio residual；profiled intercept = sampler signal mean + sampler-to-trainer KL；critic-free token PMD终端return；独立辅助抽样MC-KL；序列unbiased estimator需要crossfit/U-statistic至少两个辅助draw，保留完整非过滤rollout且token求和、response求均值。", "evaluation_evidence": "PDF第15页结论将large-scale empirical evaluation、finite-sample/convergence analysis列为next steps；本次相关页没有可用于宣称本地或原文LLM benchmark增益的实验数值。未通读53页所有附录，不作整篇零实验断言。" } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/yifanzhang-pro/KLPO" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结16篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "agentic-rl", "llm-agent", "tool-agent" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.08959", "title": "GraphOPD: Graph-Augmented On-Policy Distillation for LLM Agents", "published": "2026-10-06T18:25:23Z", "status": "reviewed_deferred", "decision_rationale": "GraphOPD用environment entity Footprints构dependency graph，stationary random walk结构credit与divergence融合轨迹相对mask；ALFWorld/WebShop/SearchQA有共享GRPO与executed-replay audit，下一批必须真实footprints/因果回放而非gold plan。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.08959v1", "sha256": "5395b73753261a0ae4f9f6ad90d1788fc0f2fa114d8054f5de172934ff6a2619", "local_html_sha256": "5395b73753261a0ae4f9f6ad90d1788fc0f2fa114d8054f5de172934ff6a2619", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.08959v1", "result": "retrieved", "sha256": "5395b73753261a0ae4f9f6ad90d1788fc0f2fa114d8054f5de172934ff6a2619" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://anonymous.4open.science/status/GraphOPD-5011" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "agentic-rl", "llm-agent" ], "tracks": [ "agent", "post-training" ] }, { "arxiv_id": "2610.13485", "title": "Proactive Agentic Orchestration", "published": "2026-10-09T12:18:49Z", "status": "reviewed_deferred", "decision_rationale": "POrchestra允许subagent主动申请harness变化并保留局部执行，OPOD从hindsight request回归orchestrator决定作过程监督；GAIA/动态数字助手/软件任务，转agent orchestrator下一批需真实控制与训练而非静态标签。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.13485v1", "sha256": "9649e66ff2a76853ae90b9e13560b1e1271b96ee33f6a0e6289df5a16e850df3", "local_html_sha256": "9649e66ff2a76853ae90b9e13560b1e1271b96ee33f6a0e6289df5a16e850df3", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.13485v1", "result": "retrieved", "sha256": "9649e66ff2a76853ae90b9e13560b1e1271b96ee33f6a0e6289df5a16e850df3" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/XueyangFeng/POrchestra", "https://github.com/huggingface/trl" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "long-context" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.13202", "title": "S$^3$R: Fast and Faithful RL Rollouts via Self-Sparse Speculative Decoding", "published": "2026-10-09T08:22:40Z", "status": "reviewed_deferred", "decision_rationale": "S3R同参数sparse drafting+dense exact speculative verification，保留完整dense group分布及KV-aware scheduler；8H100 DAPO math GRPO throughput对照，下一批需实际dense correction和GPU latency/分布验证。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.13202v1", "sha256": "2c4704fec367030df4d39d71da112f0f5dd160df8bf803e558a25adf417c33e1", "local_html_sha256": "2c4704fec367030df4d39d71da112f0f5dd160df8bf803e558a25adf417c33e1", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.13202v1", "result": "retrieved", "sha256": "2c4704fec367030df4d39d71da112f0f5dd160df8bf803e558a25adf417c33e1" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Problem Formulation and Design Principles", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "6 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/apoorvumang/prompt-lookup-decoding/", "https://huggingface.co/AngelSlim/Qwen3-1.7B_eagle3", "https://huggingface.co/AngelSlim/Qwen3-8B_eagle3", "https://huggingface.co/datasets/MathArena/aime_2026", "https://huggingface.co/datasets/MathArena/arxivmath" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "long-context" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.13035", "title": "GeoGR: Spatial-Semantic Token Compression and Group Routing for Efficient 3D VLMs", "published": "2026-10-09T05:46:54Z", "status": "reviewed_deferred", "decision_rationale": "GeoGR voxel GeoSemZip与LLM recoverable GroupRoute结合compression-aware 67k posttraining，ScanNet五benchmark/真实prefill评测；foundation3D VLM下一批需RGB/depth/pose和可恢复token状态，不用文本mask proxy。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.13035v1", "sha256": "8c0d8440bbe5b6cc81360366e17d139bd3312b70d0680b777a432e06a369fcb6", "local_html_sha256": "8c0d8440bbe5b6cc81360366e17d139bd3312b70d0680b777a432e06a369fcb6", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.13035v1", "result": "retrieved", "sha256": "8c0d8440bbe5b6cc81360366e17d139bd3312b70d0680b777a432e06a369fcb6" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "4 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "5 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/EIT-NLP/GeoGR" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "efficient-inference", "model-compression", "vision-language" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.12995", "title": "What Should SFT Learn? Conditional Information Gain for Generalizable Post-Training", "published": "2026-10-09T04:57:55Z", "status": "reviewed_deferred", "decision_rationale": "CIG-SFT frozen reference有/无task条件likelihood在gold训练prefix上估计conditional information gain，用KL projection等价weighted SFT；math/code公开任务与baseline，离线train supervision合法但gold不得进入被评估rollout，下一批。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12995v1", "sha256": "2a7ba2a327ec42e00bb6dc364e69b71fad8d09a7b773b921b72d262c21b0d374", "local_html_sha256": "2a7ba2a327ec42e00bb6dc364e69b71fad8d09a7b773b921b72d262c21b0d374", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12995v1", "result": "retrieved", "sha256": "2a7ba2a327ec42e00bb6dc364e69b71fad8d09a7b773b921b72d262c21b0d374" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/NiuKequan/CIG-SFT-public" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "pretraining-data" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.12707", "title": "Unsupervised Multimodal Reward Modeling", "published": "2026-10-08T20:38:25Z", "status": "reviewed_deferred", "decision_rationale": "冻结Qwen3-VL intermediate-layer linear head用same-image caption perturbations自监督，标签只用于layer选择；reward benchmark与verifier-driven bootstrap downstream不等价，下一批需train/val选择层隔离和真实multimodal heads。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12707v1", "sha256": "bf8fe4d9ecf041ecb1a8efb2caee64631527f1aa5627e740cfaeca358e8e6eb8", "local_html_sha256": "bf8fe4d9ecf041ecb1a8efb2caee64631527f1aa5627e740cfaeca358e8e6eb8", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12707v1", "result": "retrieved", "sha256": "bf8fe4d9ecf041ecb1a8efb2caee64631527f1aa5627e740cfaeca358e8e6eb8" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "3 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "multimodal-llm", "vision-language" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.12582", "title": "One Model to Loop Them All: Elastic Latent Recurrence for Language Models", "published": "2026-10-08T18:05:42Z", "status": "reviewed_deferred", "decision_rationale": "ELR实际共享latent block、learned gated step、loop/budget conditioning与每update mixed-depth训练；4B dense/30B-A3B评测及route reuse/KV tradeoff，foundation recurrence下一批，不能只插循环标签冒充模型训练。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12582v1", "sha256": "299b298ab2cc380e7ee9c422e1f7d74ac3823b7643a84481ec681ff41fb87d16", "local_html_sha256": "299b298ab2cc380e7ee9c422e1f7d74ac3823b7643a84481ec681ff41fb87d16", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12582v1", "result": "retrieved", "sha256": "299b298ab2cc380e7ee9c422e1f7d74ac3823b7643a84481ec681ff41fb87d16" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 One-Loop Accuracy and Gains from Additional Loops", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://huggingface.co/collections/nvidia/nemotron-post-training-v3" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "pretraining-data" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.12248", "title": "EgoVoice: Proactive Spoken Assistance from Egocentric Multimodal Streams", "published": "2026-10-08T16:24:12Z", "status": "reviewed_deferred", "decision_rationale": "EgoVoice用HoloAssist source-separated/resynth audio和listen-or-speak SFT/DPO训练主动语音助手，170 sessions onset/content/end评测；多模态/语音后训练下一批须实际流式音视频与timing policy。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12248v1", "sha256": "739fa088b70399874fa19e26e6d2ef62271c43adc60495b02c979cdeb5d7298a", "local_html_sha256": "739fa088b70399874fa19e26e6d2ef62271c43adc60495b02c979cdeb5d7298a", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12248v1", "result": "retrieved", "sha256": "739fa088b70399874fa19e26e6d2ef62271c43adc60495b02c979cdeb5d7298a" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/snakers4/silero-vad", "https://huggingface.co/594zyc/ProAssist-Model-L4096-I10", "https://huggingface.co/IndexTeam/IndexTTS-2", "https://huggingface.co/Qwen/Qwen3-ASR-1.7B", "https://huggingface.co/Qwen/Qwen3-Embedding-0.6B", "https://huggingface.co/Qwen/Qwen3-ForcedAligner-0.6B", "https://huggingface.co/facebook/sam-audio-large-tv", "https://huggingface.co/openbmb/MiniCPM-o-4_5" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "multimodal-llm", "vision-language" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.12235", "title": "Language Models as AI Research World Models", "published": "2026-10-08T16:18:54Z", "status": "reviewed_deferred", "decision_rationale": "Research World Models用真实实验记录预测unseen interventions并按预算选实验，九个pre/post/inference环境；适合auto-research controller下一批，必须保持outcome heldout、真实执行再验证，预测不能当已完成实验。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12235v1", "sha256": "d1d254f14a7c1b85aeddbc1a5261848692d0d8ca16eea25086b1d061ddebad24", "local_html_sha256": "d1d254f14a7c1b85aeddbc1a5261848692d0d8ca16eea25086b1d061ddebad24", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12235v1", "result": "retrieved", "sha256": "d1d254f14a7c1b85aeddbc1a5261848692d0d8ca16eea25086b1d061ddebad24" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 Language Models as Research World Models", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "3 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/allenai/OLMo-core", "https://github.com/huggingface/transformers/tree/main/src/transformers/models/qwen3", "https://github.com/karpathy/autoresearch", "https://github.com/karpathy/nanochat", "https://github.com/marin-community/marin" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "pretraining-data" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.11543", "title": "Learning the Loop, Not Just the Page: Execution-Grounded Loop Learning for Web Generation", "published": "2026-10-08T09:11:17Z", "status": "reviewed_deferred", "decision_rationale": "WebLoop共享Generator/Critic/Refiner通过browser execution rewards学习loop，critic先diagnostic后downstream helpfulness信用；WebRise/WebGen-Bench空workspace控制，下一批需真实网页执行与Critic训练。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11543v1", "sha256": "54e4531c7ca2ddb678f6a5a45e2bd24b369b59c7694352d1f97627ccba18e4a4", "local_html_sha256": "54e4531c7ca2ddb678f6a5a45e2bd24b369b59c7694352d1f97627ccba18e4a4", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11543v1", "result": "retrieved", "sha256": "54e4531c7ca2ddb678f6a5a45e2bd24b369b59c7694352d1f97627ccba18e4a4" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 WebLoop: Learning the Generation–Critique–Refinement Loop", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://huggingface.co/moonshotai/Kimi-K2.6", "https://huggingface.co/tesslate/UIGEN-T2", "mailto:1@l@[][c]https://github.com/IIGROUP/WebLoop" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "multimodal-llm" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.11461", "title": "Beyond Speech Captions: Speech-Rewarded Style Planning for Conversational Text-to-Speech", "published": "2026-10-08T08:12:28Z", "status": "reviewed_deferred", "decision_rationale": "SRSP冻结TTS，以teacher-forced target speech-token likelihood作GRPO style-planner训练reward，推理只dialogue/text；scene/movie-disjoint dev/test划分明确，下一批需真实声学模型与语音评测。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11461v1", "sha256": "4d82b63ea2353f5f8d7151e291b4730899b3d373d8e8f20e52ef48498b683050", "local_html_sha256": "4d82b63ea2353f5f8d7151e291b4730899b3d373d8e8f20e52ef48498b683050", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11461v1", "result": "retrieved", "sha256": "4d82b63ea2353f5f8d7151e291b4730899b3d373d8e8f20e52ef48498b683050" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Speech-Rewarded Style Planning", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/huggingface/trl", "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct", "https://huggingface.co/nvidia/audio-flamingo-next-captioner-hf" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "model-compression" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.11371", "title": "SignRAG: Unified Retrieval-Augmented Gloss-Free Sign Language Translation", "published": "2026-10-08T07:03:24Z", "status": "reviewed_deferred", "decision_rationale": "SignRAG hierarchical sign/LLM alignment + train-only retrieval gallery + quality/utility rewards GRPO，四SLT公开split评测；下一批需实际sign encoder、retrieval与utility变化，不能gold translation进test检索。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11371v1", "sha256": "4b2b208b797987fbb34f71f87aef62c307f6bce1ba1c27f44b5bc5e2c73bf618", "local_html_sha256": "4b2b208b797987fbb34f71f87aef62c307f6bce1ba1c27f44b5bc5e2c73bf618", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11371v1", "result": "retrieved", "sha256": "4b2b208b797987fbb34f71f87aef62c307f6bce1ba1c27f44b5bc5e2c73bf618" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiment", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/open-mmlab/mmpose", "https://github.com/shahelaojieraozhi/SignRAG" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "vision-language" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.11354", "title": "AdaptEvo: Adaptive Agent Learning with Evolving Supervision", "published": "2026-10-08T06:47:41Z", "status": "reviewed_deferred", "decision_rationale": "AdaptEvo CA-GRPO按reference confidence混合outcome/process，并演化decision knowledge/rubric；工业多模态moderation In/Out-Period评测，private data/changed rules使完整复现受限，下一批需公开替代语义与真实rubric闭环。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11354v1", "sha256": "27419bf408cb2407fedc5c74d59dd28ec0dd9c75a35c3c9a0a1f808b55476a12", "local_html_sha256": "27419bf408cb2407fedc5c74d59dd28ec0dd9c75a35c3c9a0a1f808b55476a12", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11354v1", "result": "retrieved", "sha256": "27419bf408cb2407fedc5c74d59dd28ec0dd9c75a35c3c9a0a1f808b55476a12" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "4 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "5 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/AllSpark-Research/AdaptEvo" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "multimodal-llm" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.10526", "title": "Rephrase Before You Act: Characterizing and Mitigating Language Sensitivity in Vision-Language-Action Models", "published": "2026-10-07T17:57:38Z", "status": "reviewed_deferred", "decision_rationale": "VLA rephrasing从training任务phrase success提炼规则、deployment一次重写，无权重训练；SIMPLER/LIBERO heldout phrasing/tasked控制，转foundation/embodied robustness下一批，不能登记参数posttrain。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10526v1", "sha256": "b6c7e43fe6dfa6ad65ede0685f45386bbdaa153094e68550b8b96bcafa4efe6c", "local_html_sha256": "b6c7e43fe6dfa6ad65ede0685f45386bbdaa153094e68550b8b96bcafa4efe6c", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10526v1", "result": "retrieved", "sha256": "b6c7e43fe6dfa6ad65ede0685f45386bbdaa153094e68550b8b96bcafa4efe6c" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "V Reducing Language Sensitivity with LLM-Distilled Rephrasing Rules", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "VI Evaluating the Rulebooks", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/sttawm/vla-rephrasing-artifacts" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "vision-language" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.10641", "title": "Coverage-Aware Reasoning with Medical Tokens for Diagnosis Prediction", "published": "2026-10-07T14:39:18Z", "status": "reviewed_deferred", "decision_rationale": "CARing ontology-enriched residual-quantized medical SID tokens、alignment+reasoning training、coverage RL和multi-positive监督，MIMIC3/4多标签评测；健康数据受准入协议限制，下一批应先满足data access与多label覆盖contract。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10641v1", "sha256": "38d8000670fab99702044d4287c0368f9bf2c728765a2b91de3dad25ed79cd22", "local_html_sha256": "38d8000670fab99702044d4287c0368f9bf2c728765a2b91de3dad25ed79cd22", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10641v1", "result": "retrieved", "sha256": "38d8000670fab99702044d4287c0368f9bf2c728765a2b91de3dad25ed79cd22" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/zmlxzyh/CARing-Codes-Logs" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "efficient-inference" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.10145", "title": "InterView-C: A Synchronized Multimodal Corpus of VR Avatar-Mediated Survey Interviews", "published": "2026-10-07T14:23:32Z", "status": "rejected_or_out_of_scope", "decision_rationale": "InterView-C主要贡献是German VR interview语料、ASR/negation transfer诊断；读摘要及study procedure，没有当前批次LLM posttraining/agent核心算法，不因Meta Quest设备误判Meta机构论文。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10145v1", "sha256": "3d28484991584187b3a00b3f409dd848592a0996e447225d5fa67e7931a35830", "local_html_sha256": "3d28484991584187b3a00b3f409dd848592a0996e447225d5fa67e7931a35830", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10145v1", "result": "retrieved", "sha256": "3d28484991584187b3a00b3f409dd848592a0996e447225d5fa67e7931a35830" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Experiment Setup", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/TheBv/faces-dataset-pre-study", "https://github.com/snakers4/silero-vad" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "当前任务范围不纳入", "matched_queries": [ "multimodal-llm" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.09914", "title": "RollVerify: Bridging Efficiency and Accuracy in Long-Tail Rollout Reinforcement Learning", "published": "2026-10-07T12:05:33Z", "status": "reviewed_deferred", "decision_rationale": "RollVerify以OPS衡量partial rollout off-policy偏移，并在训练前sequence/token验证截断repair suffix；VeRL Qwen3-8B math/tool-math及初步code对照，下一批需真实stale sampling/repair和throughput，非仅importance loss。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09914v1", "sha256": "4c34b39e848c01e4f41107803340bcbca7458df59abed5858d423ef8e62c7035", "local_html_sha256": "4c34b39e848c01e4f41107803340bcbca7458df59abed5858d423ef8e62c7035", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09914v1", "result": "retrieved", "sha256": "4c34b39e848c01e4f41107803340bcbca7458df59abed5858d423ef8e62c7035" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "4 RollVerify", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "5 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/THUDM/slime" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "long-context" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.09841", "title": "ORCA: Hunting Compositional Failures in Text-to-Image Diffusion", "published": "2026-10-07T11:03:58Z", "status": "reviewed_deferred", "decision_rationale": "ORCA将T5-CLIP residual条件化orthogonal low-rank visual readout作为diffusion latent辅助loss；COCO DiT/U-ViT FID/GenEval训练对照，转foundation T2I下一批，必须真实diffusion/visual encoder与QR投影。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09841v1", "sha256": "cc643f18eee8de1d5c81929ff89816f48b202b92f66d68a7b76b68f9dcf7e57c", "local_html_sha256": "cc643f18eee8de1d5c81929ff89816f48b202b92f66d68a7b76b68f9dcf7e57c", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09841v1", "result": "retrieved", "sha256": "cc643f18eee8de1d5c81929ff89816f48b202b92f66d68a7b76b68f9dcf7e57c" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-architecture", "transformer-architecture", "vision-language" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.09788", "title": "Judging in Latent Space: Efficient Generative Reward Modeling via Semantics-Preserving Compression", "published": "2026-10-07T10:05:06Z", "status": "reviewed_deferred", "decision_rationale": "LatentGRM semantic-chunk compression/reconstruction训练连续latent judge，另interpreter重构criterion评测，matched4B/8B preference和latency控制；下一批需真实continuous steps不能短文本proxy。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09788v1", "sha256": "3def777973d069f686e65d1c5752fdc2519351c8ad456ed15bded5581905b51d", "local_html_sha256": "3def777973d069f686e65d1c5752fdc2519351c8ad456ed15bded5581905b51d", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09788v1", "result": "retrieved", "sha256": "3def777973d069f686e65d1c5752fdc2519351c8ad456ed15bded5581905b51d" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/EADMO/LatentGRM" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "model-compression" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.09647", "title": "When Rank Rises as LLMs Degrade", "published": "2026-10-07T08:22:30Z", "status": "reviewed_deferred", "decision_rationale": "RankMe/effective-rank在Qwen3-0.6B degradation controls和leave-one-seed-out calibration/test下双向诊断；没早于heldout loss且健康seed false alarm，适合training-health诊断下一批，不能宣传早预警或能力算法。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09647v1", "sha256": "f1d0015f7de8cdb6c3d64b99680a0e78572b559f609df58d4d975fb628581267", "local_html_sha256": "f1d0015f7de8cdb6c3d64b99680a0e78572b559f609df58d4d975fb628581267", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09647v1", "result": "retrieved", "sha256": "f1d0015f7de8cdb6c3d64b99680a0e78572b559f609df58d4d975fb628581267" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Setup", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "6 A two-sided monitor under held-out false-alarm evaluation", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "pretraining-data" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.09517", "title": "It Is Not Seeing the Hazard: A Frozen Vision-Language Safety Score Measures Its Caption Bank", "published": "2026-10-07T06:15:23Z", "status": "reviewed_deferred", "decision_rationale": "冻结CLIP prompt-margin safety passive scorer，用VLM-free trajectories、geometry-matched pre-contact controls分离caption/viewpoint混杂；下一批evaluation integrity候选，不以policy returns当hazard perception证据。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09517v1", "sha256": "dc8be4442badc95545d08301558b75b8750b01b944d83ed822185fff9a1c4fe6", "local_html_sha256": "dc8be4442badc95545d08301558b75b8750b01b944d83ed822185fff9a1c4fe6", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09517v1", "result": "retrieved", "sha256": "dc8be4442badc95545d08301558b75b8750b01b944d83ed822185fff9a1c4fe6" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Measurement protocol", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "5 Results", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "vision-language" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.09471", "title": "When Should an In-Context Learner Expand Its Hypothesis Space?", "published": "2026-10-07T05:26:15Z", "status": "reviewed_deferred", "decision_rationale": "SRE exact Bayesian one-shot structural family expansion价值边界与learner utility训练，cost/horizon独立控制；受控理论诊断符合机制研究但非现成agent capability，下一批需gold-free action policy与明确diagnostic等级。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09471v1", "sha256": "26b2e479679184e5339afb2b5405e08a7290e5a75280b693c5e88005fc3d9957", "local_html_sha256": "26b2e479679184e5339afb2b5405e08a7290e5a75280b693c5e88005fc3d9957", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09471v1", "result": "retrieved", "sha256": "26b2e479679184e5339afb2b5405e08a7290e5a75280b693c5e88005fc3d9957" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Structural Revision Environment", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "5 Results", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://huggingface.co/Qwen/Qwen3-8B-Base/tree/49e3418fbbbc", "https://huggingface.co/Qwen/Qwen3-8B/tree/b968826d9c46", "https://huggingface.co/allenai/Llama-3.1-Tulu-3-8B-DPO/tree/a7beb67e33ff", "https://huggingface.co/allenai/Llama-3.1-Tulu-3-8B-SFT/tree/f2a0b46b0cfd", "https://huggingface.co/allenai/Llama-3.1-Tulu-3-8B/tree/666943798adb", "https://huggingface.co/allenai/Llama-3.1-Tulu-3.1-8B/tree/46239c2d07db", "https://huggingface.co/allenai/Olmo-3-1025-7B/tree/373bad25002f", "https://huggingface.co/allenai/Olmo-3-1025-7B/tree/a81bae42db39", "https://huggingface.co/allenai/Olmo-3-7B-Instruct-DPO/tree/b33130b7de49", "https://huggingface.co/allenai/Olmo-3-7B-Instruct-SFT/tree/e1452fc572d5", "https://huggingface.co/allenai/Olmo-3-7B-Instruct/tree/6e5971d9eba4", "https://huggingface.co/allenai/Olmo-3-7B-Think-DPO/tree/7b18bf927b43", "https://huggingface.co/allenai/Olmo-3-7B-Think-SFT/tree/6ff857587e04", "https://huggingface.co/allenai/Olmo-3-7B-Think/tree/d97e442d7cc6", "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct/tree/0e9e39f249a1", "https://huggingface.co/meta-llama/Llama-3.1-8B/tree/d04e592bb4f6" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "long-context" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.09355", "title": "Multimodal LLMs Can Learn to Read Brain Signals: A Vision--Language Model for Unified Multi-Task EEG Decoding", "published": "2026-10-07T03:13:03Z", "status": "reviewed_deferred", "decision_rationale": "BraVista把EEG时频结构图作为VLM接口进行joint multitask后训练，HMC/SEED/EEGMAT/TUAB三seed评测并排除SEED预训练污染；下一批需真实EEG preprocessing/VLM，医疗数据准入与subject隔离。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09355v1", "sha256": "40671903f841bf1fb498f7517e69a8469952e4a1a9529e6c89d103800981ddd6", "local_html_sha256": "40671903f841bf1fb498f7517e69a8469952e4a1a9529e6c89d103800981ddd6", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09355v1", "result": "retrieved", "sha256": "40671903f841bf1fb498f7517e69a8469952e4a1a9529e6c89d103800981ddd6" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "3 Experiments and Results", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "multimodal-llm", "vision-language" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.09346", "title": "OnlineQAT: On-Policy Distillation for Ultra-Low-Bit Large Language Models", "published": "2026-10-07T03:05:42Z", "status": "reviewed_deferred", "decision_rationale": "OnlineQAT先blockwise W2/W3 QAT再quantized student prefixes OPD恢复，Qwen3-1.7B六benchmark明确group quantization/frozen teacher；下一批需实际低bit forward和两阶段优化，不能浮点假quantization标签。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09346v1", "sha256": "93d2d8dcc7571d9e0017da89ccf5c3d6b0d9f116369c68da21c9c7b82d155093", "local_html_sha256": "93d2d8dcc7571d9e0017da89ccf5c3d6b0d9f116369c68da21c9c7b82d155093", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09346v1", "result": "retrieved", "sha256": "93d2d8dcc7571d9e0017da89ccf5c3d6b0d9f116369c68da21c9c7b82d155093" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "model-compression" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.09328", "title": "Visual Jev Rewards: Reference-Bound Verification for Multi-Subject Image Generation", "published": "2026-10-07T02:34:09Z", "status": "reviewed_deferred", "decision_rationale": "Visual Jev把reference identities和condition联合Yes标签训练Qwen3.5 verifier后GRPO多主体生成；手工selected897 tasks且单GRPO run/human差异未显著，下一批只能复现参考绑定reward与诊断，不扩大稳定能力结论。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09328v1", "sha256": "bafd194495e686966ab92efa108b85e7d8987ef44d50ded4de852724d269c946", "local_html_sha256": "bafd194495e686966ab92efa108b85e7d8987ef44d50ded4de852724d269c946", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09328v1", "result": "retrieved", "sha256": "bafd194495e686966ab92efa108b85e7d8987ef44d50ded4de852724d269c946" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Reference-Bound Visual Jev Rewards", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Data and Experimental Setup", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://huggingface.co/blog/jgalego/rljf" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "vision-language" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.09179", "title": "LayerRoPE: Dynamic Depth-wise Magnitude & Angular Superposition", "published": "2026-10-06T22:30:58Z", "status": "reviewed_deferred", "decision_rationale": "LayerRoPE共享gain向量加depth rotation/scalars替换layerwise gamma，C4同顺序scaling/depth/LR/looped模型控制；主要foundation预训练architecture而非OPD，下一批需真实前向与scaled训练。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09179v1", "sha256": "7a67b989be87aa337de905348072114bc41c0ff742d603007e19e3d042ad506c", "local_html_sha256": "7a67b989be87aa337de905348072114bc41c0ff742d603007e19e3d042ad506c", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09179v1", "result": "retrieved", "sha256": "7a67b989be87aa337de905348072114bc41c0ff742d603007e19e3d042ad506c" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 LayerRoPE", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments and Results", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-architecture", "transformer-architecture", "vision-language" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.10552", "title": "Freeze the Decoder, Heal the Encoder: Parameter-Efficient Adaptation for SVD-Based KV-Cache Compression", "published": "2026-09-25T20:48:04Z", "status": "reviewed_deferred", "decision_rationale": "SVD KV compression encoder-only healing在per-arm调LR后三seed与decoder/full parity但减optimizer memory；属于有效parameter-efficiency及调参混杂诊断，下一批需实际factorized cache、独立val LR选择，不能shared LR制造收益。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10552v1", "sha256": "d8dc1679babf8917c9e4a2202c9369499edc0e8bd1089972aae72a17bd9f1cc5", "local_html_sha256": "d8dc1679babf8917c9e4a2202c9369499edc0e8bd1089972aae72a17bd9f1cc5", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10552v1", "result": "retrieved", "sha256": "d8dc1679babf8917c9e4a2202c9369499edc0e8bd1089972aae72a17bd9f1cc5" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "3 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "model-compression" ], "tracks": [ "foundation-model", "post-training" ] }, { "arxiv_id": "2610.13896", "title": "Structural Automata for Neural Semantic Parsing", "published": "2026-10-09T17:42:14Z", "status": "reviewed_deferred", "decision_rationale": "Structural automaton rejected-token training与APD保证固定length budget下accepting path；PMB train/dev/test清楚并dev loss选checkpoint，下一批需真实automaton completion witness与DRS/AMR parser，不用gold完成fixture。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.13896v1", "sha256": "4cdb01c4e378eae3305eca830fe266aa17a2e5cd6dd0c930519d0d775e4e5a5d", "local_html_sha256": "4cdb01c4e378eae3305eca830fe266aa17a2e5cd6dd0c930519d0d775e4e5a5d", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.13896v1", "result": "retrieved", "sha256": "4cdb01c4e378eae3305eca830fe266aa17a2e5cd6dd0c930519d0d775e4e5a5d" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 One Automaton, Three Stages", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experimental Setup", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "post-training" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.13821", "title": "Can Language Models Learn When to Abstain? Policy Learning and Calibrated Answer Selection", "published": "2026-10-09T16:33:27Z", "status": "reviewed_deferred", "decision_rationale": "Qwen3-14B factual abstention比较RL policy和trained selector，独立training/calibration与heldout thresholds；selector TriviaQA改善但NQ仍不及always abstain，下一批保留negative transfer和utility baseline。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.13821v1", "sha256": "714264f7afe089a6666ad89798fb294a4f5e522467cf6a434fd2b965fc45e3bd", "local_html_sha256": "714264f7afe089a6666ad89798fb294a4f5e522467cf6a434fd2b965fc45e3bd", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.13821v1", "result": "retrieved", "sha256": "714264f7afe089a6666ad89798fb294a4f5e522467cf6a434fd2b965fc45e3bd" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Results", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/jrterven/Can-LLMs-learn-to-abstain" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-rl", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.13744", "title": "A Privileged Teacher That Only Scores Cannot Search: Critique-Conditioned On-Policy Self-Distillation for Reasoning", "published": "2026-10-09T15:36:47Z", "status": "reviewed_deferred", "decision_rationale": "Critique-conditioned OPSD先显式生成纠错critique再score，比privileged score-only错误定位更好；gold error spans只能外部诊断，test critique不可读取reference解；下一批需区分privileged training与gold-free test/runtime。 论文oracle critiques含reference只算privileged诊断；本地不能作为gold-free test能力证据。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.13744v1", "sha256": "bb77f821556417d32712123fba118abc815c7dd4105e0c8a4a5a3e668160db8e", "local_html_sha256": "bb77f821556417d32712123fba118abc815c7dd4105e0c8a4a5a3e668160db8e", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.13744v1", "result": "retrieved", "sha256": "bb77f821556417d32712123fba118abc815c7dd4105e0c8a4a5a3e668160db8e" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "7 Critique-Conditioned Self-Distillation Delivers", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "4 Setup", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "on-policy-distillation", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.13653", "title": "A Critical Look at Power Sampling when Test Time Sharpening LLMs", "published": "2026-10-09T14:29:28Z", "status": "reviewed_deferred", "decision_rationale": "Power sampling practical MCMC/SMC不收敛对照Best-ofN，并用Power-RLOO直接拟合target探索目标价值；MATH500/GPQA模型特异噪声结果，下一批需真正target sampling/梯度、不能用BoN冒称power distribution。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.13653v1", "sha256": "65486a1deedbb259bf7c63d5f002da225a2402b2400b78c159019dbc67a4a4ae", "local_html_sha256": "65486a1deedbb259bf7c63d5f002da225a2402b2400b78c159019dbc67a4a4ae", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.13653v1", "result": "retrieved", "sha256": "65486a1deedbb259bf7c63d5f002da225a2402b2400b78c159019dbc67a4a4ae" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 The Pitfalls of Existing Sharpening Schemes", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "5 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "test-time-rl" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.13617", "title": "Training User Simulators with Latent Intrinsic Feedback", "published": "2026-10-09T14:00:36Z", "status": "reviewed_deferred", "decision_rationale": "官方PDF第一页确认Google Research/Georgia Tech；LES user simulator以real logs监督thought/rating/reaction，Sec4独立4000对话反馈/realism，Sec5实际multiturn assistant RL对比intrinsic/external rewards。Google优先合格候选，冻结16外defer下一批；须真实simulator与assistant training，不能伪造user反馈。", "priority_reviewed": true, "verified_organization": "Google Research / Georgia Institute of Technology", "source": { "url": "https://arxiv.org/html/2610.13617v1", "sha256": "e5fdc46eca3a6e95b37bba0e6e2fc249c995775dc9391357d9005946c6f095af", "local_html_sha256": "e5fdc46eca3a6e95b37bba0e6e2fc249c995775dc9391357d9005946c6f095af", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.13617v1", "result": "retrieved", "sha256": "e5fdc46eca3a6e95b37bba0e6e2fc249c995775dc9391357d9005946c6f095af" } ], "alternate_evidence": { "url": "https://arxiv.org/pdf/2610.13617v1", "sha256": "f831e52cb9bbb95884277533bcae71cb2dabc47eb598cbc78501d7100c9efbf3", "location": "PDF page 1 author affiliations", "read_scope": "author/institution block only; method/evaluation sections read from official HTML" } }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 User Simulators with Latent Evaluation States", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Evaluating the Feedback Behavior of User Simulators", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "5 Training LLM Assistants with Simulators via Multi-Turn RL", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/huggingface/trl" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.13058", "title": "Travel Back To Move Forward: Retrospective Exploration and Evolving for Long Horizon Agents", "published": "2026-10-09T06:11:39Z", "status": "reviewed_deferred", "decision_rationale": "ReTRAIL真实恢复环境状态并保留abandoned exploration经验，以State-Anchored GRPO与branch advancement reward共同训练；MineSweeper/Sokoban/ALFWorld/ScienceWorld评测，具Facebook署名线索优先审查，需真实rewind接口，冻结16外下一批。", "priority_reviewed": true, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.13058v1", "sha256": "dac2da4a6c200b7b10b09a96b67249a22ccc39e56dfe33189fffc5e86dea9c57", "local_html_sha256": "dac2da4a6c200b7b10b09a96b67249a22ccc39e56dfe33189fffc5e86dea9c57", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.13058v1", "result": "retrieved", "sha256": "dac2da4a6c200b7b10b09a96b67249a22ccc39e56dfe33189fffc5e86dea9c57" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Retrospective Exploration", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Learning from Retrospective Exploration", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "5 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/whi497/ReTRAIL" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.12885", "title": "Personalized Federated Direct Preference Optimization via LoRA Adapters", "published": "2026-10-09T02:20:29Z", "status": "reviewed_deferred", "decision_rationale": "pFedDPO同步global base、private LoRA保留client preference，SHP/hh5 clients非IID实验、主要gradient convergence；下一批需真实FedAvg与local adapters，private数据不能centralize，训练loss不等于user效用。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12885v1", "sha256": "b6753f95bca1a75347d72e607738bd7b4248bdaaf6fa7123fc7a67b52a2dae62", "local_html_sha256": "b6753f95bca1a75347d72e607738bd7b4248bdaaf6fa7123fc7a67b52a2dae62", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12885v1", "result": "retrieved", "sha256": "b6753f95bca1a75347d72e607738bd7b4248bdaaf6fa7123fc7a67b52a2dae62" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "Personalized FedDPO (pFedDPO)", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "preference-optimization", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.12757", "title": "MADS: Multi-Agent Dialog Simulation for Robust Conversational Agents", "published": "2026-10-08T22:08:10Z", "status": "reviewed_deferred", "decision_rationale": "MADS三agent模拟user/assistant/tool含adversarial personas并验证dialogue，SFT+RLVR训练robust tool dialogs；下一批需检查106 training trajectories与1095总conversations split isolation，真实tool验证避免泄漏。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12757v1", "sha256": "053fe9f848f542a6258a943471764084edddcd5ba9f07284da03682037c146f0", "local_html_sha256": "053fe9f848f542a6258a943471764084edddcd5ba9f07284da03682037c146f0", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12757v1", "result": "retrieved", "sha256": "053fe9f848f542a6258a943471764084edddcd5ba9f07284da03682037c146f0" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 MADS-Bench", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://huggingface.co/sentence-transformers/all-mpnet-base-v2" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-rl", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.12566", "title": "Sample-Adaptive Zeroth-Order Post-Training", "published": "2026-10-08T18:00:37Z", "status": "reviewed_deferred", "decision_rationale": "官方PDF已取得并限定阅读算法及评测。SAZO复用two-forward directional difference，按per-example EMA估计梯度范数重要性，带uniform floor importance sampling及1/(Np_i)无偏缩放，仅更新抽中样本score；原文有RoBERTa/OPT/Phi/Llama matched-budget实证。冻结16篇之外，下一批需真实ZO参数扰动/恢复、随机方向与采样账本，不能常量loss或固定重权代理。", "priority_reviewed": false, "verified_organization": "Peking University", "source": { "url": "https://arxiv.org/html/2610.12566v1", "sha256": null, "local_html_sha256": null, "local_html_sha256_matches": null, "attempts": [ { "url": "https://arxiv.org/html/2610.12566v1", "result": "failed", "error": "404 Client Error: Not Found for url: https://arxiv.org/html/2610.12566v1" }, { "url": "https://arxiv.org/html/2610.12566v1", "result": "failed", "error": "404 Client Error: Not Found for url: https://arxiv.org/html/2610.12566v1" }, { "url": "https://arxiv.org/pdf/2610.12566v1", "result": "retrieved", "retrieved_on": "2026-10-12", "command": "curl -sSL --fail --max-time 45 https://arxiv.org/pdf/2610.12566v1 -o tmp/pdfs/2610.12566.pdf", "sha256": "109824aa290f19c36607545f5579a3472def7a2bf9c5793f0b13c31aa5a4b2f2" } ], "alternate_evidence": { "url": "https://arxiv.org/pdf/2610.12566v1", "format": "official_arxiv_pdf", "sha256": "109824aa290f19c36607545f5579a3472def7a2bf9c5793f0b13c31aa5a4b2f2", "local_path": "tmp/pdfs/2610.12566.pdf", "local_sha256_verified": true, "total_pdf_pages": 23 } }, "read_scope": "abstract_plus_bounded_official_pdf_method_and_evaluation_pages", "full_article_read": false, "evidence_locations": [ { "heading": "Title/author and arXiv v1 margin", "pdf_page": 1, "scope": "full page text read; manuscript cover date is not substituted for arXiv v1 date" }, { "heading": "Method and evaluation/conclusion", "pdf_pages": [ 3, 4, 6, 7, 8, 9 ], "scope": "listed page text read; not all article pages or appendices", "method_evidence": "s_i=(f(x+mu z)-f(x-mu z))/(2mu)；p_i=(1-eps)m_i/sum(m)+eps/N；update=s_i z/(N p_i)；仅selected m_i做(1-beta)m_i+beta sqrt(pi/2)|s_i|，其他score不变；无额外ZO查询、O(N)scalar state。", "evaluation_evidence": "§4/Table2: RoBERTa-large350M六任务full/PEFT；Table3: OPT1.3B/Phi2/Llama3-8B八任务并同模型LR search grid；Figure2在4k/6k/6k forward passes达到MeZO24k端点，不能直接称wallclock同倍加速；Table4 latency与MeZO相近；population scaling三seed；beta1发散，.005/.05/.5稳定且.05最佳。" } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结16篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "post-training" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.12545", "title": "AutoIndexer: Flexible-Order Decoding by Training Causal Models on Chains of Edits", "published": "2026-10-08T17:59:52Z", "status": "reviewed_deferred", "decision_rationale": "AutoIndexer用inverse chains of edits训练causal model cursor/insert/delete/substitute，Qwen3-8B LoRA extra100k CPT与原AR/LLM benchmarks；下一批真实cursor与edit生成语义，不能后处理edit heuristic冒充模型。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12545v1", "sha256": "5e0bec696fb8d5d4ebc4c72d8cd44acc5c5ff77f4f4f2fa4b1750081772ed5ff", "local_html_sha256": "5e0bec696fb8d5d4ebc4c72d8cd44acc5c5ff77f4f4f2fa4b1750081772ed5ff", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12545v1", "result": "retrieved", "sha256": "5e0bec696fb8d5d4ebc4c72d8cd44acc5c5ff77f4f4f2fa4b1750081772ed5ff" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 AutoIndexer", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "3 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/AutodeskAILab/autoindexer", "https://github.com/Sean-Diab/Reviser", "https://github.com/ZHZisZZ/dllm" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "post-training" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.12420", "title": "A Unified Bellman Operator for Safety-Critical Reinforcement Learning", "published": "2026-10-08T17:53:02Z", "status": "rejected_or_out_of_scope", "decision_rationale": "统一Bellman safety/task value与两时间尺度收敛主要面向neural continuous control；未涉及LLM/foundation/agent语言后训练，当前研究批次out of scope，不因强化学习关键词纳入。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12420v1", "sha256": "7beeff8ccaa4d70298d7ed5f54eba7a7bc07043a4f08f0a0c99272f11ac0c8cd", "local_html_sha256": "7beeff8ccaa4d70298d7ed5f54eba7a7bc07043a4f08f0a0c99272f11ac0c8cd", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12420v1", "result": "retrieved", "sha256": "7beeff8ccaa4d70298d7ed5f54eba7a7bc07043a4f08f0a0c99272f11ac0c8cd" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Learning Safety-Constrained Task Action-Value Function", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "7 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/PKU-Alignment/safety-gymnasium/tree/main/benchmarks#important-notes" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "当前任务范围不纳入", "matched_queries": [ "test-time-rl" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.12410", "title": "Predicting Alignment Generalization with Value Representations", "published": "2026-10-08T17:47:26Z", "status": "reviewed_deferred", "decision_rationale": "Value representations通过single-value DPO/activation spaces预测66-value alignment generalization，并研究multivalue robustness；下一批需真实finetune与heldout value matrix，predictive correlations不能当causal controls保证。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12410v1", "sha256": "86b49a8ca953b909c9ee9eaeb0e3450be53276ebb8a9973dcc0f10cb2d9d3c9a", "local_html_sha256": "86b49a8ca953b909c9ee9eaeb0e3450be53276ebb8a9973dcc0f10cb2d9d3c9a", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12410v1", "result": "retrieved", "sha256": "86b49a8ca953b909c9ee9eaeb0e3450be53276ebb8a9973dcc0f10cb2d9d3c9a" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Computing Alignment Generalization Effects", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Which Value Representations Best Predict Alignment Generalization?", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/andyjliu/value-generalization", "https://github.com/huggingface/trl", "https://huggingface.co/datasets/allenai/tulu-3-sft-olmo-2-mixture" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "post-training" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.12367", "title": "Which Skill to Distill? SGUID: Selecting a Compact Skill Bank for Model-Skill Co-Evolution", "published": "2026-10-08T17:27:53Z", "status": "accepted", "decision_rationale": "SGUID已implemented：按skill-conditioned teachers对student trajectory的有效learning signals选compact bank，并updated-model rounds重建bank；DAPO/math benchmark与bank controls，保留existing实现状态，不宣称本批新实现。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12367v1", "sha256": "44936fbd5a055ab8961f343749b98311cbe44d4b1f79d75d367cc23b3757a0c7", "local_html_sha256": "44936fbd5a055ab8961f343749b98311cbe44d4b1f79d75d367cc23b3757a0c7", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12367v1", "result": "retrieved", "sha256": "44936fbd5a055ab8961f343749b98311cbe44d4b1f79d75d367cc23b3757a0c7" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "implemented", "new_implementation_claimed": false, "next_action": "保留已有登记；不重复实施或补称本次 GPU 通过", "matched_queries": [ "on-policy-distillation" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.12274", "title": "HarnessSQL: Harness-Native Training for SQL Agents in Realistic Database Environments", "published": "2026-10-08T16:36:39Z", "status": "reviewed_deferred", "decision_rationale": "HarnessSQL在hermetic live DB/harness中教师verified全interaction SFT→execution RL，79db训练与100val，Spider2/BIRD/LiveSQL transfer；下一批必须hide execution oracle、readonly DB sandbox和4tool真实执行。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12274v1", "sha256": "fe5902a21b323d42333a4533b879121cafc8f2d75484f0374d1b3fcd7cb461ce", "local_html_sha256": "fe5902a21b323d42333a4533b879121cafc8f2d75484f0374d1b3fcd7cb461ce", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12274v1", "result": "retrieved", "sha256": "fe5902a21b323d42333a4533b879121cafc8f2d75484f0374d1b3fcd7cb461ce" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experimental Setup", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/YangHaolin0526/HarnessSQL", "https://github.com/bird-bench/livesqlbench/tree/main/LiveSQLBench-CLI", "https://github.com/deepseek-ai/deepseek-harness" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-rl", "post-training", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.12207", "title": "SciTBERT: A family of chronologically consistent language models for scientific and technological language processing", "published": "2026-10-08T15:57:45Z", "status": "reviewed_deferred", "decision_rationale": "SciTBERT年份2013-2025 cutoff pretrain/citation posttrain保持chronology，SciRepEval/PatRepEval embeddings任务；foundation encoder nextbatch需历史cutoff与citation pairs隔离，不能把现有晚期模型当无lookahead。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12207v1", "sha256": "ba6bd944317d96a918277a1e11fd0215ea6d2f1941cc6ceb7569146bb8b4bd90", "local_html_sha256": "ba6bd944317d96a918277a1e11fd0215ea6d2f1941cc6ceb7569146bb8b4bd90", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12207v1", "result": "retrieved", "sha256": "ba6bd944317d96a918277a1e11fd0215ea6d2f1941cc6ceb7569146bb8b4bd90" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "Contrastive post-training with chronologically consistent citation data", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "Evaluation setup", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/KPSS2017/Technological-Innovation-Resource-Allocation-and-Growth-Extended-Data", "https://huggingface.co/datasets/gebhart/patrepeval" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "post-training" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.12201", "title": "SteerablePlex: Can We Steer Full-Duplex Models?", "published": "2026-10-08T15:56:28Z", "status": "reviewed_deferred", "decision_rationale": "SteerablePlex full-duplex actor+asynchronous director，GDPO LoRA训练保turn-taking并跟多stage文本goal；200 SimIF scenarios/音频ASR-VAD评测，下一批需真实full-duplex streaming/director异步和rewarddecoupling。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12201v1", "sha256": "c30f3cb3c91006da12cb43ababc3531a23fc30dd70faf972921ff345abf21edb", "local_html_sha256": "c30f3cb3c91006da12cb43ababc3531a23fc30dd70faf972921ff345abf21edb", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12201v1", "result": "retrieved", "sha256": "c30f3cb3c91006da12cb43ababc3531a23fc30dd70faf972921ff345abf21edb" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiment", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://huggingface.co/Qwen/Qwen3.8-27B", "https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.11854", "title": "GRPODropout: Less is More for Online Reinforcement Learning Rollouts", "published": "2026-10-08T12:38:17Z", "status": "accepted", "decision_rationale": "GRPODropout已有实现：删高概率positive-advantage rollout并recenter retained credit，matched budget math/generalQA三scale比较；保留existing范围，不因survey再称本批新增。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11854v1", "sha256": "0c0f923d5467463326ae73186b4159a965ceb9358e7b6cf8ef011ee1080d8d43", "local_html_sha256": "0c0f923d5467463326ae73186b4159a965ceb9358e7b6cf8ef011ee1080d8d43", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11854v1", "result": "retrieved", "sha256": "0c0f923d5467463326ae73186b4159a965ceb9358e7b6cf8ef011ee1080d8d43" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/hexuandeng/GRPODropout/" ], "repository_status_at_assignment": "implemented", "new_implementation_claimed": false, "next_action": "保留已有登记；不重复实施或补称本次 GPU 通过", "matched_queries": [ "llm-rl", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.11659", "title": "DIAL-OPD: Learning More from Fewer Tokens in On-Policy Distillation", "published": "2026-10-08T10:36:00Z", "status": "accepted", "decision_rationale": "DIAL-OPD已有实现：logarithmic mean与density exponent调节token gap，retention40%多个teacher/student；本次核对论文选择机制与seven-benchmark口径，不新增实现。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11659v1", "sha256": "7e980643b14c22e3c39588e084182f74ad107b56ad6aa96333a9aaa5cb7ae3a9", "local_html_sha256": "7e980643b14c22e3c39588e084182f74ad107b56ad6aa96333a9aaa5cb7ae3a9", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11659v1", "result": "retrieved", "sha256": "7e980643b14c22e3c39588e084182f74ad107b56ad6aa96333a9aaa5cb7ae3a9" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "4 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "5 Experimental Setup", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/EIT-NLP/DIAL-OPD", "https://github.com/THUDM/slime" ], "repository_status_at_assignment": "implemented", "new_implementation_claimed": false, "next_action": "保留已有登记；不重复实施或补称本次 GPU 通过", "matched_queries": [ "on-policy-distillation", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.11638", "title": "UXBench Pro: Benchmarking Personalized User Experience in Multi-Turn Dialogue Interactions", "published": "2026-10-08T10:15:55Z", "status": "reviewed_deferred", "decision_rationale": "UXBench Pro真实交互test+FACTORS profiles，并URM/Sim4Eval及两meta-bench校验human alignment；private人口日志/用户profile要额外data协议，下一批evaluation/per-user reward候选，不与gold评分混用。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11638v1", "sha256": "0067ba18f5a076d9b70a088728fb4340ba0c6eeb1039769579ea5af911a04e60", "local_html_sha256": "0067ba18f5a076d9b70a088728fb4340ba0c6eeb1039769579ea5af911a04e60", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11638v1", "result": "retrieved", "sha256": "0067ba18f5a076d9b70a088728fb4340ba0c6eeb1039769579ea5af911a04e60" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 UXBench Pro", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "5 Meta-Benchmarking", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "preference-optimization", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.11519", "title": "Residual Advantage: Student-Relative Teacher Guidance for RL with Verifiable Rewards", "published": "2026-10-08T08:52:55Z", "status": "accepted", "decision_rationale": "Residual Advantage/CoRA已有实现：bounded teacher-student probability residual减state value、response中心化保verifier mean，CoRA teacher LoRA逐轮更新；math三seed对照，保留existing实现不宣称新GPU证据。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11519v1", "sha256": "6c9bd85a439410f9ebce47dbb01f5f449ecc175878b03b55defd783eb10967de", "local_html_sha256": "6c9bd85a439410f9ebce47dbb01f5f449ecc175878b03b55defd783eb10967de", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11519v1", "result": "retrieved", "sha256": "6c9bd85a439410f9ebce47dbb01f5f449ecc175878b03b55defd783eb10967de" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 Residual Advantage", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "3 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "implemented", "new_implementation_claimed": false, "next_action": "保留已有登记；不重复实施或补称本次 GPU 通过", "matched_queries": [ "llm-rl", "on-policy-distillation", "post-training", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.11366", "title": "SpatialOPSD: Self-Distilling Spatial Intelligence from Verified Coding Agent Traces", "published": "2026-10-08T06:58:30Z", "status": "reviewed_deferred", "decision_rationale": "SpatialOPSD用verified coding-agent traces为privileged teacher生成spatialCoT，repetition-aware mask+unlikelihood控leakage；5008原训练题student只见原input，SFT用不同4459数量应记录，不把等protocol称等data预算；下一批。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11366v1", "sha256": "3c5c06be29a5c47463999447782e96826ec01e077f3094939f6deae50733410e", "local_html_sha256": "3c5c06be29a5c47463999447782e96826ec01e077f3094939f6deae50733410e", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11366v1", "result": "retrieved", "sha256": "3c5c06be29a5c47463999447782e96826ec01e077f3094939f6deae50733410e" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Spatial Reasoning via OPSD", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "on-policy-distillation" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.11352", "title": "RL-ARC: Calibrating Large Reasoning Models via Reasoning-guided Uncertainty", "published": "2026-10-08T06:45:24Z", "status": "reviewed_deferred", "decision_rationale": "RL-ARC以reasoning/answer verbalized confidences对correct regularize/incorrect overconfident penalize，在ID/OOD比较calibration与accuracy；下一批需proper scoring+真实reasoningconfidence，不能硬编码置信值。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11352v1", "sha256": "0c105e996bf31fe0ddcdbbf8daee3426e630b0e92b70f50a66f866862b717507", "local_html_sha256": "0c105e996bf31fe0ddcdbbf8daee3426e630b0e92b70f50a66f866862b717507", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11352v1", "result": "retrieved", "sha256": "0c105e996bf31fe0ddcdbbf8daee3426e630b0e92b70f50a66f866862b717507" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 RL-ARC: Reinforcement Learning with Auxiliary Reasoning Confidence", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/2gukhyeon/RL-ARC" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-rl", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.11332", "title": "ReCal: Calibrating Structured Pruning for On-Policy Distillation Recovery", "published": "2026-10-08T06:28:38Z", "status": "accepted", "decision_rationale": "ReCal已有实现：teacher与pruned probe前向KL重权calibration，引导现有结构剪枝并SFT→OPD恢复；六benchmark需保持实剪/恢复接口，本次原文复核不声称新增运行。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11332v1", "sha256": "2ec4f8c652d52479e2604f4e976dccc1c5b93cf137c10ee56ce131b918e78493", "local_html_sha256": "2ec4f8c652d52479e2604f4e976dccc1c5b93cf137c10ee56ce131b918e78493", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11332v1", "result": "retrieved", "sha256": "2ec4f8c652d52479e2604f4e976dccc1c5b93cf137c10ee56ce131b918e78493" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "4 Recovery-Aware Calibration for Pruning", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "5 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/jianghoucheng/ReCal", "https://huggingface.co/datasets/allenai/IF_multi_constraints_upto5", "https://huggingface.co/datasets/nvidia/Nemotron-RL-Math-v2", "https://huggingface.co/datasets/nvidia/Nemotron-RL-Science-v1", "https://huggingface.co/datasets/nvidia/Nemotron-RL-coding-competitive_coding", "https://huggingface.co/datasets/nvidia/Nemotron-SFT-Competitive-Programming-v2", "https://huggingface.co/datasets/nvidia/Nemotron-SFT-Instruction-Following-Chat-v3", "https://huggingface.co/datasets/nvidia/Nemotron-SFT-Math-v4", "https://huggingface.co/datasets/nvidia/Nemotron-SFT-Science-v2" ], "repository_status_at_assignment": "implemented", "new_implementation_claimed": false, "next_action": "保留已有登记；不重复实施或补称本次 GPU 通过", "matched_queries": [ "on-policy-distillation" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.11291", "title": "When Do We Need On-Policy Distillation? Distilling on Offline Student Rollouts Is Often Better", "published": "2026-10-08T05:54:03Z", "status": "accepted", "decision_rationale": "Semi-OPD已implemented，固定initial student离线rollout teacher token cache，以teacher/student overlap诊断何时持续OPD；17 pairs对照覆盖准确率/效率，本次保留既有而非新重复实施。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11291v1", "sha256": "7d840425b74c01b525bbb24d7b11ce46ee153db415e38ea03f00ebac6e330f21", "local_html_sha256": "7d840425b74c01b525bbb24d7b11ce46ee153db415e38ea03f00ebac6e330f21", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11291v1", "result": "retrieved", "sha256": "7d840425b74c01b525bbb24d7b11ce46ee153db415e38ea03f00ebac6e330f21" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 An Efficient OPD Alternative: Semi-On-Policy Distillation", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "2 An Efficient OPD Alternative: Semi-On-Policy Distillation", "scope": "additional Semi-OPD formula, caching and paired-model results excerpt", "characters_read_at_least": [ 1300, 6500 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/Qihoo360/Light-R1", "https://huggingface.co/POLARIS-Project/Polaris-7B-Preview", "https://huggingface.co/Qwen/Qwen3-1.7B", "https://huggingface.co/Qwen/Qwen3-14B", "https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507", "https://huggingface.co/Qwen/Qwen3-235B-A22B-Thinking-2507", "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507", "https://huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507", "https://huggingface.co/Qwen/Qwen3-4B", "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507", "https://huggingface.co/Qwen/Qwen3-8B", "https://huggingface.co/agentica-org/DeepCoder-14B-Preview", "https://huggingface.co/datasets/open-r1/DAPO-Math-17k-Processed", "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B", "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B", "https://huggingface.co/hbx/JustRL-DeepSeek-1.5B", "https://huggingface.co/qihoo360/Light-R1-14B-DS" ], "repository_status_at_assignment": "implemented", "new_implementation_claimed": false, "next_action": "保留已有登记；不重复实施或补称本次 GPU 通过", "matched_queries": [ "on-policy-distillation" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.11281", "title": "How to post-train on a surrogate: Envelope sampling mitigates reward hacking", "published": "2026-10-08T05:48:31Z", "status": "reviewed_deferred", "decision_rationale": "Envelope sampling在judge L2 uncertainty ball下优化regret bound，用reject或modified-reward finetune采样标注recalibration，clinical notes/sycophancy反rewardhacking；下一批需expert labels与heldout calibration边界，不能oracle答案给被评估policy。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11281v1", "sha256": "0702714fd93108be1cc5bf2b54efe4b4df4141eef22d6a5616d3b7db3a497451", "local_html_sha256": "0702714fd93108be1cc5bf2b54efe4b4df4141eef22d6a5616d3b7db3a497451", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11281v1", "result": "retrieved", "sha256": "0702714fd93108be1cc5bf2b54efe4b4df4141eef22d6a5616d3b7db3a497451" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "Theoretical results", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/sanjitdp/envelope-sampling" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-rl", "post-training", "preference-optimization", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.11247", "title": "Why On-Policy Distillation Sometimes Fails: Vanishing Learning Signals", "published": "2026-10-08T04:52:05Z", "status": "accepted", "decision_rationale": "OPD-learning-signals已implemented diagnostic：small-LR dynamics与true logs gradient proxy、near-student teacher条件保证；early plateau/CKA相关性不证明原因，保持diagnostic而非capability method。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11247v1", "sha256": "614cf91e0d89f57fedbb7e4d6494e2c96cb2e945bdb7edb63df2399b023c2e05", "local_html_sha256": "614cf91e0d89f57fedbb7e4d6494e2c96cb2e945bdb7edb63df2399b023c2e05", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11247v1", "result": "retrieved", "sha256": "614cf91e0d89f57fedbb7e4d6494e2c96cb2e945bdb7edb63df2399b023c2e05" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Preliminaries", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Early Loss Plateaus in On-Policy Distillation", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/leizhao7/opd-learning-signals" ], "repository_status_at_assignment": "implemented", "new_implementation_claimed": false, "next_action": "保留已有登记；不重复实施或补称本次 GPU 通过", "matched_queries": [ "llm-rl", "on-policy-distillation", "test-time-rl" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.11226", "title": "When Lower Reconstruction Loss Hurts: Distributionally Robust Refinement for Low-Bit LLM Quantization", "published": "2026-10-08T04:22:29Z", "status": "reviewed_deferred", "decision_rationale": "DRQ固定quant grid只refine integer codes，worst-case activation distributions纠正calibration reconstruction泛化，六PTQ methods dense/MoE公开ppl/tasks；下一批须真实整数码/量化forward/uncertainty inner problem。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11226v1", "sha256": "e4a077366e7ad0dd0c0dc82e9a3e1d78dc26353c7e10ef4ae50750ec58a8ef27", "local_html_sha256": "e4a077366e7ad0dd0c0dc82e9a3e1d78dc26353c7e10ef4ae50750ec58a8ef27", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11226v1", "result": "retrieved", "sha256": "e4a077366e7ad0dd0c0dc82e9a3e1d78dc26353c7e10ef4ae50750ec58a8ef27" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 DRQ: Distributionally Robust Quantization", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experimental Evaluation", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/z-lab/paroquant" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "post-training" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.11167", "title": "PIVOT: Perplexity-Informed KD-to-RL Transition Scheduling for Vertical-Domain Few-Shot Distillation", "published": "2026-10-08T03:24:55Z", "status": "accepted", "decision_rationale": "PIVOT已有实现：teacher sequence perplexity per-sample OPD→GRPO切换，在Banking77/HWU64 fewshot同postwarmup student steps比较；原文明确train seed与officialtest，保留existing，追加实验不混选test。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11167v1", "sha256": "16faf341f9b98e02844d8185693c43fe02019cc26f828f2b698ebfffeed54399", "local_html_sha256": "16faf341f9b98e02844d8185693c43fe02019cc26f828f2b698ebfffeed54399", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11167v1", "result": "retrieved", "sha256": "16faf341f9b98e02844d8185693c43fe02019cc26f828f2b698ebfffeed54399" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "3 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "implemented", "new_implementation_claimed": false, "next_action": "保留已有登记；不重复实施或补称本次 GPU 通过", "matched_queries": [ "llm-rl", "on-policy-distillation", "post-training", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.11111", "title": "Lapras: Latent Reasoning for Time Series Language Models", "published": "2026-10-08T02:33:37Z", "status": "reviewed_deferred", "decision_rationale": "Lapras真实continuous latent thoughts、CoT teacher selfdistill在answer hidden states对齐；4 TSLM/5 time-series QA公开评测，trainingtrace可引用annotated labels但test learner必须仅signals/questions，下一批需真实time series encoder与latent recurrence。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11111v1", "sha256": "7cc934642f9dfd128458969a4bca9380b2c8d394c6680968d399fc351efc2e9a", "local_html_sha256": "7cc934642f9dfd128458969a4bca9380b2c8d394c6680968d399fc351efc2e9a", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11111v1", "result": "retrieved", "sha256": "7cc934642f9dfd128458969a4bca9380b2c8d394c6680968d399fc351efc2e9a" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiment", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/hiyouga/LLaMA-Factory", "https://github.com/yuc0805/Lapras", "https://huggingface.co/leochen085/Lapras" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "post-training" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.11064", "title": "Measuring and Mitigating Solution Mode Collapse in RLVR", "published": "2026-10-08T01:25:16Z", "status": "reviewed_deferred", "decision_rationale": "ModeBench correctness+mode verifier揭示RLVRmode collapse，ReMax每discovered correctmode存oneexemplar均匀replay；五seed三scale两RLobjectives同sampling预算但增加replayupdate需记录，下一批真实mode verifier和训练。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.11064v1", "sha256": "c4f22dd0282c6ad09e040fc5e6a3bc17cd17a67c7024517299f03742e659062f", "local_html_sha256": "c4f22dd0282c6ad09e040fc5e6a3bc17cd17a67c7024517299f03742e659062f", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.11064v1", "result": "retrieved", "sha256": "c4f22dd0282c6ad09e040fc5e6a3bc17cd17a67c7024517299f03742e659062f" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Re:Max: Learn from New Samples and Remember Successes", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experimental Design", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/liv-daliberti/modeBench", "https://github.com/liv-daliberti/remax", "https://huggingface.co/datasets/od2961/ModeBench", "https://huggingface.co/datasets/od2961/ModeBench/blob/main/EVALUATION.md", "https://huggingface.co/datasets/od2961/ModeBench/blob/main/README.md#loading", "https://huggingface.co/od2961/maxent-grpo-models", "https://huggingface.co/od2961/maxent-grpo-models/blob/main/DATA.md", "https://huggingface.co/od2961/maxent-grpo-models/blob/main/EXPERIMENTS.md", "https://huggingface.co/od2961/maxent-grpo-models/blob/main/RESTORE.md" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-rl", "post-training", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.10974", "title": "Budgeted Multi-Source Counterfactual Annotation for Off-Policy Evaluation", "published": "2026-10-07T22:55:15Z", "status": "rejected_or_out_of_scope", "decision_rationale": "多源counterfactual annotation成本分配与MM/DP主要contextual-bandit OPE，LLM只是教育annotation source；读formulation/numerical experiments，无LLM agent后训练核心，不纳入冻结实施。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10974v1", "sha256": "2cda31d53083bb7e488dd83b1a35d46b3dd943b18946e5b94bed772151340832", "local_html_sha256": "2cda31d53083bb7e488dd83b1a35d46b3dd943b18946e5b94bed772151340832", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10974v1", "result": "retrieved", "sha256": "2cda31d53083bb7e488dd83b1a35d46b3dd943b18946e5b94bed772151340832" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Problem Formulation", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "7 Numerical Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/Lygist/Budgeted-Multi-Source-Counterfactual-Annotation-for-Off-Policy-Evaluation" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "当前任务范围不纳入", "matched_queries": [ "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.10947", "title": "RH-Detect: A Unified Benchmark for Reward Hacking Detection", "published": "2026-10-07T21:52:33Z", "status": "reviewed_deferred", "decision_rationale": "RH-Detect统一11 sources schema并source-heldout detectorSFT/GRPO，pooled AUROC不能掩multi-turn跌落；下一批evaluation integrity需保originalprovenance/labels且heldoutsource选参隔离。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10947v1", "sha256": "e4700bbc349d485cefb8e96053c2971d567f6cd51bd6d35dda0ee3a0276f4bac", "local_html_sha256": "e4700bbc349d485cefb8e96053c2971d567f6cd51bd6d35dda0ee3a0276f4bac", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10947v1", "result": "retrieved", "sha256": "e4700bbc349d485cefb8e96053c2971d567f6cd51bd6d35dda0ee3a0276f4bac" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Dataset Construction", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Zero-Shot Detector Evaluation", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://huggingface.co/datasets/steinad/CheatBench" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.10932", "title": "World-Model Policy Arbiter for Goal-Conditioned Reinforcement Learning", "published": "2026-10-07T21:41:33Z", "status": "rejected_or_out_of_scope", "decision_rationale": "WMPA learned statespaceworldmodel arbitrate frozen goal-conditioned continuouscontrolpolicies，18 OGBench state datasets；不是LLM/foundation model后训练，portfolio/value机制可参考但当前paperimplementation outofscope。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10932v1", "sha256": "e4db227678e0cbc1048b6f1b10ff5fba0e6612f0bc9eb91b377e1cd5ef35add6", "local_html_sha256": "e4db227678e0cbc1048b6f1b10ff5fba0e6612f0bc9eb91b377e1cd5ef35add6", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10932v1", "result": "retrieved", "sha256": "e4db227678e0cbc1048b6f1b10ff5fba0e6612f0bc9eb91b377e1cd5ef35add6" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method: World-Model Policy Arbiter", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/junwei0102/World-Model-Policy-Arbiter" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "当前任务范围不纳入", "matched_queries": [ "test-time-rl" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.10868", "title": "Conversational Voice Aesthetic Model with Reinforcement Learning from Human Listeners", "published": "2026-10-07T20:13:03Z", "status": "reviewed_deferred", "decision_rationale": "CVAM Qwen2.5-Omni voiceaesthetics SFT+humanlabelGRPO，objective pitch/pacing与subjective emotion/delivery区分，CANDOR human majority评测；下一批需真实音频与多listener训练reward，不把singlerater当确定gold。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10868v1", "sha256": "f8c082d200d9d6a87d479fc573d2846122f42869a9dcaa1845bb14963100324c", "local_html_sha256": "f8c082d200d9d6a87d479fc573d2846122f42869a9dcaa1845bb14963100324c", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10868v1", "result": "retrieved", "sha256": "f8c082d200d9d6a87d479fc573d2846122f42869a9dcaa1845bb14963100324c" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "3 Results", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-rl" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.10854", "title": "KDFP: A first-principles approach to knowledge distillation in large language models", "published": "2026-10-07T19:59:07Z", "status": "reviewed_deferred", "decision_rationale": "KDFP whiteboxgeneralknowledgeKD在continualpretrainPythia和9tasks对照、ephemeralparameterreduction提升训练效率；foundationKD下一批需查actual目标/参数临时冻结恢复，不能与instructionOPD混称。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10854v1", "sha256": "5c486aafb0efc59620272162a5403161289ee9e5ecabd1591d625e9631877548", "local_html_sha256": "5c486aafb0efc59620272162a5403161289ee9e5ecabd1591d625e9631877548", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10854v1", "result": "retrieved", "sha256": "5c486aafb0efc59620272162a5403161289ee9e5ecabd1591d625e9631877548" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "5 Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "6 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/TransformerLensOrg/TransformerLens" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "post-training" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.10536", "title": "Decoupling Exploration from Optimization in RLVR", "published": "2026-10-07T17:59:26Z", "status": "reviewed_deferred", "decision_rationale": "ExpDis separate explorer RND noveltyRL→correctness/qualityfilter→studentdistill→noNoveltyRL多轮，math同wallclockDAPO和passk比较；下一批需真实separatepolicies/RND和过滤成本，不能常量noveltyreward。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10536v1", "sha256": "22869b7a7c4ec36ef4ffd76c33fa33d0cfef6d29c16772a09b65b62d0b190d88", "local_html_sha256": "22869b7a7c4ec36ef4ffd76c33fa33d0cfef6d29c16772a09b65b62d0b190d88", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10536v1", "result": "retrieved", "sha256": "22869b7a7c4ec36ef4ffd76c33fa33d0cfef6d29c16772a09b65b62d0b190d88" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experimental Setup", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/SaifPunjwani/Exploration-Distillation", "https://huggingface.co/SaifPunjwani/expdis-checkpoints" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-rl", "on-policy-distillation", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.10426", "title": "CoTrace: Data Recipes for Training Terminal Agents with Harness-Model Co-Evolution", "published": "2026-10-07T17:06:59Z", "status": "reviewed_deferred", "decision_rationale": "CoTrace componentwisepromotion的harness/model交替loop，以runtimeprovenance匹配verifiedSFT或freshRL，Tmax promotion/evolve/reservoir互斥与OODtransfer；下一批需真实terminalcontainers、控制器execute/evaluate闭环。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10426v1", "sha256": "8fd0c440501cf7ec0672b97cb4d8b173566b3b5dfd3e76a6fe76d0126d2c99d5", "local_html_sha256": "8fd0c440501cf7ec0672b97cb4d8b173566b3b5dfd3e76a6fe76d0126d2c99d5", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10426v1", "result": "retrieved", "sha256": "8fd0c440501cf7ec0672b97cb4d8b173566b3b5dfd3e76a6fe76d0126d2c99d5" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "3 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-rl" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.10422", "title": "Which Rollout Taught It That? BehaviorTrace and the Limits of Training-Data Attribution in Online RL", "published": "2026-10-07T17:01:37Z", "status": "reviewed_deferred", "decision_rationale": "BehaviorTrace GRPO plantedtrigger causalgroundtruth+fullgradient sketching对比norm/fluency/headroom/seedcontrols；作者不提新attribution estimator，下一批diagnostic harness不得把plantedlabel提供policy或当能力提升。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10422v1", "sha256": "468af7a5c0251545333a035b907a9718b8070849b49c1347dc90a2bca220488d", "local_html_sha256": "468af7a5c0251545333a035b907a9718b8070849b49c1347dc90a2bca220488d", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10422v1", "result": "retrieved", "sha256": "468af7a5c0251545333a035b907a9718b8070849b49c1347dc90a2bca220488d" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Methods", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Results", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/AmitoVrito/BehaviorTrace" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-rl", "test-time-rl" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.10411", "title": "Training Parallel Speculative Draft Models by Directly Minimizing Expected Decoding Rounds", "published": "2026-10-07T16:56:29Z", "status": "reviewed_deferred", "decision_rationale": "EDR用speculativedecoding Markovrewardprocess stateoccupancy精确expectedrounds，TDgradient训parallelDSpark/DFlydraft；九benchmark下一批需真实rollout/drafter与decoderoundevaluator，理论MainResults不代替实测。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10411v1", "sha256": "38bb29d63922e2068c4a8561287cbc2603d2acf7cd85ab281bc91c8272c53174", "local_html_sha256": "38bb29d63922e2068c4a8561287cbc2603d2acf7cd85ab281bc91c8272c53174", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10411v1", "result": "retrieved", "sha256": "38bb29d63922e2068c4a8561287cbc2603d2acf7cd85ab281bc91c8272c53174" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Main Results", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "5 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/tatsu-lab/stanford_alpaca", "https://github.com/y-x-zhao/AngelSpec-EDR" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "process-reward", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.10092", "title": "I would rather quit NLP than read another paper like this: The rise of antithesis in NLP papers", "published": "2026-10-07T13:50:50Z", "status": "rejected_or_out_of_scope", "decision_rationale": "ACL/arXiv写作antithesis频率、双annotator和rewardmodel偏好研究，posttraining只作机制猜测；无具体训练算法可实施，当前batch outofscope，可作为写作偏差背景。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10092v1", "sha256": "16de4adb67313eced254c8e2f2047a9e46b1aecb275bc28bdff469a58efe9715", "local_html_sha256": "16de4adb67313eced254c8e2f2047a9e46b1aecb275bc28bdff469a58efe9715", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10092v1", "result": "retrieved", "sha256": "16de4adb67313eced254c8e2f2047a9e46b1aecb275bc28bdff469a58efe9715" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Data", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "5 Results", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/olzama/rather-than-public" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "当前任务范围不纳入", "matched_queries": [ "post-training", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.09934", "title": "Itgan at NADI 2026 shared task: Parameter-Efficient Whisper Adaptation for Robust, Mixed-Dialect and Code-Switched Arabic ASR", "published": "2026-10-07T12:18:40Z", "status": "reviewed_deferred", "decision_rationale": "NADI WhisperwideLoRA Arabicdialect/ code-switch ASR、specialistrouting/weightspaceaverage/ROVER有bootstrap与negative results；foundation speechadaptation下一批，oracle dialect routing需标注不可用label的postevalprobe。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09934v1", "sha256": "f2ceccf28e1f629e8596b3a89ad83598551bfcda3fe2a246267f3f2d4d0f95ec", "local_html_sha256": "f2ceccf28e1f629e8596b3a89ad83598551bfcda3fe2a246267f3f2d4d0f95ec", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09934v1", "result": "retrieved", "sha256": "f2ceccf28e1f629e8596b3a89ad83598551bfcda3fe2a246267f3f2d4d0f95ec" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Common Setup", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Subtask 1.1: Robust Country-Level ASR", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/SYSTRAN/faster-whisper", "https://github.com/huggingface/peft", "https://github.com/jitsi/jiwer/releases/tag/v4.0.0", "https://github.com/snakers4/silero-vad" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "post-training" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.09835", "title": "A Deafening Silence: Catastrophic Forgetting Lives in the Output Embeddings of Tokens the Data Never Speaks", "published": "2026-10-07T11:01:01Z", "status": "reviewed_deferred", "decision_rationale": "Absenttoken outputhead one-sidedsoftmaxgrad受Adamvhat放大，head-onlyepsilon提高防forgetting并可加replay；多scale/familycontrolledfreezing与posthoc失败显示需training-time，下一批真实outputembedding optimizergroup和riskcounts。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09835v1", "sha256": "428548747a95624eaa45472690071a538d6d77eabe174d62f6b7348030990a3c", "local_html_sha256": "428548747a95624eaa45472690071a538d6d77eabe174d62f6b7348030990a3c", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09835v1", "result": "retrieved", "sha256": "428548747a95624eaa45472690071a538d6d77eabe174d62f6b7348030990a3c" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "5 The Intervention and Its Boundaries", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "5 The Intervention and Its Boundaries", "scope": "additional eight-setting intervention results excerpt", "characters_read_at_least": [ 1300, 3800 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "post-training" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.09804", "title": "BoT-GRPO: Efficient Process-Reward RL for Reasoning via Bag-of-Token Aggregation", "published": "2026-10-07T10:23:17Z", "status": "reviewed_deferred", "decision_rationale": "BoT-GRPO按source-sequencelength逆权聚合tokenrewards/weightedgroupstats、criticfree；Reactcompile/VLMjudge和AIME对照，下一批需真实稳定tokenRM及compileexecutions，防cosmeticrewardhacking。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09804v1", "sha256": "7dfe9b575388eac9aab6d07d0f136e7952342d0c13f85bd9c20265d27806981a", "local_html_sha256": "7dfe9b575388eac9aab6d07d0f136e7952342d0c13f85bd9c20265d27806981a", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09804v1", "result": "retrieved", "sha256": "7dfe9b575388eac9aab6d07d0f136e7952342d0c13f85bd9c20265d27806981a" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://huggingface.co/blog/smollm3" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-rl", "process-reward", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.09795", "title": "MIRROR: From Imitation to Internalization in LLM Personalization", "published": "2026-10-07T10:10:01Z", "status": "reviewed_deferred", "decision_rationale": "MIRROR reference-conditioned frozen selfteacher对studentrollouts分布蒸馏，MIRRORF informative-referencetoken focalanchor；LaMP/LongLaMP+OOD个性化，训练privilege合法但test不得读referencewording，下一批。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09795v1", "sha256": "dc63b1f4486fbe4e59a155691c3cacb6fe7ea43faaa96effa7af7fdbc3850c28", "local_html_sha256": "dc63b1f4486fbe4e59a155691c3cacb6fe7ea43faaa96effa7af7fdbc3850c28", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09795v1", "result": "retrieved", "sha256": "dc63b1f4486fbe4e59a155691c3cacb6fe7ea43faaa96effa7af7fdbc3850c28" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "on-policy-distillation" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.10623", "title": "Recurrent Self-Improvement: Dynamic Cross-Loop On-Policy Distillation for Looped Language Models", "published": "2026-10-07T08:54:30Z", "status": "reviewed_deferred", "decision_rationale": "LoopOPDterminalcomputeloop frozen teacher指导intermediateloop student trajectories，dynamic版按sharedparameters刷新teacher；OuroThinking DAPO与general/code评测，下一批必须真loopbackbone/dynamicweights而非两个不同名普通LM。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10623v1", "sha256": "829d72fd3aaf1e4f2db3e733b2715d91418a0086979a1f34d669f56050c7a870", "local_html_sha256": "829d72fd3aaf1e4f2db3e733b2715d91418a0086979a1f34d669f56050c7a870", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10623v1", "result": "retrieved", "sha256": "829d72fd3aaf1e4f2db3e733b2715d91418a0086979a1f34d669f56050c7a870" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "4 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "5 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://[https://huggingface.co/datasets/ByteDance-Seed/BeyondAIME(https://huggingface.co/datasets/ByteDance-Seed/BeyondAIME)" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "on-policy-distillation", "post-training", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.09684", "title": "From Pareto to Preference: Personalized Test-Time Scaling via Amortized Agentic Policy Discovery", "published": "2026-10-07T08:46:56Z", "status": "reviewed_deferred", "decision_rationale": "PersonTTS executablecontroller discovery按accuracy/latency/costjoint满足度，requirement-matched经验reuse与frozenGuide，AIME/HMMT source/targetprofiles和heldoutproblems；下一批controller须actualexecution/evaluation，replayrecords不能泄漏test答案。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09684v1", "sha256": "a03218b24a19a33b6a272be8fd51b3c11e010361e13acd3d991a86ca46c42f47", "local_html_sha256": "a03218b24a19a33b6a272be8fd51b3c11e010361e13acd3d991a86ca46c42f47", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09684v1", "result": "retrieved", "sha256": "a03218b24a19a33b6a272be8fd51b3c11e010361e13acd3d991a86ca46c42f47" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/WangXinglin/PersonTTS" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "on-policy-distillation", "preference-optimization" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.09665", "title": "SAPD: Step-Aligned Privileged Distillation", "published": "2026-10-07T08:31:37Z", "status": "reviewed_deferred", "decision_rationale": "SAPD rolloutfree固定demonstrations按reasoningstep对privileged信息提供context-dependent distributiontargets，Qwen3-4B math/OODeval和loop-speed对照；下一批真实offline teacherdistribution，gold仅traincontext允许。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09665v1", "sha256": "9ee60d653a0c3c142bc45011fb59982ba7eca7312172b04251cc0d0633f718df", "local_html_sha256": "9ee60d653a0c3c142bc45011fb59982ba7eca7312172b04251cc0d0633f718df", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09665v1", "result": "retrieved", "sha256": "9ee60d653a0c3c142bc45011fb59982ba7eca7312172b04251cc0d0633f718df" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Step-Aligned Privileged Distillation", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/Miaow-Lab/SAPD", "https://github.com/siyan-zhao/OPSD" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-rl", "on-policy-distillation", "post-training" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.09639", "title": "On-Policy Distillation Teaches New Skills but Not New Knowledge", "published": "2026-10-07T08:15:10Z", "status": "reviewed_deferred", "decision_rationale": "控制randomfacts/compositionalstructures独立teacher知识/技能，reverse/forwardKL和studentrolloutfactorial分离OPDtransfer；结论限制四model实验，下一批diagnostic synthetic训练不能goldtable进evaluatedrollout。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09639v1", "sha256": "80a3d7103e7a5f4bd6d303213468cd412b47385f08cbd0181a338489a15e023d", "local_html_sha256": "80a3d7103e7a5f4bd6d303213468cd412b47385f08cbd0181a338489a15e023d", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09639v1", "result": "retrieved", "sha256": "80a3d7103e7a5f4bd6d303213468cd412b47385f08cbd0181a338489a15e023d" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Experimental Setting: Separating Knowledge and Skill", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Synthetic Experiment Results", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "5 Real-World Validation", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://[https://huggingface.co/AI-MO/NuminaMath-CoT](https://github.com/project-numina/aimo-progress-prize/blob/main/report/numina_dataset.pdf)" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "on-policy-distillation" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.09597", "title": "COPC: Coupled Off-Policy Correction for Asynchronous LLM Reinforcement Learning", "published": "2026-10-07T07:43:49Z", "status": "reviewed_deferred", "decision_rationale": "COPC协调tokenratiomasking与TDresidual two-sidedclippedreturns，policy/advantagestaleness耦合，math/search异步64stepstale与synchronousspeed对照；下一批需真实actorcritic/stale rollout和jointval参数选择。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09597v1", "sha256": "f3db33889b29fb0954fa0bf94e17eab1943fc7ea5caf580217ae73ac7a1c6a48", "local_html_sha256": "f3db33889b29fb0954fa0bf94e17eab1943fc7ea5caf580217ae73ac7a1c6a48", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09597v1", "result": "retrieved", "sha256": "f3db33889b29fb0954fa0bf94e17eab1943fc7ea5caf580217ae73ac7a1c6a48" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "5 COPC: Coupled Off-Policy Correction", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 1300 ] }, { "heading": "6 experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-rl", "post-training" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.09587", "title": "Collaborative Reasoning Distillation via Cross-Feedback and Coherent Curation", "published": "2026-10-07T07:31:33Z", "status": "reviewed_deferred", "decision_rationale": "CRD multi-teachercrosscritique、stepqualityjudge、coherentstitch/RQO预算curriculum，50k math/code/logic/STEM及contamination检查；下一批需真实judge/教师生成与coherence操作，保留benchmarkclaims可验证边界。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09587v1", "sha256": "a897e8e1a5618bdf7ff1d97f9b45dcc6f1626b5e07954624127cdad58f910b74", "local_html_sha256": "a897e8e1a5618bdf7ff1d97f9b45dcc6f1626b5e07954624127cdad58f910b74", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09587v1", "result": "retrieved", "sha256": "a897e8e1a5618bdf7ff1d97f9b45dcc6f1626b5e07954624127cdad58f910b74" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "3 Evaluation", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.09563", "title": "EvoSignal: LLM-Guided Evolutionary Design of Modular Traffic Signal Control Programs", "published": "2026-10-07T07:04:22Z", "status": "reviewed_deferred", "decision_rationale": "EvoSignalLLM演化modulartrafficcontrol程序，以拥堵反馈保存多objective tradeoffs并zeroshotnetworktransfer；agent-programsearch下一批需实际CityFlowsimulator/程序执行与heldoutflow，不能只有registry标签。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09563v1", "sha256": "7b1ee8ae5f784aef00d1e3ded614c2ac4ddd5a181b4d6158c3815023f07f3439", "local_html_sha256": "7b1ee8ae5f784aef00d1e3ded614c2ac4ddd5a181b4d6158c3815023f07f3439", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09563v1", "result": "retrieved", "sha256": "7b1ee8ae5f784aef00d1e3ded614c2ac4ddd5a181b4d6158c3815023f07f3439" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/algorithmicsuperintelligence/openevolve", "https://github.com/georgewanglz2019/EvoSignal" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-rl" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.09384", "title": "The Persona Hierarchy Model: Understanding Contextual Generalization in Fine-Tuning LLMs", "published": "2026-10-07T03:40:44Z", "status": "reviewed_deferred", "decision_rationale": "Persona Hierarchy/PPR在reference persona下约束initial-model行为，120个prefix×behavior模型/heldout评测及SFT/RL应用；Google DeepMind线索为作者当前去向而非本研究官方机构，优先审查但不夸大Google论文署名；下一批PPR需跨prefix真实KL与reward-hacking评测。", "priority_reviewed": true, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09384v1", "sha256": "aec40368df6af778ccaafc5b1e31a8d443c369f8aa7ecb6812e0c0066c5084a6", "local_html_sha256": "aec40368df6af778ccaafc5b1e31a8d443c369f8aa7ecb6812e0c0066c5084a6", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09384v1", "result": "retrieved", "sha256": "aec40368df6af778ccaafc5b1e31a8d443c369f8aa7ecb6812e0c0066c5084a6" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 Persona Hierarchy Model", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "6 Persona-preserving regularization", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "3 Experimental Setup", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 The narrowness of contextual generalization is correlated with the proximity with default persona", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/andotalao24/Persona_hierarchy_model" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.09337", "title": "Ask the Expert: LLM-Guided Reinforcement Learning for Autonomous Cyber Defense", "published": "2026-10-07T02:53:26Z", "status": "reviewed_deferred", "decision_rationale": "AskExpert LLM对hardcyberdefensestates给受限host建议转换tieredrewardshaping训PPO，deployment不依赖LLM，CAGE1/2benigndefensebenchmark；下一批需可审计限定actions与真实simrollouts，属于LLMguidance非语言权重训练。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09337v1", "sha256": "60160fae8bf8bb76cdbd0f47a7022977e7b21c311fdcd21104ba21f37f580c65", "local_html_sha256": "60160fae8bf8bb76cdbd0f47a7022977e7b21c311fdcd21104ba21f37f580c65", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09337v1", "result": "retrieved", "sha256": "60160fae8bf8bb76cdbd0f47a7022977e7b21c311fdcd21104ba21f37f580c65" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "IV Ask the Expert Framework", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "V Experimental Evaluation", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "llm-rl", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.09247", "title": "An Informational Curse of Horizon in Goal-Conditioned Policy Learning", "published": "2026-10-07T00:20:45Z", "status": "rejected_or_out_of_scope", "decision_rationale": "goal-conditionedBC/RL longhorizon informationalcurse与inputJacobian Sobolevbootstrapping在OGBenchcontinuouscontrol和oracleplanners诊断；无LLM/foundation/语言agent后训练，当前implementation scope外。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09247v1", "sha256": "5836a8230b01dc8a1fea3eecfda5f8bdc6f76b16fd6965d8865a2963fb2eeb83", "local_html_sha256": "5836a8230b01dc8a1fea3eecfda5f8bdc6f76b16fd6965d8865a2963fb2eeb83", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09247v1", "result": "retrieved", "sha256": "5836a8230b01dc8a1fea3eecfda5f8bdc6f76b16fd6965d8865a2963fb2eeb83" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 An Informational Curse of Horizon", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "5 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "当前任务范围不纳入", "matched_queries": [ "test-time-rl" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.09227", "title": "Quantize by Drift: Label-Free Mixed-Precision Post-Training Quantization for Text Embedders", "published": "2026-10-06T23:44:32Z", "status": "reviewed_deferred", "decision_rationale": "Textembedder label-freemodulequantization输出drift作hardpackedbytebudgetallocation，five development/three sealedheldoutbackbones；mainbudget劣于uniform，下一批必须保negativeresults、真实embedding重算/bitbudget，不能把drift减少当retrieval普遍提升。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09227v1", "sha256": "b677aa9f10d68c9e54377eb1d876b84f5239ce5babbc7410bcf6a0bdc68ea1ee", "local_html_sha256": "b677aa9f10d68c9e54377eb1d876b84f5239ce5babbc7410bcf6a0bdc68ea1ee", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09227v1", "result": "retrieved", "sha256": "b677aa9f10d68c9e54377eb1d876b84f5239ce5babbc7410bcf6a0bdc68ea1ee" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "2 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "3 Experimental setup", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "post-training" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.09209", "title": "Multi-Objective Aligned Small Language Model Framework for SUD Patient Dialogue Generation", "published": "2026-10-06T23:10:21Z", "status": "reviewed_deferred", "decision_rationale": "SUDdialogue用CCD认知component检测→CADG KD/SFT/preferences/attentionreward，数据来自publicReddit personalnarratives再构patienthistories；下一批需医疗伦理/隐私审查与不臆造临床真实度，不能仅LLMhitmetric当therapy有效。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.09209v1", "sha256": "347e6cc18f9c3f316a6cd31ad668e0cb635de16ecf3be48222030e91506ded4d", "local_html_sha256": "347e6cc18f9c3f316a6cd31ad668e0cb635de16ecf3be48222030e91506ded4d", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.09209v1", "result": "retrieved", "sha256": "347e6cc18f9c3f316a6cd31ad668e0cb635de16ecf3be48222030e91506ded4d" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "III Methodology", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "IV Evaluation", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/thusharamanjari/Multi_objective_aligned_SLM" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "preference-optimization", "reward-model" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.10592", "title": "Wieszcz-XIX: A 3.1-Billion-Word Corpus of Pre-1918 Polish and Temporally Bounded Language Models Trained From Scratch", "published": "2026-10-06T18:35:26Z", "status": "reviewed_deferred", "decision_rationale": "WieszczXIX historicalPolishcorpus documentdedup/时间泄漏审计与47-349Mfromscratchmodel ladder，denseheldouttemporalBPB；foundationpretrain候选nextbatch，OCR45%不可纠正和残余latertext须保留，非posttrainingRL。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.10592v1", "sha256": "6dfce9a2d889083032063994c46b556b89141c6ea204c47e11b1f95e7f7b77db", "local_html_sha256": "6dfce9a2d889083032063994c46b556b89141c6ea204c47e11b1f95e7f7b77db", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.10592v1", "result": "retrieved", "sha256": "6dfce9a2d889083032063994c46b556b89141c6ea204c47e11b1f95e7f7b77db" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Corpus construction", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "7 Evaluation", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [ "https://github.com/SKocur/wieszcz-xix", "https://github.com/haykgrigo3/TimeCapsuleLLM", "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B", "https://huggingface.co/SKocur/wieszcz-xix-349m", "https://huggingface.co/datasets/SKocur/polish-pre1918-corpus", "https://huggingface.co/flax-community/papuGaPT2", "https://huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512", "https://huggingface.co/zai-org/GLM-5.2" ], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "post-training" ], "tracks": [ "post-training" ] }, { "arxiv_id": "2610.12474", "title": "TTS-Q: Perception-Guided Post-Training Quantization of Text-to-Speech Diffusion Transformers", "published": "2026-08-11T00:27:07Z", "status": "reviewed_deferred", "decision_rationale": "TTSQ taskmatchedlossmaxcalibration与lateFFNactivationmixedprecisionbudgetbalance改善speakerSIM，1B/3.5B speechDiT W4A4及peakmemory；下一批需真实speechsamples、nominalbits与actualmemory区分，不仅WER假设音色无损。", "priority_reviewed": false, "verified_organization": null, "source": { "url": "https://arxiv.org/html/2610.12474v1", "sha256": "dcc7a6c803e277e5ac853a7a70ef5ac679655b959acde73c2e46c93a1698405c", "local_html_sha256": "dcc7a6c803e277e5ac853a7a70ef5ac679655b959acde73c2e46c93a1698405c", "local_html_sha256_matches": true, "attempts": [ { "url": "https://arxiv.org/html/2610.12474v1", "result": "retrieved", "sha256": "dcc7a6c803e277e5ac853a7a70ef5ac679655b959acde73c2e46c93a1698405c" } ], "alternate_evidence": null }, "read_scope": "abstract_plus_bounded_method_and_evaluation_excerpts", "full_article_read": false, "evidence_locations": [ { "heading": "3 Method", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] }, { "heading": "4 Experiments", "scope": "bounded section excerpt, not the full section", "characters_read_at_least": [ 0, 650 ] } ], "reviewed_abstract": true, "fulltext_unresolved": false, "upstream_code_release_status": "not independently audited in this closure; links in source may be dependencies/baselines", "source_observed_code_links": [], "repository_status_at_assignment": "new", "new_implementation_claimed": false, "next_action": "冻结 16 篇实施之外，保留为下一批候选；未实施", "matched_queries": [ "post-training" ], "tracks": [ "post-training" ] } ], "pdf_followup": { "checked_on": "2026-10-12", "note": "官方HTML 404不等于无全文；两项经官方PDF恢复方法/评测限定审查，保留HTML失败尝试和精确PDF证据；full_articles_read仍为0。" } }