{ "schema_version": 1, "reviewed_on": "2026-10-15", "track": "post-training", "reviews": [ { "id": "2610.20597", "title": "Co-Evolving Acting and Forecasting for Language Agents", "published": "2026-10-14T17:59:28Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "daaf74202e484916259563cf5ccb43749ad1998c8244b340c7c03b2219e2b601", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "CoPE 以成功 on-policy 轨迹的多步后续动作监督同一 backbone，不能替换成固定成功动作表。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§5 CoPE multi-step forecasting", "§6 Experimental Settings", "Appendix B.2 Training Hyperparameters" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.20597v1", "source_sha256": "41d3b2bfca8efafbddf4111b8eaa1ccd729a243e45d25fafc64bf2ba2abda697", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.20597v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。Qwen2.5-7B/14B，五个长时程环境，策略训练与多步预测辅助目标；完整多任务训练总时未核定", "scale": "Qwen2.5-7B/14B，五个长时程环境，策略训练与多步预测辅助目标；完整多任务训练总时未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.20576", "title": "One Task, Many Interfaces: Simple Cross-Interface Consistency Training Improves LLM Agent Robustness", "published": "2026-10-14T17:56:06Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "208086140b8aca99abfab5a56ea28eac26b6dc5f9055b9449de7e3820801281f", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "CIPR 对去语义工具 ID、孪生接口和等价动作分布施加一致性 KL；原始双接口训练不能只用一个 KL 数值样例冒充。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 Cross-Interface Policy Regularization", "Hyperparameters" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.20576v1", "source_sha256": "e9f83a17be7fc69b963ffeaa7bdc13f6a5b7bd98398de1825ea479be3087a0b4", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "source_attempts": [ { "url": "https://arxiv.org/html/2610.20576v1", "result": "retrieved", "sha256": "e9f83a17be7fc69b963ffeaa7bdc13f6a5b7bd98398de1825ea479be3087a0b4" } ], "cost_assessment": { "source_url": "https://arxiv.org/html/2610.20576v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。Qwen2.5-3B/Qwen3.5-4B，3,920例，15 epochs、约105更新，每接口4 rollouts及双路重评分；总运行时未披露", "scale": "Qwen2.5-3B/Qwen3.5-4B，3,920例，15 epochs、约105更新，每接口4 rollouts及双路重评分；总运行时未披露", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.20572", "title": "Recovering Reliable Supervision via Critique-Guided On-Policy Distillation", "published": "2026-10-14T17:54:34Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "6acad8623f1e79ee2a220f8126bbabd9b9b221fa7487106837932b0eeae52cd2", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "CriticOPD 由教师批评错误轨迹、学生重写，并仅由带特权 critique 的教师评分；学生推理不接收 critique。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 CriticOPD", "Appendix C.1–C.2", "Appendix D Computational Cost" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.20572v1", "source_sha256": "f14b87e88e0c21648936b5455c8a6bfb6f9e73b67feb5b754e9aa98aedcf29a9", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.20572v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。0.6B/1.7B学生、4B教师；每训练4 A100（学生2/教师2），1.7B报告17.2 GPU-hours；长回答与8–32次评测采样", "scale": "0.6B/1.7B学生、4B教师；每训练4 A100（学生2/教师2），1.7B报告17.2 GPU-hours；长回答与8–32次评测采样", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.20538", "title": "PrismaCoder: Pushing the Boundaries of Open Multimodal Code Intelligence through Decomposed Rubric Synthesis", "published": "2026-10-14T17:32:54Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "9a12a2af5bb09b7abdc31b03f9fc16f863e42782c189535a152c567c1c67ec46", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "PrismaForge 将分解 rubric 同时用于多模态代码数据过滤和 RL reward，PrismaCoder 包含完整数据/训练链路。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 PrismaForge", "§4 PrismaCoder-2M", "Appendix F Training Details" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.20538v1", "source_sha256": "e9092f8023f12004d5c248496dd5ce69b9b8534e06e886641cdd480f5f5b06c8", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.20538v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。PrismaCoder-2M数据，9B/35B模型，4节点×8 H200=32 GPUs，全参数SFT与RL", "scale": "PrismaCoder-2M数据，9B/35B模型，4节点×8 H200=32 GPUs，全参数SFT与RL", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.20516", "title": "VASPO: Self-Distillation with Privileged Value Estimation", "published": "2026-10-14T17:17:18Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "6e5d4cba5fe13d9788809df22e6d401ebe30570fe3d222f8e4cdf714861a431c", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "VASPO 让 privileged critic 与自教师在同一特权上下文评价学生前缀，以有界概率比调节部分正 PPO advantage。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 VASPO", "Hyperparameter Record" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.20516v1", "source_sha256": "c096e8aef4eb876e1808c56b1a12292e2ba6f31d1b4c4d1b6f6d5c23617e27a3", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.20516v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。4B/7B actor、critic/self-teacher；8 H200，全参数更新，200 rollout updates，每prompt8响应", "scale": "4B/7B actor、critic/self-teacher；8 H200，全参数更新，200 rollout updates，每prompt8响应", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.20497", "title": "Hindsight-Privileged RLVR: Single-Rollout Token Credit Without Group Sampling", "published": "2026-10-14T17:06:59Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "764ac63b4dc7369163c65f2ad022887b76579ea513fea636a940d1645259c93b", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "HP-RLVR 用生成快照在因果/后见上下文的 JS 差构造单轨迹 token credit，无参考解；相对 GRPO 更快不能替代成本门槛。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 Hindsight-Privileged RLVR", "§4.1 Experimental setup", "§4.2 Cost of replacing group sampling" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.20497v1", "source_sha256": "46b185fb3b5bee63e1206286936d843f13327d9b718f61486faf34984774674b", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.20497v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。Qwen3-1.7B/4B/8B，三任务100-step对照；原文计时为8 Ascend 910B NPUs，非本地NVIDIA运行", "scale": "Qwen3-1.7B/4B/8B，三任务100-step对照；原文计时为8 Ascend 910B NPUs，非本地NVIDIA运行", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.20413", "title": "When Does Momentum Help in Reinforcement Learning with Verifiable Rewards?", "published": "2026-10-14T16:11:14Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "bec5c89eb4810be609f5356ab4f5e3d3cb54b0513421a7e1b107c71447eec2ba", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "accepted", "priority": "P1", "resolution": "affordable_candidate_not_implemented", "reason": "只接收论文原始十臂 bandit 的 bias-corrected EMA/安全 η/(1−β) 机制和解析性质检验，可用 CPU 数值运算保留原目标；不接收4B/32B RLVR或50,000-replicate完整统计活动，不据此声称LLM提升；接受为低成本审查候选，本轮未实现、未实验，亦不自动进入执行/evolve。", "locations": [ "§3–§5 bandit model and momentum results", "§6.1 bandit validation", "Appendix I.1 Bandit results" ], "fulltext_review_completed": true, "review_level": "primary_method_and_affordability_review_bounded_scope", "source_url": "https://arxiv.org/html/2610.20413v1", "source_sha256": "91d74ddf092c63286e49fad0e7acf8139421312144f24852711e5f04d3bfa337", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.20413v1", "basis": "只接收下列公开原始机制的有界范围，非整篇多模型/多臂研究复现；预算判断依据原文配置而非本地运行。十臂softmax/KL-regularized bandit，T=32…4096；廉价范围限原始机制及有限数值核验，非全套50,000次重复", "scale": "十臂softmax/KL-regularized bandit，T=32…4096；廉价范围限原始机制及有限数值核验，非全套50,000次重复", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.20401", "title": "Training Language Models to Fail Gracefully", "published": "2026-10-14T16:03:12Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "5f189dd4e678c3601a16529e34519bec0d581bc663233b227ae77847d4c64337", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "GracefulRL 以 success/fallback/error 的 log-barrier/TailRL 目标和 smoothing 避免仅fallback坍缩；不能用手写回退规则替代策略学习。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4.1 GracefulRL algorithm", "§5.1 SafeMaze", "Appendix D.7 more training compute", "Acknowledgements compute resources" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.20401v1", "source_sha256": "70c604cc7277d176ed821a68e1713ab0675efde463d18ab12300e9172a131bd4", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "source_attempts": [ { "url": "https://arxiv.org/html/2610.20401v1", "result": "retrieved", "sha256": "70c604cc7277d176ed821a68e1713ab0675efde463d18ab12300e9172a131bd4" } ], "cost_assessment": { "source_url": "https://arxiv.org/html/2610.20401v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。SafeMaze 20k训练步及多reward/rollout超参，Qwen3-4B的200/400/800步×五fallback rewards，另Minesweeper；项目使用GH200，完整总时未核定", "scale": "SafeMaze 20k训练步及多reward/rollout超参，Qwen3-4B的200/400/800步×五fallback rewards，另Minesweeper；项目使用GH200，完整总时未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.20318", "title": "Efficient Test-Time Adaptation For Robot Policies", "published": "2026-10-14T15:05:24Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "85d3177a7ba891fb7b4bd3a5c3a493732a6ca8d281c6fe49defd772265b8853b", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "non_llm_robot_control_out_of_scope", "reason": "TEMPO 是 Go1/Go2/G1 机器人运动策略的无奖励 test-time adaptation，包含实体Go2验证，不是本队列的LLM偏好/蒸馏/语言策略后训练；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "official arXiv abstract: scope triage" ], "fulltext_review_completed": false, "review_level": "official_abstract_scope_triage", "source_url": "https://arxiv.org/abs/2610.20318v1", "source_sha256": null, "evidence_limitations": "决定仅基于官方元数据/摘要范围或明确的来源例外；未核验正文算法或资源，不以metadata hash冒充文档hash。" }, { "id": "2610.20303", "title": "SpectrumQuant: Spectrum Alignment for LLM Quantization", "published": "2026-10-14T14:54:16Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "09ad0199d9208e3a8b9df801207a4c2062c8c131f11b0cc0e00e1d02c191f99c", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "accepted", "priority": "P1", "resolution": "affordable_candidate_not_implemented", "reason": "SpectrumQuant 用校准激活协方差构造闭式 alignment+mixing 正交旋转；原文7B校准约0.26 GPU小时、4.51 GiB，可保留128条真实校准样本的7B原算法路径；不承诺70B或全套多模型下游扫评；接受为低成本审查候选，本轮未实现、未实验，亦不自动进入执行/evolve。", "locations": [ "§4 covariance alignment and rotation construction", "§5.1 Implementation Details", "§5.2 Main Results cost table" ], "fulltext_review_completed": true, "review_level": "primary_method_and_affordability_review_bounded_scope", "source_url": "https://arxiv.org/html/2610.20303v1", "source_sha256": "d7ef15bfb21f18c495103578351d0a5fa3106461b54687eb60831113e06ea03a", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.20303v1", "basis": "只接收下列公开原始机制的有界范围，非整篇多模型/多臂研究复现；预算判断依据原文配置而非本地运行。7B原始校准：128 WikiText2样本×2048 tokens；作者报告0.26 GPU-hours、4.51 GiB峰值（不是本地测量）", "scale": "7B原始校准：128 WikiText2样本×2048 tokens；作者报告0.26 GPU-hours、4.51 GiB峰值（不是本地测量）", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.20293", "title": "CASI: Credit Assignment via Set Intervention for Agentic Retrieval", "published": "2026-10-14T14:47:46Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "f8325f572d0c5d0561946243bf10a0620c04ad3609ed1f9265d9f200239f6315", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P0", "resolution": "not_reproduced_high_cost", "reason": "CASI 经固定检索快照的集合干预、Shapley归因与有界零和优势投影分配信用；§4.5确有随机七天10%生产流量A/B、Advv +3.51%，工业证据通过，但不等于公开可复现。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3.2 snapshot intervention and audit boundary", "§3.5 Computational Cost", "§4.5 Online A/B Evaluation" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.20293v1", "source_sha256": "b9d5f5ff912eb7a126b05d29030d3f633777783317fb2bd942cb912ec0e94ee7", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.20293v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。五动作时31个coalition scores，1,200请求生产回放审计、五次训练；生产回放、冻结evaluator及完整训练配置未公开，单卡总成本无法廉价核定", "scale": "五动作时31个coalition scores，1,200请求生产回放审计、五次训练；生产回放、冻结evaluator及完整训练配置未公开，单卡总成本无法廉价核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.20174", "title": "RM-OPD: On-Policy Distillation for Generative Reward Models", "published": "2026-10-14T13:20:55Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "cc31cfc5b1fc144510d8e0373a9fcdc40f805b4635fcfcefc760df1ea3f09fe2", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "RM-OPD 先以单领域GRPO培养教师，再路由OPD与领域贡献重加权整合生成式奖励模型。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§5 RM-OPD", "§6.1 Experimental setup" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.20174v1", "source_sha256": "e3e2b330b7beb2940bfea379e1a736ad1fb312b932b073f4da45f04e346cd4e9", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.20174v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。74k偏好、7B教师/学生；训练2节点×8 H800，评测另8 H800", "scale": "74k偏好、7B教师/学生；训练2节点×8 H800，评测另8 H800", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19920", "title": "MolHop: Scaffold-Hopping Molecular Design with Agentic Reinforcement Learning", "published": "2026-10-14T10:11:10Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "61c23cca43eb84e5bf014023d3fd843f5150cb8273527320eec3ae8d5814c35a", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "MolHop 在约160万scaffolds层次空间内导航/编辑，并通过Agentic RL训练跨骨架发现策略。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 Methods", "Appendix B.3 SFT and Agentic RL Training" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19920v1", "source_sha256": "4c440b650eec4e67b8a2d8c63042a04b6f1d0b726a00569f98a33c5fbaa01841", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19920v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。三模型规模，共享4 A800 80GB配置，SFT约3小时、RL约18小时；9B主要模型", "scale": "三模型规模，共享4 A800 80GB配置，SFT约3小时、RL约18小时；9B主要模型", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19832", "title": "Rent-Seeking Language Models: Reward Models Set the Prices, Frequent Audits Curb the Rent", "published": "2026-10-14T09:06:00Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "c69a1bcfcb786829fcabddcf85c9a397b0e10bf6980f73ea50ff9b4e7033d235", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Rent-Seeking 以proxy/gold reward和审计频率研究RL过优化，须保留固定测试床、同准确率proxy及真实训练，不能只模拟租金标量。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 Framework", "§4 Experimental Setup", "RL. compute accounting" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19832v1", "source_sha256": "ed982842a3ac29ed5c41aa6579c2ae8caf47d93916702819f29ae96a1dbb5243", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19832v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。每run4 A100 80GB、600步7.4–8.8小时；全项目约1,280 A100 GPU-hours", "scale": "每run4 A100 80GB、600步7.4–8.8小时；全项目约1,280 A100 GPU-hours", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19827", "title": "MAPLE: A Multi-Axis Parametric Knowledge Extraction Lens for LLM Evaluation", "published": "2026-10-14T09:02:05Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "c335839599b5b9360ff394ae15879b0b003a11fcbb06ca08f102e2a7811ac9c3", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "evaluation_analysis_reference_only", "reason": "MAPLE 是知识提取恢复/破坏、gold-token rank及beyond-sampling的诊断评价框架，不是新后训练策略；可借鉴指标，但本轮不把评测论文登记成能力实现；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "official arXiv abstract: complete abstract scope review" ], "fulltext_review_completed": false, "review_level": "official_abstract_scope_triage", "source_url": "https://arxiv.org/abs/2610.19827v1", "source_sha256": null, "evidence_limitations": "决定仅基于官方元数据/摘要范围或明确的来源例外；未核验正文算法或资源，不以metadata hash冒充文档hash。" }, { "id": "2610.19789", "title": "More Methods, Better Agents? A Reality Check on Self-Distillation in Agentic Reinforcement Learning", "published": "2026-10-14T08:36:31Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "522e17385bd2acdfa6654fc1995d8fd681b4dd9d946d643493b4179cab43a98a", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "该公平比较研究对十种self-distillation方法与GRPO作同协议比较；保留“收益不普适/稳定性和成本分离”的结论，不整套重跑。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "official PDF page 1 study scale", "PDF experimental setup and compute accounting" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/pdf/2610.19789v1", "source_sha256": "2038e0601b8cab431720f40b3542991236e8a47f25d5a213d6c3716c1e445b80", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "source_attempts": [ { "url": "https://arxiv.org/html/2610.19789v1", "result": "unavailable", "error": "HTTP Error 404: Not Found" }, { "url": "https://arxiv.org/pdf/2610.19789v1", "result": "retrieved", "sha256": "2038e0601b8cab431720f40b3542991236e8a47f25d5a213d6c3716c1e445b80" } ], "cost_assessment": { "source_url": "https://arxiv.org/pdf/2610.19789v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。132训练runs、12,124.4 H800 GPU-hours、四环境/三规模，并重复八个rollout seeds评测", "scale": "132训练runs、12,124.4 H800 GPU-hours、四环境/三规模，并重复八个rollout seeds评测", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19687", "title": "Function-Space Temporal Self-Distillation for Policy Optimization", "published": "2026-10-14T07:10:49Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "681d8c99d1b9a642988ef025cbc01885f7a5ae1d73ca97f224ed366635e2e09f", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "FSPO 保留稀疏历史next-token分布进行函数空间时序巩固，并结合特权纠偏与可靠性调节。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§2 Function-Space Policy Optimization", "§3.1 Experimental Setting" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19687v1", "source_sha256": "a50eeb75ac1977955e46f9298645a272a21818af16551ed4be6ea16ff5d430c0", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19687v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。每run4 GH200，1小时/5小时训练预算；7B峰值约56GB不代表总训练便宜", "scale": "每run4 GH200，1小时/5小时训练预算；7B峰值约56GB不代表总训练便宜", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19642", "title": "Model What Matters: Decision-Guided Selective World Modeling for Language Agents", "published": "2026-10-14T06:11:08Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "e05c9c7b747a3a9365964da49b4beaf9919a35ad17e6324409076d4a3bec0814", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "DGS-WM 依据后续策略决策选择observation单元，并用预测置信度调整辅助世界模型损失。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 Methodology", "Implementation Details.", "Appendix A Task Setup." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19642v1", "source_sha256": "7a117b71ee28fae182603a57602e2598de9bbe286ee2b0f9471baec6c5dc6705", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19642v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。Qwen3-1.7B/8B、LoRA rank32；QA检索约500万Wiki18文章/E5+FAISS，另交互WebShop；完整索引/多任务训练总成本未核定", "scale": "Qwen3-1.7B/8B、LoRA rank32；QA检索约500万Wiki18文章/E5+FAISS，另交互WebShop；完整索引/多任务训练总成本未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.19614", "title": "Divergent Paths: Learning an Outcome-Predictive Representation of LLM Agent Trajectories", "published": "2026-10-14T05:36:22Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "4d47a05f4c63ab39385269f8752d89a9b66c82a48d652118e873b92fc4c2dbc6", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Precog-RM 学习轨迹表征并以few-shot outcome预测提前终止GRPO；47.3%是相对FLOPs节约，不是低成本绝对预算。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 Precog-RM", "§4 Experimental Setup", "Where the compute goes." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19614v1", "source_sha256": "7c01e4bfb6864bd8e3e6503274dadddafb0b8a3874c2da5516c2473c068c5919", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19614v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。三agent基准×三策略模型，跨任务meta-learning轨迹表征，再在Tau2-Bench训练GRPO；生成占基线FLOPs89%，全链路总时未核定", "scale": "三agent基准×三策略模型，跨任务meta-learning轨迹表征，再在Tau2-Bench训练GRPO；生成占基线FLOPs89%，全链路总时未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.19568", "title": "INTACT: Inducing Latent Reasoning on a Frozen Backbone", "published": "2026-10-14T04:35:34Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "254345b969473940bb8d1cd93bfd9d7f4f27e25f1359ed025dadf082ed4c6812", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "INTACT 在冻结backbone上训练外部latent encoder，先蒸馏written reasoning再以答案正确性RL强化；小可训练参数不等于短训练。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 INTACT training", "Trainable parameters and training cost." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19568v1", "source_sha256": "33264e53530ff5c87419c3459d939420801cd3c90c05d206ee6fc9e35345bd45", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19568v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。1B–8B四backbones；Llama-3.2-3B在单H100 NVL上Stage1 6.2h、Stage2 10.3h，合计约16.5h", "scale": "1B–8B四backbones；Llama-3.2-3B在单H100 NVL上Stage1 6.2h、Stage2 10.3h，合计约16.5h", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19552", "title": "Test-Time Forecasting: Recurrent Depth Enables In-Context Dynamics Prediction", "published": "2026-10-14T04:08:44Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "11e3adeee4d21e8409799dfa08b822ba563e5697b895203778b918fd56d38421", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "dynamical_forecasting_out_of_scope", "reason": "TTF 的主问题为ODE/PDE/时间序列propagator与循环深度预测，不是LLM后训练；本轮只完成摘要范围分流，未审正文或判定其成本；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "official arXiv abstract: scope triage" ], "fulltext_review_completed": false, "review_level": "official_abstract_scope_triage", "source_url": "https://arxiv.org/abs/2610.19552v1", "source_sha256": null, "evidence_limitations": "决定仅基于官方元数据/摘要范围或明确的来源例外；未核验正文算法或资源，不以metadata hash冒充文档hash。" }, { "id": "2610.19528", "title": "Retrieve What Helps, Only When It Helps: Agentic Reinforcement Learning for Selective Retrieval-Augmented Machine Translation", "published": "2026-10-14T03:45:40Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "26ec710be3a8a23d3183e5a5560b03f6727561bc1d7d5a9220dd560a5ce95b85", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "SRAT 通过agentic RL学习何时检索/检索何物，并联合Plackett–Luce与质量对比训练retriever。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 SRAT", "Appendix A.2 Technical Details" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19528v1", "source_sha256": "fe1d4e137140e46c84a47eca4ae7f05203c1741d0b83ed9dd066293b168d02b0", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19528v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。7B GRPO、每prompt8 rollouts/最多5turn；8 A100节点，作者报告各任务约47–382 GPU-hours", "scale": "7B GRPO、每prompt8 rollouts/最多5turn；8 A100节点，作者报告各任务约47–382 GPU-hours", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19525", "title": "Diagnosing and Correcting Reward Hacking in Verifiable Instruction Following", "published": "2026-10-14T03:42:59Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "34f401c21faacfee0585a58c923dbdbe8edd0600f35be4649f89cff4a8943e2d", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "CaLIF 用请求本身的冻结reference likelihood，在两条离线anchor界定的区域内补奖励，抑制空泛约束满足和likelihood inflation。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4 CaLIF reward", "Appendix B Models and Training." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19525v1", "source_sha256": "9e491dc48a64a00e42231117895dfe5c39fd4fcad0085c2006cb3062e59a11ab", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "source_attempts": [ { "url": "https://arxiv.org/html/2610.19525v1", "result": "retrieved", "sha256": "9e491dc48a64a00e42231117895dfe5c39fd4fcad0085c2006cb3062e59a11ab" } ], "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19525v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。95,373 prompts；1.7B/4B/8B，8 H100，3–4epochs，每prompt8 rollouts，思考响应最长8192", "scale": "95,373 prompts；1.7B/4B/8B，8 H100，3–4epochs，每prompt8 rollouts，思考响应最长8192", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19457", "title": "MIPS: Mitigating Mode Collapse in Reinforcement Learning for Large Outcome Spaces", "published": "2026-10-14T02:42:13Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "8f7c5b03d4b9d4da2d27b5fc653ff40daaf14516e3ded59c218b9d0bcf6a6a9d", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "MIPS 以学习的反向轨迹模型做marginal IPS，不能把真实结果概率直接送给被评测策略来替代定义机制。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 MIPS", "§4.1 Hypergrid scalability", "Appendix C Experimental details" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19457v1", "source_sha256": "1396841e23e69b67b38e58194ab59be5aa9fab9cbf8666cd418a10b1d9c913ec", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "source_attempts": [ { "url": "https://arxiv.org/html/2610.19457v1", "result": "retrieved", "sha256": "1396841e23e69b67b38e58194ab59be5aa9fab9cbf8666cd418a10b1d9c913ec" } ], "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19457v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。二维Hypergrid H=4…512、G=16…2048，共192配置训练至收敛，另分子合成/系统发育；全套总时未能廉价核定", "scale": "二维Hypergrid H=4…512、G=16…2048，共192配置训练至收敛，另分子合成/系统发育；全套总时未能廉价核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.19404", "title": "ActMOF: Mapping Reticular Synthesis through Active Learning", "published": "2026-10-14T01:52:55Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "d080d784400e1b8e4850fb935f33981dbd753e18fe08b8ad4393249023859e75", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "materials_active_learning_out_of_scope", "reason": "ActMOF 是铝羧酸MOF结晶实验/主动学习benchmark，GP等策略与LLM的合成搜索比较不构成新的LLM后训练方法；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "official arXiv abstract: complete abstract scope review" ], "fulltext_review_completed": false, "review_level": "official_abstract_scope_triage", "source_url": "https://arxiv.org/abs/2610.19404v1", "source_sha256": null, "evidence_limitations": "决定仅基于官方元数据/摘要范围或明确的来源例外；未核验正文算法或资源，不以metadata hash冒充文档hash。" }, { "id": "2610.19362", "title": "Tail Competition Explains Scaling in Best-of-N Sampling for Verifiable Problems", "published": "2026-10-14T01:09:16Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "1c53d48b0900792da2c78860dcf3ce49e5eb6f122c1899d4015c240da172c30b", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "evaluation_analysis_reference_only", "reason": "Tail Competition 是Best-of-N正确/错误reward尾部竞争的理论与诊断研究，提出tail-tension停止指标而非新后训练目标；可以保留推理评测边界，不能宣称策略能力实现；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "§3 Tail-Competition", "§4 Scaling Law", "§5 Tail-Tension Early Stopping", "§6 Experiments" ], "fulltext_review_completed": true, "review_level": "primary_scope_or_evaluation_passages_review", "source_url": "https://arxiv.org/html/2610.19362v1", "source_sha256": "e0a6b50912163ead7f47cb6ffab5439f90e6e606682c71511b0317e82bdbe768", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。" }, { "id": "2610.19349", "title": "Where to Start, How to Move: Analyzing Curricula under GRPO for Long-Horizon LLM Agents", "published": "2026-10-14T00:42:53Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "3a1fd3f4882f210087195626eae6b8722a7d299f711863e4d420c9f9ee2d1113", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Guide-assisted GRPO 比较训练起点、handoff粒度与curriculum移动规则；成功reference前缀只限训练，不能在无辅助评测泄漏。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§2 Study Design", "Appendix B.2 Optimization and Rollout Generation" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19349v1", "source_sha256": "605e10ae5bdc1c171484a9bd74c56eae8b70e6784169c64229fd7b8b207a8754", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19349v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。三模型规模、BFCL/AppWorld长时程训练；原文每实验4 B200或4 H100，多起点/规则对照", "scale": "三模型规模、BFCL/AppWorld长时程训练；原文每实验4 B200或4 H100，多起点/规则对照", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19302", "title": "DTDD: Divergence-Triggered Dynamic Distillation for Reliable On-Policy Supervision", "published": "2026-10-13T23:26:26Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "5b963e2c68366941b38e00825efe9ac10414938f1a2e72a5b67da76e83d64c5a", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "DTDD 以片段分歧触发教师接管：学生前缀仍OPD、教师恢复后缀用behavior cloning，不能将教师tokens误标为学生on-policy。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 DTDD", "§4.3 Q2 base-student cost", "Appendix D.3 timing" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19302v1", "source_sha256": "4e3ac1546d09b4a1cb29dd8d9a4d952d288b509409733f9c851446bfc3df26c8", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19302v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。23实验臂；base-student设置约22 serial hours、43M teacher-written tokens，另长推理与多轮agent对照", "scale": "23实验臂；base-student设置约22 serial hours、43M teacher-written tokens，另长推理与多轮agent对照", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19298", "title": "Verifier-Grounded Agentic RL for Provably Correct Code", "published": "2026-10-13T23:17:25Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "2af8f8ce20f34114717924669d737e84c4c804ab238136fa0f9e1d906ab7292f", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "形式验证RL使用proof-obligation进度、隐藏probe和优先完整性审计，并有冻结proof specialist；简单单测奖励不能替代正式验证链路。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 Verifier-Grounded Reward", "§4 Experiments", "Compute." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19298v1", "source_sha256": "5044ebbcb6bf8d9c1183b23cdee201a688b613fe218eb8a69174ce668f85674a", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19298v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。1.7B–14B及35B MoE，项目约30,000 A100-hours，已报告runs约占一半", "scale": "1.7B–14B及35B MoE，项目约30,000 A100-hours，已报告runs约占一半", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19271", "title": "Outcome-Routed Distillation and Reinforcement Learning for Diffusion Language Models", "published": "2026-10-13T22:44:18Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b5332264243624e0f71e43211fed0c1122bdb8b28ac22d8b3d637a5a463657a7", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Outcome-routed dLLM训练对mixed groups用ratio-free REINFORCE，全失败但教师成功时走canvas KL，并以锚定KL保留能力。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 Setting and method", "§4 Experimental discipline" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19271v1", "source_sha256": "7a4825d6d582265aa975573114421d063747dbf78ea71fdd5d1e037eaf644254", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19271v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。单GH200级GPU每臂240 updates需约1–1.5天，每scoring sweep约14小时；多预注册臂和第二seed", "scale": "单GH200级GPU每臂240 updates需约1–1.5天，每scoring sweep约14小时；多预注册臂和第二seed", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19215", "title": "Narrowing the Collaboration Gap, Probably", "published": "2026-10-13T21:27:48Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "37435abee90ce39955d4db4a88afb92d82de58cbb373ea2f42b471aa98537cec", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "cross_track_communication_reference", "reason": "该文主线为概率通信优于动作通信的条件和校准，属于多Agent决策/推理界面研究；本次后训练审查只保留其理论与校准边界，不将它误当作新LLM后训练算法；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "§3 probability communication", "§4 calibration", "Appendix H Compute Resources" ], "fulltext_review_completed": true, "review_level": "primary_scope_or_evaluation_passages_review", "source_url": "https://arxiv.org/html/2610.19215v1", "source_sha256": "20d50663f6ffa9c88679d22edc175add972fe95197502b4ed0b061d79b00f208", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。" }, { "id": "2610.19145", "title": "Demystifying On-Policy Distillation: Where Does the Teacher's Advantage Live?", "published": "2026-10-13T20:11:48Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "cbd93aa9eecf4eebf80b7190d61671fdb0be5413c2c409cf0415d768df79118b", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "OPD教师优势研究通过受控token flips和分叉续写分析纠错价值；保持诊断设计，不把选中过金标的失败分支当可泛化能力。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4 Setup", "§5 Empirical Analysis", "Appendix A.1 Implementation", "Appendix A.3 Single Flip" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19145v1", "source_sha256": "0893daa56c1200cce16fde56b4bb12440b6dc61d0ba7d18196e86ff74db46d7d", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19145v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。5,926 forks，各4次教师最多8192tokens，再4次学生最多3072tokens，并有16续写resample、多模型分支；完整总时未核定", "scale": "5,926 forks，各4次教师最多8192tokens，再4次学生最多3072tokens，并有16续写resample、多模型分支；完整总时未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.19103", "title": "Contextual Credit Policy Optimization", "published": "2026-10-13T19:39:52Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "04510cb06f95438e1d1281d013ebaf3a519f79a7a9b66ede7adf596fbdf57ff1", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "CCPO 以冻结LLM表征给observation-matched peers加上下文权重，再结合历史与未来credit。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 CCPO", "Implementation details.", "Appendix B training configuration" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19103v1", "source_sha256": "bb452925a35f3fc74922d1aeef29b980f46cc8d04d4ae60c6962f92fcf74c806", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19103v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。1.5B/7B，150次迭代×16tasks×8轨迹，ALFWorld/WebShop；每run8 H200，三seeds", "scale": "1.5B/7B，150次迭代×16tasks×8轨迹，ALFWorld/WebShop；每run8 H200，三seeds", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19080", "title": "Backpropagating Verifier Rewards Through Continuous Flow Language Models", "published": "2026-10-13T19:12:33Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b50d255b458a62c106ba80396d860d773809c3e01439d6d436fa2861fb22eaf5", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "连续flow LM以中间embedding critic绕过离散decode、反传pathwise verifier reward，须保留critic随policy更新的链路。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4 CFRL critic/pathwise methods", "Compute." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19080v1", "source_sha256": "b80d7716d664680ffbf3b6f3c4a189e9eff0df77143c730f746a3f5efa80e0bc", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19080v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。原研究约300 GPU-hours（A100/H200/L40），单pathwise run约1 H200-hour，还含多任务/多轮比较", "scale": "原研究约300 GPU-hours（A100/H200/L40），单pathwise run约1 H200-hour，还含多任务/多轮比较", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19035", "title": "Who Should We Listen to More? Welfare-Aware Preference Acquisition for Pluralistic Alignment", "published": "2026-10-13T18:27:17Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "571610cd6e98fa6c6aaf5b6217914e2b10d93710a581c307af8787553d14b485", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "WAPA 用辅助reference估计个体supervision utility和marginal welfare gain，预算标签效率与完整推理/训练成本必须分开；oracle标签仅用于比较上界。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 welfare-aware acquisition", "Appendix F.2 Reference-Service Workload", "Appendix F.5 Downstream Alignment Training" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19035v1", "source_sha256": "57e138e2b07373b21c22919966e95ccc22297ba1ff40a17fa0727570d98d2b0e", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "source_attempts": [ { "url": "https://arxiv.org/html/2610.19035v1", "result": "retrieved", "sha256": "57e138e2b07373b21c22919966e95ccc22297ba1ff40a17fa0727570d98d2b0e" } ], "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19035v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。Qwen3-4B LoRA DPO，36个最终policy runs，单A100每policy10.00–11.56h、合计387.12 GPU-hours；reference/generation另计", "scale": "Qwen3-4B LoRA DPO，36个最终policy runs，单A100每policy10.00–11.56h、合计387.12 GPU-hours；reference/generation另计", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19011", "title": "PMO-Dock: Benchmarking Docking, Specificity, and Generalization in Molecular Optimization", "published": "2026-10-13T18:07:37Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "9a756bb8d835de6fcfb29aa64c63db0b66f9c64461657dff1a57e6bd7c58499a", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "molecular_benchmark_out_of_scope", "reason": "PMO-Dock 是25个docking任务的分子优化benchmark与严格validation/test oracle预算协议，不是新的LLM后训练方法；可参考隔离协议，不启动分子oracle扫参；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "§2 benchmark and splits", "§3 experiments", "Appendix I.3 Compute Resources" ], "fulltext_review_completed": true, "review_level": "primary_scope_or_evaluation_passages_review", "source_url": "https://arxiv.org/html/2610.19011v1", "source_sha256": "a532eb33ac76b42473325fa7f53ff8644ad790054844e093c17c167d7d286f73", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。" }, { "id": "2610.17623", "title": "RELAY: Selective Self-Distillation with Adaptive Hand-off for Rubric-Based Reinforcement Learning", "published": "2026-10-13T02:45:54Z", "revision": 2, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "7683527bf59a696de6a00c82d2f037d1a72e6f156ee2c708e1b3f3e1e86535f5", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "RELAY v2按positive/low-zero-advantage选择蒸馏rollouts，并以loss plateau停用教师后继续GRPO；已按v2而非v1审查。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "v2 §3 Method", "v2 Appendix A Implementation Details" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.17623v2", "source_sha256": "810b945e680f9d8bdf48ca10ac34df6cb02b1e97e5c3dabd1fe0a06972e2ff5d", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.17623v2", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。1.7B/4B多领域，8 H200、Megatron-Core/vLLM，teacher hand-off在约122–171步，四设置对照", "scale": "1.7B/4B多领域，8 H200、Megatron-Core/vLLM，teacher hand-off在约122–171步，四设置对照", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.15778", "title": "Who Sets the Pace of Judge Hacking? Judge Capability versus the Policy-Judge Gap", "published": "2026-10-11T17:05:39Z", "revision": 2, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "ea41c1fdcdc344d264eb1749936ceabf925ee0ef3101ca7ff0966f0257127b7f", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《Who Sets the Pace of Judge Hacking? Judge Capability versus the Policy-Judge Gap》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15778v2", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14959", "title": "AlphaRLM: Aligning Language Models for Literature-Grounded Symbolic Factor Discovery", "published": "2026-10-10T14:32:10Z", "revision": 2, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "0830514fa767f795bcf711cae171d3d5241ad9d81f97295da2a621380a98c109", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《AlphaRLM: Aligning Language Models for Literature-Grounded Symbolic Factor Discovery》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14959v2", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.18883", "title": "Token Utility for On-Policy Distillation via Local Trust-Region Optimization", "published": "2026-10-01T20:01:20Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b830af9574069d7fdb60ba06114f2d78d639a3a1c584905ef7d70ef676688959", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P1", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-01，早于2026-10-13窗口起点；《Token Utility for On-Policy Distillation via Local Trust-Region Optimization》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.18883v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.18881", "title": "Following the Flow the Law Requires: Training Voice Agents to Transfer to Unseen Regulated Procedures in Debt Collection", "published": "2026-10-01T03:08:52Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "8cce0f3c9a81f78098b7fac2ce74ac35ed4c00074deb689510f205a2f427dc59", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-01，早于2026-10-13窗口起点；《Following the Flow the Law Requires: Training Voice Agents to Transfer to Unseen Regulated Procedures in Debt Collection》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.18881v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.18878", "title": "Inverse Reasoning Models for Efficient Post-Training and Verification from Answer-Only Data", "published": "2026-10-01T01:56:34Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "5c174c8964dd4f6832a96659b140e7d6ce8673bbeeb1d5203a8617ec8587517b", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-01，早于2026-10-13窗口起点；《Inverse Reasoning Models for Efficient Post-Training and Verification from Answer-Only Data》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.18878v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.20570", "title": "Manacá-1B-Instruct-2: Reproducible Safety Alignment for a Native Brazilian-Portuguese LLM. Negative DPO Results, a Controlled Composition Scan, and a Silent Failure Mode of the LLM Judge", "published": "2026-10-14T17:53:46Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "1bd7767fff30ed24b54c8262636548318eb8baaadd9c1bf0eea76bcb4fb3a6a3", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Manacá公开三次负DPO、三轮Safety-SFT与数据配比扫描；负结果和judge拒绝评分的有效N问题应保留，不能以reward margin冒充生成能力。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 Methodology", "§4.2 Three DPO Attempts", "§4.3 Safety-SFT", "§4.7 Compute Budget" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.20570v1", "source_sha256": "d3f00ac7fbfd18690b7110ade9cf886d18d79ffb66b900529d503bfee9fb5dfb", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.20570v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。1.72B模型、291,223条SFT样本，三DPO尝试/三Safety-SFT迭代/10%–60%混合扫描；全活动总时未核定", "scale": "1.72B模型、291,223条SFT样本，三DPO尝试/三Safety-SFT迭代/10%–60%混合扫描；全活动总时未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.20195", "title": "What Makes a Social Agent's Interactions Effective? Learning from Verifiable Outcomes in Paired Worlds", "published": "2026-10-14T13:44:30Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "a0f8f77a0c1b372f59c58f81a4e367f89a9e6781ba160fceec09709b40e63550", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "PaCPO 在paired worlds中以逐turn暂定计划的可行性差异分配信用；隐藏私有约束由checker持有，被评测Agent不能直接读取两世界金标计划。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 HICO-Bench protocol", "§4 Paired-Credit Policy Optimization", "Appendix F A Paired Scenario" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.20195v1", "source_sha256": "daad4169463e2eb83bf2d7e38d82befc9e8791ae65b242af3bac6485f81fb7fb", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.20195v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。180 HICO情境、13模型评测，Qwen3-8B多轮paired-world训练及LLM integrity judge；完整rollout/API总成本未核定", "scale": "180 HICO情境、13模型评测，Qwen3-8B多轮paired-world训练及LLM integrity judge；完整rollout/API总成本未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.20155", "title": "LLM-Based Policy Compliance Verification of Loan Applications", "published": "2026-10-14T13:12:51Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "490c568be48366713f3cae08de0d3277f236fb00a8667f1c7a49ae2c4f9d24ad", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "application_prototype_not_post_training", "reason": "银行贷款政策checker是规则抽取加申请核验的原型与11份真实申请评估，未提出新的LLM后训练方法；真实业务文档不自动构成后训练或生产A/B能力证据；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "official arXiv abstract: complete abstract scope review" ], "fulltext_review_completed": false, "review_level": "official_abstract_scope_triage", "source_url": "https://arxiv.org/abs/2610.20155v1", "source_sha256": null, "evidence_limitations": "决定仅基于官方元数据/摘要范围或明确的来源例外；未核验正文算法或资源，不以metadata hash冒充文档hash。" }, { "id": "2610.19928", "title": "MemDuet: Learning Dual-Process Memory for Long-Context Reasoning", "published": "2026-10-14T10:16:13Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "c30e642cd2baa7844705f87cde58168832e0814477044619b5155b70c47234e4", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "MemDuet 以小模型parallel evidence selection与大模型memory verification/update组成三阶段SFT/GRPO/joint策略训练。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 MemDuet training stages", "Appendix B Training and Implementation Details" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19928v1", "source_sha256": "a77fddb1b1580d390bf79aca93f7fd8bc29ac43f5f9bec45d8cf731da54abc40", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19928v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。原文训练8 A100 80GB；推理单A100只是一个环节，不能据9.74x相对推理加速判断完整复现便宜", "scale": "原文训练8 A100 80GB；推理单A100只是一个环节，不能据9.74x相对推理加速判断完整复现便宜", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19915", "title": "Not All Criteria Matter: Importance-Aware Criteria Weighting for Rubric-Based LLM Evaluation", "published": "2026-10-14T10:05:43Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "d8d1ebe2828c833f88da8693e0be9f99016518e0f04414a7b1b466327be068ff", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "ICW 通过LLM两两重要性比较、一致性修订和principal eigenvector导出rubric权重，不能以随机构造比较矩阵替代LLM比较再宣称后训练增益。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 ICW workflow", "§4.1 Experimental Setup", "Downstream RL Settings." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19915v1", "source_sha256": "d2cb3d715264afca80ee81433d49b5a46088da67a2e9224a20693844455e6241", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19915v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。六偏好benchmark、多种LLM权重比较/修订，另下游RL最长9216响应；全链路调用和训练总时未核定", "scale": "六偏好benchmark、多种LLM权重比较/修订，另下游RL最长9216响应；全链路调用和训练总时未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.19767", "title": "Likelihood Is Not Authorship: Can We Reliably Measure Human Contribution in AI-Assisted Writing?", "published": "2026-10-14T08:18:02Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b1c781761b8c3e853a3f68ce4b24186dbdd4fe26d2419bbc834ca2b666ed4877", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "authorship_measurement_out_of_scope", "reason": "该文研究AI辅助写作贡献度量的不可识别性和specificity-padding审计，而非后训练方法；不能把likelihood分数等同人类作者贡献；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "§2 Contribution as Predictability", "§3 context-swap audit", "Appendix A.3 Models and Computations" ], "fulltext_review_completed": true, "review_level": "primary_scope_or_evaluation_passages_review", "source_url": "https://arxiv.org/html/2610.19767v1", "source_sha256": "d3597573a663091a6131f50d96e3615511be11b609574e518705a569988ba051", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。" }, { "id": "2610.19758", "title": "Qwen-MT-Lumen at WMT2026: Instruction-Specialized Translation via Quality-Aware Training and On-Policy Distillation", "published": "2026-10-14T08:12:10Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "9160951ab59a2bdbf8ca56655f2d41785fed4ec58aab7f4fa83ec44fedecc5b0", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Qwen-MT-Lumen包含CPT/SFT、online M2PO、教师OPD和结构修复的四阶段翻译系统，不是单一轻量损失函数。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§2 System Description", "§2.2 CPT and SFT", "§2.3 online M2PO", "§2.4 OPD", "Translation-specific instruction data." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19758v1", "source_sha256": "7c154cc887b4960ce3b2dffecb30748e3c75165c2a2fda48bc6e6cfc24dd31d0", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19758v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。14B学生，23翻译方向，Qwen3.7-Max合成/筛选教师数据与更大教师OPD；完整数据构造/训练总成本未核定", "scale": "14B学生，23翻译方向，Qwen3.7-Max合成/筛选教师数据与更大教师OPD；完整数据构造/训练总成本未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.19705", "title": "RouteWeaver: Weaving Mode Selection and Execution into Unified LLM Routing", "published": "2026-10-14T07:33:05Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "e7a94b5094e932b867ecdccc3349afc616f5229003e6887a3207ddbd91241898", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "RouteWeaver 用COMET-GRPO拆分mode/execution信用，并以progressive exploration学习三类routing mode。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4 RouteWeaver", "§5.1 Experimental Setup", "Workers." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19705v1", "source_sha256": "325fdf28a0c04f0f6046efdb5a51159087fc62ca5d760f87c80ed3b610c9567e", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19705v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。六冻结workers：两个本地workers已需2 A100，另四个经OpenRouter API；还需router RL和五域评测", "scale": "六冻结workers：两个本地workers已需2 A100，另四个经OpenRouter API；还需router RL和五域评测", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19601", "title": "PA1-OPD: One-Step Prefix-Augmented On-Policy Distillation for Language Model Pretraining", "published": "2026-10-14T05:23:00Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "7a0b658a51d7a4e5fcda1641c0442ded9330c797578f123dff748cd4ffbfed5b", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "PA1-OPD 在语料prefix后增加一步学生转移的教师监督，使用共享prefix tree-attention而非离线KD替代。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 PA1-OPD", "§4 Experiments", "Implementation details." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19601v1", "source_sha256": "08652fb0dc8e0dd2d8b4067b2a1d06028bfc7996a5494e5fe444298d96f76e18", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19601v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。8 accelerator cards，2048序列、global batch约2.1M tokens，每run4769更新至10B tokens，两个continued-pretraining设置", "scale": "8 accelerator cards，2048序列、global batch约2.1M tokens，每run4769更新至10B tokens，两个continued-pretraining设置", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19560", "title": "$τ$-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions", "published": "2026-10-14T04:19:15Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b6375d3a78a1b7244204ad8332f7bf346109e379ef805c1c26b049d53acc1dbb", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "τ-Emotion-Bench/Tracer以可中止的情绪用户交互筛选监督轨迹，并训练真实tool-calling policy，不是只改情绪标签。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 benchmark and Tracer", "Appendix B.1 Fine-Tuning Hyperparameters" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19560v1", "source_sha256": "f1f03ea97a81c44dd59fde12da1bbf0671eb36d3a19595a4bbd4915456790585", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19560v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。十domain、Qwen3多规模；SFT用4 H100，最大sequence32768，另轨迹生成/双user simulators评测", "scale": "十domain、Qwen3多规模；SFT用4 H100，最大sequence32768，另轨迹生成/双user simulators评测", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19475", "title": "ActICL: Post-Training for Agentic Multimodal In-Context Learning", "published": "2026-10-14T02:54:40Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "fe8bf1f9e4325e8821e82d77ef419e3507eb8308b6fc2130956a717409bd3c7f", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "ActICL 联合学习多模态demonstration acquisition和answering，必须保留多turn检索/筛选与两阶段训练。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 ActICL", "Appendix E Implementation Details" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19475v1", "source_sha256": "0fea774c6f6f35eac5dd27534930725b8b5d1ed322bf030e08869cf6e80f1840", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19475v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。8B/9B/30B-A3B，全语言backbone参数更新，两个阶段均8 H200，七benchmark", "scale": "8B/9B/30B-A3B，全语言backbone参数更新，两个阶段均8 H200，七benchmark", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19408", "title": "ForkGRPO: Efficient Exploration and Fine-Grained Credit Assignment for Diffusion Language Models", "published": "2026-10-14T01:56:46Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "92536affb2c281c50e1c37e740aacdfbf9f8fc5b6a1c77fe3c4068eb30314554", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "ForkGRPO 在高熵中间状态fork，多组续写提供state内与state间segment credit，Fast分支仍需真实rollout训练。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 ForkGRPO", "Appendix C.4 Optimization" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19408v1", "source_sha256": "cf2e26e4b4ce4db770acfd011f91c19009506e648a2db37d228a3259d41c29c8", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19408v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。8 H200，64 prompts/update，每prompt默认32 scored completions，100更新及额外消融", "scale": "8 H200，64 prompts/update，每prompt默认32 scored completions，100更新及额外消融", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19380", "title": "Learning What to Keep: Transfer-Aware Personal Skill Learning from User Feedback", "published": "2026-10-14T01:27:19Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "efab763971201741303c41139a99e2865d0009bc4f930797aab7a61540bfebe5", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "SkillGrain 以原子rule transferability、rubric reward和细分credit联合优化skill induction/conditioned generation。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4 Method", "Appendix B.6 Comparison of Computational Cost" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19380v1", "source_sha256": "97971d0c238c56645d908e0167e7a1088b1e3e0b7a51dfebf36ac69bb9e554e0", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19380v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。三base models、五writing tasks、in/out-of-domain；Qwen3-4B成本表在4 H100上测得，非单卡配置", "scale": "三base models、五writing tasks、in/out-of-domain；Qwen3-4B成本表在4 H100上测得，非单卡配置", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19367", "title": "Beyond Corrected Actions: Training LLM Agents on Textual Critiques", "published": "2026-10-14T01:12:03Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "0634497ba426b0da540a9616e566318e6198a151ac38dcafe62263d33c5f7439", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "CAFT 对学生错误action同时生成专家纠正和文字critique再真实微调，并比较RL warm start。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§2 CAFT supervision", "Appendix C.1 Trajectory collection and annotation" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19367v1", "source_sha256": "c11af6ff24fbbb3dba11e28c83a973c3c6a78a45ad9c296b51b9e13b863e1ecf", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19367v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。四学生模型，bf16全参数FSDP，8 RTX PRO 6000 96GB单节点；1–2epochs、多环境与迭代对照", "scale": "四学生模型，bf16全参数FSDP，8 RTX PRO 6000 96GB单节点；1–2epochs、多环境与迭代对照", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.19210", "title": "Titan-Llama: Retrofitting Pre-trained LLMs with External Neural Memory Modules", "published": "2026-10-13T21:23:58Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "10dabe96e54f5ea720d27f64f3bfdc8c701ea998ec8aba4dfdc7fb5e79219c5e", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Titan-Llama 将部分attention换为可测试时更新的神经memory，需保留冻结/适配层和真实长上下文训练，非普通KV缓存演示。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 Titan-Llama architecture", "§4.1 Post-Training for Long Context Reasoning", "Appendix A Training Details", "Appendix B Inference Implementation" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19210v1", "source_sha256": "14ce7fea58976af67c45e428c088ca3b459ffb506989e76b5cf2016f878003db", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19210v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。Llama-3.2-1B-Instruct、长上下文训练/多chunk-size和CUDA Graph推理比较；小backbone不代表完整训练与扫评总成本已核定", "scale": "Llama-3.2-1B-Instruct、长上下文训练/多chunk-size和CUDA Graph推理比较；小backbone不代表完整训练与扫评总成本已核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.19108", "title": "Towards a Data-Centric Understanding of Spatial Reasoning in Multimodal Large Language Models", "published": "2026-10-13T19:42:30Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "08cf9421caf2db015677fecf80fcad8773c89cdebe066c42bbc97cdcf3f2ea23", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "evaluation_analysis_reference_only", "reason": "该文是空间推理数据集构建/标注/评价的survey，未提出待复现的新后训练算法；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "§1 Introduction", "§3–§4 survey taxonomy" ], "fulltext_review_completed": true, "review_level": "primary_scope_or_evaluation_passages_review", "source_url": "https://arxiv.org/html/2610.19108v1", "source_sha256": "625f6a7384498323cb731586412e03a55b28f3424916462c4263bc3262d9019f", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。" }, { "id": "2610.19100", "title": "GroundGUI: GUI Grounding by High-Quality Data and Distillation-to-Reinforcement Learning", "published": "2026-10-13T19:37:49Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "d0b5a2670eb0979fa521efc8fcec28345f35aa2f3d06cd599f42248f90fab0f4", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "GroundGUI 以ScreenRef真实图文grounding数据进行SFT，再OPSD→RLOO的D2RL，不能只测point-in-box reward取代模型训练。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 ScreenRef", "§4 D2RL", "§5.1 Implementation Details" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19100v1", "source_sha256": "990c291d9e0db96f3cc933087a4dcdc82fa2e285edf8d87808a88adc91d8cb25", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19100v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。586K screenshots/633K referring instances；Qwen3-VL-8B，SFT4 B300×2epochs，D2RL8 B300×2epochs", "scale": "586K screenshots/633K referring instances；Qwen3-VL-8B，SFT4 B300×2epochs，D2RL8 B300×2epochs", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18869", "title": "What Matters in Quality-Aware VLA Pretraining: A Benchmark and Empirical Study", "published": "2026-10-13T17:59:52Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "7ecd0626e56be8563fab9a029c53e60599798d41756c0198e742fc7b1e0098d0", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "embodied_pretraining_comparison_reference", "reason": "AXIS-Bench比较三VLA在不同示范质量weighting/filtering/conditioning下的预训练，是具身质量数据benchmark/实证研究；本后训练队列不把机器人pretraining结果登记为LLM后训练能力；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "§3 benchmark", "§4 State-Derived Quality and Pretraining", "§5 Simulation Experiments" ], "fulltext_review_completed": true, "review_level": "primary_scope_or_evaluation_passages_review", "source_url": "https://arxiv.org/html/2610.18869v1", "source_sha256": "ac042f05026e48454b696afe735345bed374a60558e698d6a72340398b311597", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。" }, { "id": "2610.18868", "title": "Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards", "published": "2026-10-13T17:59:51Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "e92f160d3eec6cb12d3cabb724c0e9bc40aa170d395e02d0cabb04832dc7121f", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Tev 以latent-world-first合成可验证环境，并用progress potential shaping训练长时程Agent；隐藏完整世界规范不能给被评测Agent读取。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§2 environment synthesis", "§3 PPRS", "§4.1 Setup" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18868v1", "source_sha256": "d44dff44fb9ec644dfc6313981993613ec28d2a2a3224eff014e4f556d224dce", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18868v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。9B/27B及320B扩展；27B训练4节点B200或8节点A100约4天，每task24 rollouts、最多65536tokens/64工具调用", "scale": "9B/27B及320B扩展；27B训练4节点B200或8节点A100约4天，每task24 rollouts、最多65536tokens/64工具调用", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18861", "title": "Hybrid Reasoning Agents", "published": "2026-10-13T17:59:31Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "9f98e83ffdda40f482e573e8b97e06724f65f5561125c9d67aeff24ee169557d", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "HybridAgent 以HybridLink在latent reasoning与文本tool interaction间切换，需保留HybridSFT/HybridRL和跨turn连续memory。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 hybrid architecture", "§4 HybridSFT/HybridRL", "Appendix H.3 Implementation Details" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18861v1", "source_sha256": "7b3cfbfe6aede417b69a6fd36b4bb00b6110818a38010133203c62452b59d664", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18861v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。三模型规模/11 benchmarks；所有训练和推理在32 H100，agent trajectory最多24000steps/tokens并含search API/judge", "scale": "三模型规模/11 benchmarks；所有训练和推理在32 H100，agent trajectory最多24000steps/tokens并含search API/judge", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18851", "title": "MFG: Multi-turn Feedback Guided Molecular Optimization", "published": "2026-10-13T17:58:41Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "312cb2b7efe1de586195d63ad04794473260f9e3c5f5835f8bbf96b658f6615f", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "MFG 以跨turn property/similarity feedback与反馈对齐reward训练分子优化LLM，不能只用现成property打分器模拟学习。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 MFG", "Training" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18851v1", "source_sha256": "421621c404a5eb12c35ddaf3822c9e29e713d2e52e9f4378c921194ac5bf7243", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18851v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。GeLLM³O，GRPO group4加动态重采样；原文训练2 A100约36小时，十个多目标任务", "scale": "GeLLM³O，GRPO group4加动态重采样；原文训练2 A100约36小时，十个多目标任务", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18807", "title": "SkillScaffold: Scaffolding Agentic Reinforcement Learning with Policy-Adaptive Skill Credit Assignment and Selective Injection", "published": "2026-10-13T17:46:25Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "6e1e75df72a973c537b4629f2f9465836471a8e771f8dfa90cf2dea38b9ed32c", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "SkillScaffold 用verified trajectories的outcome-aware likelihood估计skill utility与选择/弃用，联合Generator/Executor训练；不允许金标成功轨迹被评测Agent直接读取。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4 SkillScaffold", "Baselines and Implementation Details." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18807v1", "source_sha256": "1d9f30f7faf8fade87c3aa7a7c356e809496445cf4207dffaa9bd43849ace651", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18807v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。Qwen3-8B，ALFWorld/WebShop/ScienceWorld/τ²-Bench，generator-executor训练及多类baseline；完整总成本未能廉价核定", "scale": "Qwen3-8B，ALFWorld/WebShop/ScienceWorld/τ²-Bench，generator-executor训练及多类baseline；完整总成本未能廉价核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.18683", "title": "ProAS: A Progress-Aware Advantage Shaping Framework for Multi-Turn Agentic Reinforcement Learning", "published": "2026-10-13T16:39:38Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "d9cf0aacb3627a27a39b008e3061cf751a89adb3d358ec239956e7a34e2d925b", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "ProAS 动态Agentic Reference Graph支持无效正信用mask与不对称负信用重分配，保留ARG定义而非用金标计划直接决策。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4 ProAS", "Appendix B.1 Training Framework" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18683v1", "source_sha256": "9d4e0454bf353f38be8a486de515b217bfcb1990395193fcf0a793e3db82ab9b", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18683v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。ALFWorld/WebShop，1.5B每run2 GPUs、7B每run4 GPUs，多group-RL基线", "scale": "ALFWorld/WebShop，1.5B每run2 GPUs、7B每run4 GPUs，多group-RL基线", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18642", "title": "MultiAlloc: Budget-Aware Precision Allocation for KV Cache Quantization", "published": "2026-10-13T16:10:45Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "1488ed24ea0b77dae7626a2fa5b8590e36a1134298ca4c4678cab9fdf4c92070", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "MultiAlloc 通过离线layer sensitivity/token importance profiling，在固定KV预算内联合分配层与token保真精度。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4 MultiAlloc", "§5.1 Experimental Setup" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18642v1", "source_sha256": "aeaae95ae9d091e3b86fc21fe9a083d1ef213d3b16c2e690833b6d923421799d", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18642v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。7B/8B/14B/32B五模型、GSM8K/GPQA/LongBench-v2，A100执行；完整profiling、多budget、多模型长生成总时未核定", "scale": "7B/8B/14B/32B五模型、GSM8K/GPQA/LongBench-v2，A100执行；完整profiling、多budget、多模型长生成总时未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.18614", "title": "Can an Agent Learn to Predict Its Own Failure? Outcome-Grounded Confidence Calibration in LLM Agents", "published": "2026-10-13T16:00:01Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "95ac6f8fcf6e1addf64a85b90635bd4c4131202663f871b3721e98c73a339617", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Outcome-grounded confidence用每step置信度对最终episode outcome的proper scoring rule训练，必须分离任务难度、深度与attempt foresight。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 estimands and protocol", "§4 method", "Hardware and compute budget." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18614v1", "source_sha256": "b73110b5869e51640513a6c6a80130fc705b3d512194b444993af52485365d51", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18614v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。每job单H200 LoRA，但7B ALFWorld约8h、14B约12h、TextWorld约5h；总training约200 GPU-hours", "scale": "每job单H200 LoRA，但7B ALFWorld约8h、14B约12h、TextWorld约5h；总training约200 GPU-hours", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18572", "title": "LLM Agents for Automated Post-Training via Structured Recipe Optimization", "published": "2026-10-13T15:39:34Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "7b34ca962612544490cee7aa5204c00c6790ab41208ef0d9a563c16582254c69", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "SRO 让Recipe-Space Designer显式定义可组合选择，再用结构化trial evidence和multi-objective BO搜索多域post-training配方。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 Structured Recipe Optimization", "Training and compute budget." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18572v1", "source_sha256": "84c08928dd726678e4e30ab9b908d193003e645094108b065d193fb64838bb3b", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18572v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。单GPU每个single-domain run预算10h，multi-domain20h，含多trial搜索与预算外full-set评测", "scale": "单GPU每个single-domain run预算10h，multi-domain20h，含多trial搜索与预算外full-set评测", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18504", "title": "How Low Can You Go: Fine-Tuning LLMs with Bayesian Optimization (LlMBO)", "published": "2026-10-13T14:58:27Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "ed1e020fcbf0b5305da99ad2983bf8226c737cb7d7938d8064d56ca21cd23d96", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "LlMBO 在TinyLoRA低维参数上做Gaussian-process BO并直接优化greedy accuracy；优化参数很少仍需大量LLM生成。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4 LlMBO", "§5.4 compute efficiency", "Appendix I.3 Compute" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18504v1", "source_sha256": "6d2421f8c982ae426a39b1af20f5c91db5f8ac34fa7fd448c64a8c03c7b229f0", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18504v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。十模型、单GPU逐run调度；原文训练超过14,000 GPU-hours且未含全部中断恢复/评测", "scale": "十模型、单GPU逐run调度；原文训练超过14,000 GPU-hours且未含全部中断恢复/评测", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18457", "title": "Kairos: Transactional Security Enforcement at the System Boundary for LLM Agents", "published": "2026-10-13T14:24:26Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "609a14f10de4abe6dc74c02a7d2fce3b5a7e55f280b4bb9fe93dc00ad0a15f35", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "cross_track_runtime_security_reference", "reason": "Kairos 是隔离边界的事务式reference monitor与延迟effect释放系统，不更新opaque Agent参数；本条后训练审查不将runtime安全收益算成后训练算法收益；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "§III Design", "§IV Evaluation" ], "fulltext_review_completed": true, "review_level": "primary_scope_or_evaluation_passages_review", "source_url": "https://arxiv.org/html/2610.18457v1", "source_sha256": "db19a6ef5416284dfbb17d588ad1cf89811e9aa4e9de2cfb8cca2ee1c7b9a1c4", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。" }, { "id": "2610.18442", "title": "Copying Is Not Grounding: Training Faithful Long-Form RAG with Authored-Sentence Credit", "published": "2026-10-13T14:16:38Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "08e19418c25119d14b0d14f53fcb9787a2d05454f33a076e99bbd9c2b399fa36", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "ASC 将faithfulness reward/advantage移到authored sentence，排除copy与空泛句并基于同query所有sentences作比较。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§5 Authored-Sentence Credit", "Appendix B Setup Details" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18442v1", "source_sha256": "ead54815028c242997220faae304a2ed252563f05f0130ceb05c1cae9649e2a6", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18442v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。3B–9B四模型、922 training steps、32 answers/step；8 H100（6训练+2checker），9B单run约3.5h，评测另计", "scale": "3B–9B四模型、922 training steps、32 answers/step；8 H100（6训练+2checker），9B单run约3.5h，评测另计", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18415", "title": "Do Factuality Methods Make LLMs More Factual? Reproducibility and Fair Comparison", "published": "2026-10-13T13:54:51Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "2b146feab08348b92c1416a8a9a7017923be9d03b292bbe6b521fbe99994e568", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "evaluation_analysis_reference_only", "reason": "该文独立比较五种factuality alignment的复现和共用benchmark，报告不能稳定优于base及reward/judge脆弱性；这是评价与复现审计，不是新后训练能力实现；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "§4 Reproduction in Original Setups", "§5 Common Benchmark", "Ethical Considerations" ], "fulltext_review_completed": true, "review_level": "primary_scope_or_evaluation_passages_review", "source_url": "https://arxiv.org/html/2610.18415v1", "source_sha256": "7f37ea0423bece74ec7046f70612d41587df721f4a339ed7a903a322bd4906ba", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。" }, { "id": "2610.18343", "title": "Lightweight but Linguistically Fair? Quantisation Robustness across Bavarian, German, and English", "published": "2026-10-13T13:15:59Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "179343fee376878254ff65ed7d5133214e42b49dbbd788fa72d7e1694e5556ba", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "evaluation_analysis_reference_only", "reason": "该文对14个7B–10B模型的量化语言鲁棒性作对齐比较，没有提出新后训练算法；保留“选择题稳定不等于自由生成稳定”的评测边界；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "§3 Experimental Setup", "§4 Results" ], "fulltext_review_completed": true, "review_level": "primary_scope_or_evaluation_passages_review", "source_url": "https://arxiv.org/html/2610.18343v1", "source_sha256": "46b42788a6916a3f4fc60a50b1b4793e03c6332c732a1ea1232ac7ecb1379971", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。" }, { "id": "2610.18326", "title": "Jev-Guard: Fast Routing and Auditing Agent Actions over Editable Policy Trees", "published": "2026-10-13T13:03:37Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "677d7a6488545690ec7f7ad1038db4123f72a771c4e943d96a017ae9809fd457", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "frozen_checkpoint_system_study_reference_only", "reason": "Jev-Guard正文明确只评估一个冻结checkpoint，是editable policy-tree审计的经验系统研究；其结论指出routing正确后仍有分类/编辑失败、flat dispatch有竞争力，未建立安全泛化收益。保留这些边界，不把58ms局部计时或240msreview误称新训练方法或安全能力复现；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "Decision models and agent evaluation.", "Evaluation.", "7 Conclusion", "Latency." ], "fulltext_review_completed": true, "review_level": "primary_scope_or_evaluation_passages_review", "source_url": "https://arxiv.org/html/2610.18326v1", "source_sha256": "25d54f21ac93d5f578787814ff330ab323824c60d4a0fe8bbbd579d56c4f50d6", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。" }, { "id": "2610.18299", "title": "LegalRewardBench: Benchmarking Reward Models for Legal Preference Evaluation", "published": "2026-10-13T12:47:13Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "699de6cbc0941f02ac31a34e0154a52a4a2212c274a776a29a862295e8f56de7", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "evaluation_analysis_reference_only", "reason": "LegalRewardBench 是人工核验的法律reward-model preference benchmark和训练策略比较，非新后训练算法；仅根据已读摘要归类，不声称正文或资源核验；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "official arXiv abstract: scope triage" ], "fulltext_review_completed": false, "review_level": "official_abstract_scope_triage", "source_url": "https://arxiv.org/abs/2610.18299v1", "source_sha256": null, "evidence_limitations": "决定仅基于官方元数据/摘要范围或明确的来源例外；未核验正文算法或资源，不以metadata hash冒充文档hash。" }, { "id": "2610.18256", "title": "Teach Me After You Learn It: Lookahead Self-Distillation for RLVR", "published": "2026-10-13T12:22:54Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "326264972c41d31956562b6802611fccab2ad69963afd16839dca95f90227b62", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "LEAD 以disjoint on-policy子组做virtual RL lookahead teacher，再cross-group token distillation并调和GRPO信号。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 LEAD", "Appendix B.1 Detailed Implementation" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18256v1", "source_sha256": "68c89297028b72e80878e0f578bee7256262860d7432c5cc69d00c714052e995", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18256v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。每query8 rollouts，8 H800/FSDP bf16，多科学/agent基准及模型规模", "scale": "每query8 rollouts，8 H800/FSDP bf16，多科学/agent基准及模型规模", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18216", "title": "TextualRL: Textual Reinforcement Learning for Context Management", "published": "2026-10-13T11:47:32Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "32889e33d065c3f6f22267d935799e315f6de0210c0a63d92a76aba3db3028b1", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "TextualRL 以同任务rollout组做critique，并跨组证据审查bounded context patches，候选采用由held-out validation决定，不能换成金标脚本策略。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 TextualRL", "Implementation Details.", "Appendix B.5 Token Accounting" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18216v1", "source_sha256": "e500ea766b827c33faeb5b9f4e0dd04a57777d8df4b1cfcc7315d54016c32a3c", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18216v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。五target models（含GPT-5.5/5.4、27B/35B-A3B），每target×六bench独立优化4epochs，每batch40tasks×4rollouts；还需GPT-5.5 optimizer/API", "scale": "五target models（含GPT-5.5/5.4、27B/35B-A3B），每target×六bench独立优化4epochs，每batch40tasks×4rollouts；还需GPT-5.5 optimizer/API", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.18200", "title": "A Few More Tokens Ahead: Reinforcing Parallelism in Diffusion LLMs", "published": "2026-10-13T11:37:04Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "fc5b7f21209c41bbbc9b3a3766d05cd65e53c974efa2f6a18327deb43233a357", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "RePara 用连续短token blocks的group-relative policy updates强化dLLM并行生成，不能只改decode quota冒充RL训练。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 RePara", "Reference training cost." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18200v1", "source_sha256": "393667b66025440ee204853ec0d6cccf6228f3478319cbda0d99bfab5b22261c", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18200v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。128updates/MATH/256-token rollouts，作者报告LLaDA约497、SDAR约479 H200 GPU-hours", "scale": "128updates/MATH/256-token rollouts，作者报告LLaDA约497、SDAR约479 H200 GPU-hours", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18019", "title": "Machine--human agreement in marking undergraduate assignments is specific to model, prompt and institution", "published": "2026-10-13T09:14:58Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "67789a42649160df5a74ccb424203798d0e7de35f74ec2a3e790698da84b7450", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "education_evaluation_out_of_scope", "reason": "本科作业机器/人类评分一致性研究比较模型、prompt和大学分布，无新的LLM后训练方法，也不能用单学校校准表示跨机构泛化；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "§3 Methods", "§4.3 shared errors" ], "fulltext_review_completed": true, "review_level": "primary_scope_or_evaluation_passages_review", "source_url": "https://arxiv.org/html/2610.18019v1", "source_sha256": "db5805a639099a59e25d68ebe0a0d762c2129c5b6a9ff4a1410ac11fe38fc5e6", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。" }, { "id": "2610.17953", "title": "Agent-Reflex: Teaching Language Agents to Act Reflectively", "published": "2026-10-13T08:36:20Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "5d58f70b5bfff348387253ca869f3b53457362aaaf778ab79d0fa94ce3b87741", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Agent-Reflex包含噪声grounded skill discovery、ReSFT与reflection-guided on-policy RL三阶段，不以提示反思模板替代训练。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 Agent-Reflex", "§4.1 Experimental Setups" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.17953v1", "source_sha256": "2a386715c04ade524c872cf66ab1cd8db1a2f6c787991cbd7fc59b93587cd9ba", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.17953v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。Qwen3-8B，六组长时程/噪声tool-use suites、成功失败paired data挖掘及三阶段训练；完整生成和训练总时未核定", "scale": "Qwen3-8B，六组长时程/噪声tool-use suites、成功失败paired data挖掘及三阶段训练；完整生成和训练总时未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.17893", "title": "Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards", "published": "2026-10-13T07:46:48Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "fdff8f581ca7a4ffe4b99218ae0862e62a6ea570914ee251ff7f84c091579a69", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "CoSIR 以Bayesian team-surprisal下降奖励信息披露/请求/使用，并以self-play后训练迁移合作行为；评价器可用正确option，但Agent不得读取它。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 verifiable information reward", "Appendix F.7 Learning rates and hardware" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.17893v1", "source_sha256": "6fd8cf58e65ab8f80498c59d87497ffed63ed01c5cd833a8c33c56e7a38cf2bb", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.17893v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。Qwen3-4B/8B、Gemma3-4B，三collection rounds，使用2 H100和4 Gaudi2，多域transfer评测", "scale": "Qwen3-4B/8B、Gemma3-4B，三collection rounds，使用2 H100和4 Gaudi2，多域transfer评测", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.17824", "title": "Checking the Checkers: A Synthesis of Verifier Noise and Robustness in RLVR", "published": "2026-10-13T06:44:33Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "0ea01feea99b5d24e08ae2ce13cd294c00817a7720c98cd80368491d7bf20ecd", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "evaluation_analysis_reference_only", "reason": "Checking the Checkers 综合58项verifier噪声研究，是survey/报告协议而非可登记的新训练方法；随机误差与可利用false positives的边界应保留；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "§III Survey Methodology", "§IV verifier errors", "§V training effects" ], "fulltext_review_completed": true, "review_level": "primary_scope_or_evaluation_passages_review", "source_url": "https://arxiv.org/html/2610.17824v1", "source_sha256": "c8c4972b8ba467831db88bac6f192a0db5db763b546918ec2bb0cabb9f170f63", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。" }, { "id": "2610.17804", "title": "Learning from fracture: Teaching language models to design through physical experience", "published": "2026-10-13T06:30:52Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "d59b2202713f9747a6491b3bc0f37940e25a08aa0445e32c04f5a86b67478233", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "cross_track_scientific_agent_reference", "reason": "FracLearn 通过外部知识与物理模拟经验做设计，明确不更新模型权重；属于科学Agent/材料设计流程参考，不是本队列的LLM后训练复现；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "§2 experience-abstraction-generalization", "§3 Discussion", "§5 Materials and Methods" ], "fulltext_review_completed": true, "review_level": "primary_scope_or_evaluation_passages_review", "source_url": "https://arxiv.org/html/2610.17804v1", "source_sha256": "e894c14794bccfa0911c0d4ba2ec7abb60195815bcd28e0aabb238dbcccf8694", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。" }, { "id": "2610.17734", "title": "Explore with Memory, Learn without It: In-Context Exploration and Self-Distillation for LLM Reasoning", "published": "2026-10-13T05:13:10Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "aa86462e1075b27ef0388ada10fdca23994b8477e47ae99fd979b70f9c8d5218", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "LEAF 以动态problem-specific memory辅助探索，再将有记忆行为self-distill到无记忆policy；记忆和validation/test边界不能省略。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§2 LEAF", "Appendix C Exploration setup details" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.17734v1", "source_sha256": "e056b8ba896e43f9371e02ae6c362c0d8f7798ccc8efd8d1aa6287fc5393a90e", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.17734v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。Qwen3-8B/OLMo3-7B，四bench，每problem24 verifier calls，每bench单H200限5h，另蒸馏/消融", "scale": "Qwen3-8B/OLMo3-7B，四bench，每problem24 verifier calls，每bench单H200限5h，另蒸馏/消融", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.17733", "title": "CoCA: Criterion-Aware Credit Assignment for Rubric-Based Reinforcement Learning", "published": "2026-10-13T05:13:04Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "0d5c91d9f9a6d49696ba3a0ae7b1b75507523012918087d2defb3ab4a5da5f48", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "CoCA 由criterion group-relative outcome建立credit，再按criterion-induced distribution shifts分配token，非直接模仿privileged teacher。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4 CoCA", "Performance versus training time." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.17733v1", "source_sha256": "0c9d6688be70d806c8b75f140be2853a87c531e63379e1c39f8380a167c07baf", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.17733v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。Qwen3-8B Medicine 400steps用8 A100约38.77h；125steps仍12.12h，多域/规模与消融另计", "scale": "Qwen3-8B Medicine 400steps用8 A100约38.77h；125steps仍12.12h，多域/规模与消融另计", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.17726", "title": "TSRL: Trigger-based Supervised Reinforcement Learning for Continual Agent Training", "published": "2026-10-13T05:02:58Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "2347ec084af61f410d018381204982b71e5ed854f2020c43d0ea4892edd2d89f", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "TSRL 在RL中只对失败hard states触发SFT，研究持续学习遗忘；不能用成功轨迹oracle替代真实expert supervision。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4 TSRL", "Training Setup." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.17726v1", "source_sha256": "828368dceb56ed7d219b97f15048be568491e5ed7548421c5c100d68507d94eb", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.17726v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。7B ALFWorld/WebShop用8 H20；32B DeepSearch用128 H20，rollout group8、多任务continual比较", "scale": "7B ALFWorld/WebShop用8 H20；32B DeepSearch用128 H20，rollout group8、多任务continual比较", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.17531", "title": "Towards Green AI: Evaluating Energy Efficiency for Quantized LLM Inference", "published": "2026-10-13T00:52:32Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b6c9b45391f6b618885efe53e0e4d6c58ec0c3dcfe98331bb7565b6735df580a", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "evaluation_analysis_reference_only", "reason": "该量化能耗thesis统一比较十种PTQ、五模型和四GPU，主贡献是能耗/延迟/精度评测，不是新后训练算法；不把已有量化库包装成新能力；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "official arXiv abstract: complete abstract scope review" ], "fulltext_review_completed": false, "review_level": "official_abstract_scope_triage", "source_url": "https://arxiv.org/abs/2610.17531v1", "source_sha256": null, "evidence_limitations": "决定仅基于官方元数据/摘要范围或明确的来源例外；未核验正文算法或资源，不以metadata hash冒充文档hash。" }, { "id": "2610.17456", "title": "Compositional RL: Evaluating and Training LLMs for Collaborative Reasoning", "published": "2026-10-12T22:16:45Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "1e133df9539370dd4d0ebd2cbfd24aa75b6f028d36153784fde0d4423c5479a9", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Compositional RL: Evaluating and Training LLMs for Collaborative Reasoning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.17456v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.17426", "title": "Reinforcement Learning In Partially Verifiable Environments for Colon Cancer Treatment Planning", "published": "2026-10-12T21:39:21Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "5797bff4780dead79bedf824b614adde5bdb8da533333c4c621bc71eb4cc4638", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Reinforcement Learning In Partially Verifiable Environments for Colon Cancer Treatment Planning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.17426v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.17397", "title": "JustOPD: On-Policy Distillation from Autoregressive to Diffusion Language Models", "published": "2026-10-12T21:02:26Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "3f41d927d87e59b28d1240d7237501e67e640aefdab393ac3b9539cf4d07af49", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《JustOPD: On-Policy Distillation from Autoregressive to Diffusion Language Models》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.17397v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.17306", "title": "Calibrating LLM Explanations to Human Decisions with Reinforcement Learning", "published": "2026-10-12T18:59:35Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "455fc1fca7edb9f4f95eca0f229e7e59dbf9ec2a7e6cf58feb6de74255f3b383", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Calibrating LLM Explanations to Human Decisions with Reinforcement Learning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.17306v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.17246", "title": "LLM-to-Flow Distillation: Teaching Natural Language Goals to Scientific Generators", "published": "2026-10-12T18:11:58Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b28f27314a0f1c2e11a35e948253c8a65ca970c2377046a8292a758f558cf84c", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《LLM-to-Flow Distillation: Teaching Natural Language Goals to Scientific Generators》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.17246v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.17214", "title": "Density-Matrix Reinforcement Learning", "published": "2026-10-12T18:00:01Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "823f274891d5d281c6975a96efca6dd7f434f8e7c32a782ec75fb45847b75ac9", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Density-Matrix Reinforcement Learning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.17214v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.17174", "title": "Generative Flow on Symbolic Abstractions for Embodiment Capability Discovery", "published": "2026-10-12T17:46:33Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "027996de8a00a38f29d030604c204f61374f2a01ed80c0f2fe2de459923556f5", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Generative Flow on Symbolic Abstractions for Embodiment Capability Discovery》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.17174v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.17125", "title": "ToolHACE: A Dataset for Detecting Hallucinations in Tool-Based LLM Responses", "published": "2026-10-12T17:15:43Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "725f1a9cbf61fc942d7030b6cbe0c846fb5cb7172ad8e73cc2df55000a798f09", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《ToolHACE: A Dataset for Detecting Hallucinations in Tool-Based LLM Responses》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.17125v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.17042", "title": "AD4EDU: Benchmarking University Lecture Audio Description across Accessibility Standards, Metrics, and Learner Preferences", "published": "2026-10-12T16:15:07Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "644fb4b0cfb44791f836a9da1af7e96f5bbbb41ba90f28a0a48167cf99afdb4a", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《AD4EDU: Benchmarking University Lecture Audio Description across Accessibility Standards, Metrics, and Learner Preferences》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.17042v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16866", "title": "Efficient Resource-Shiftable RL Post-training with Generative Rewards", "published": "2026-10-12T14:33:16Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "cd430fca40babae176e1cf6d1c2dfe61f2b59cd484c19cce9819f7b448cb2413", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Efficient Resource-Shiftable RL Post-training with Generative Rewards》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16866v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16592", "title": "Prorsa: Unleashing the Potential of Prefill-Decode Disaggregation in Agentic Reinforcement Learning", "published": "2026-10-12T11:16:14Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "f0c739f8e1932938e22f1047c613fd236600a720d9983a1e8d3a22eac2896a5b", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Prorsa: Unleashing the Potential of Prefill-Decode Disaggregation in Agentic Reinforcement Learning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16592v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16533", "title": "Spectral-Driven Adaptive-Rank Optimizer for Large Language Model Reinforcement Learning with Verifiable Rewards", "published": "2026-10-12T10:21:21Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "86bb9f033fa1252f3a367fb889b72b4c9e74ab96ba9cbe4896dba923ad8b21a9", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Spectral-Driven Adaptive-Rank Optimizer for Large Language Model Reinforcement Learning with Verifiable Rewards》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16533v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16523", "title": "EAU: Environment-Agent-User Co-evolution for Training User-Centric Agents", "published": "2026-10-12T10:11:38Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "26aecb8a314655765da990d49c41cd3c02dbf0ba4de1007e67476cf01949d87a", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《EAU: Environment-Agent-User Co-evolution for Training User-Centric Agents》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16523v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16517", "title": "Trust-Calibrated VLA Policy Post-Training with Evolving Imagination", "published": "2026-10-12T10:08:28Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "ed442f5a14725135a8c4b20edb0c2d8f2a96ea03844326e61cfc76ab94fb7c9d", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Trust-Calibrated VLA Policy Post-Training with Evolving Imagination》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16517v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16489", "title": "Growing and Transferring Pathways for LLM Reinforcement Learning", "published": "2026-10-12T09:43:00Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "a3bf17f3776da8950ccf8662574e7a23650b236ee44fad2080f1879d6cfe4a71", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Growing and Transferring Pathways for LLM Reinforcement Learning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16489v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16368", "title": "LoopQuant: Accurate and Efficient Post-Training Quantization for Looped Transformers", "published": "2026-10-12T08:20:04Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "2399473d4fc47093a0a1389e303fcae337ce3a78ed7a7900336b362b05fee6a5", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《LoopQuant: Accurate and Efficient Post-Training Quantization for Looped Transformers》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16368v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16352", "title": "A Production-Grade, High-Efficiency Agentic RL Framework for Frontier LLMs with Validated Recipes", "published": "2026-10-12T08:08:16Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "d78b3d93a3c3ceca559419e532a284e0953b6ae6b262163c76b3839cae10b10d", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《A Production-Grade, High-Efficiency Agentic RL Framework for Frontier LLMs with Validated Recipes》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16352v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16318", "title": "Strong yet Controllable LLM Abstention without Learning to Abstain", "published": "2026-10-12T07:41:53Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "edc63284a02e814039534e12c95aa67f6a78591db71a64201f220816eb04ebfa", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Strong yet Controllable LLM Abstention without Learning to Abstain》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16318v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16226", "title": "NetRL: Powering Network Decisions via NetGRPO-Driven Reinforcement Learning", "published": "2026-10-12T06:03:33Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b35c486d991ca60ca6104f2fc14a2934926c6d835f27f8c7d8b37ebae22f8bae", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《NetRL: Powering Network Decisions via NetGRPO-Driven Reinforcement Learning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16226v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16172", "title": "EventLens: Event-Structure Reinforcement Learning for Video Understanding", "published": "2026-10-12T04:46:04Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "9a9862f0a07aff4a53f1c669e29a36fda605ba2c9c8c10aac3f2df3e68e42c30", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《EventLens: Event-Structure Reinforcement Learning for Video Understanding》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16172v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16105", "title": "Beyond Policy Rollouts: Embedding-Space Directional Exploration with Anchored Noise for LLM RLVR", "published": "2026-10-12T03:13:09Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "7fd2e4db337928acc1da59b02a7c0f152b9e19478bd3b270f3b605a6ffa69d78", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Beyond Policy Rollouts: Embedding-Space Directional Exploration with Anchored Noise for LLM RLVR》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16105v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16043", "title": "Beyond the greenwash: do LLMs' ecologically sound decisions hold up under pressure?", "published": "2026-10-12T01:36:17Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "a48f7016797e21abc28b5608bd14bb9eb704c2d8d722828bec261e19ded80ad9", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Beyond the greenwash: do LLMs' ecologically sound decisions hold up under pressure?》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16043v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15924", "title": "Scalable Weak-to-Strong Verifier Hardening for Agent Environments", "published": "2026-10-11T21:10:19Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "0069d1913f6a3be42a11efc256ca8cc16695d5f1080d259f67359b2c0c2f19e3", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《Scalable Weak-to-Strong Verifier Hardening for Agent Environments》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15924v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15899", "title": "Lightweight Scheduling Policies for Continuously Batched LLM Serving Across GPU Architectures", "published": "2026-10-11T20:27:58Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "ad478f3e34654c4cdfcdea94bb60afad52b9b47511b19b10a20cbbc41481104f", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《Lightweight Scheduling Policies for Continuously Batched LLM Serving Across GPU Architectures》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15899v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15894", "title": "Make Every Expansion Count: Learning to Search Efficiently in LLM Code Repair", "published": "2026-10-11T20:19:05Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b76dfecb0e1a29860df75c4826be27662b81bb06dd92914664e9f3a5b10afbf7", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《Make Every Expansion Count: Learning to Search Efficiently in LLM Code Repair》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15894v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15840", "title": "LLM Value Functions for Credit Assignment in Agentic RL", "published": "2026-10-11T18:46:50Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "1a504aefd6bff73ccab6f190437f9c2a1ad66628586a42e7c045a71285a3ca08", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《LLM Value Functions for Credit Assignment in Agentic RL》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15840v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15810", "title": "Adaptive Look, Flexible Thought: Learning How and How Long to Think in Multimodal Reasoning", "published": "2026-10-11T17:57:17Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "5ded7658e0297b133f1673023b0dddfa38658651e14be553279a7b9ff449b103", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《Adaptive Look, Flexible Thought: Learning How and How Long to Think in Multimodal Reasoning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15810v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15715", "title": "History Replays Itself: Unleashing Long-Context Capability in Linear Attention", "published": "2026-10-11T15:58:04Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "2b39139ba2676bde35cc07f1ddb5b28b5562aee0306854272804a796d4178d20", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《History Replays Itself: Unleashing Long-Context Capability in Linear Attention》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15715v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15578", "title": "Towards Robust Calibration Dataset for Domain-General Mixture of Experts Pruning", "published": "2026-10-11T12:44:36Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "d1c74b17d70fc600f6a69c9422337fb573737b65a5e31fa6f9374404713c3b1b", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《Towards Robust Calibration Dataset for Domain-General Mixture of Experts Pruning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15578v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15556", "title": "Vision-Predictor-Action: Towards Efficient Embodied AI via Neuro-Symbolic JEPA Predictors", "published": "2026-10-11T11:59:18Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "ce1ebfa94007e3468e4fb606a6a51c8000d2e5fdaec7d6c8812ace5a8e6c0bcc", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《Vision-Predictor-Action: Towards Efficient Embodied AI via Neuro-Symbolic JEPA Predictors》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15556v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15542", "title": "Apertus 1.5: From Open Text Models to Open Multimodal Reasoners", "published": "2026-10-11T11:34:43Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "384517bee433d46c06d451e38ec02c27c8088b4017f95937c558a7cefe97f7ce", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《Apertus 1.5: From Open Text Models to Open Multimodal Reasoners》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15542v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15514", "title": "SARATHI: Decoupled Data-Free Probing and Adaptive Reconstruction for Structured LLM Pruning", "published": "2026-10-11T10:27:26Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "c99e2b1ae03273663467ffc121365c57b0699fb45e39994a4c0c296e9ed7e77a", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《SARATHI: Decoupled Data-Free Probing and Adaptive Reconstruction for Structured LLM Pruning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15514v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15449", "title": "Hidden in Plain Sight: Advancing Harmful Content Understanding via Perception-Aware Adversarial Self-Play", "published": "2026-10-11T08:50:07Z", "revision": 2, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "ca213d0356afa20a8ef0f033641d2fc41394f7abce64064e4c632d8de06dc3d2", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《Hidden in Plain Sight: Advancing Harmful Content Understanding via Perception-Aware Adversarial Self-Play》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15449v2", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15398", "title": "GA-SVD: Generation-aware Singular Value Decomposition for LLM Compression", "published": "2026-10-11T06:53:27Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "77fabbb3f10b6d4dcc21edb4fcbc5b3a0a25aa20c4dc6372a4c092789f808d58", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《GA-SVD: Generation-aware Singular Value Decomposition for LLM Compression》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15398v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15371", "title": "A Sovereign, Indic-First, Reasoning Model", "published": "2026-10-11T05:52:29Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "45a9bb0832bd90c6ee41ff7edf3f61264460e6969914d0a32c1e581daa971761", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《A Sovereign, Indic-First, Reasoning Model》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15371v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15328", "title": "Anchor-and-Harvest Curriculum: A Fully Unmasked Non-LLM Curriculum for Craftax", "published": "2026-10-11T04:07:53Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "07762e4e2fc836cc4d2785652a795686942b5d52a8187c1d07ffef96ef826f6a", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《Anchor-and-Harvest Curriculum: A Fully Unmasked Non-LLM Curriculum for Craftax》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15328v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15277", "title": "Do correct answers have stable latent knowledge support?", "published": "2026-10-11T02:16:29Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "5f8a5f120f16713e974ee39e07be16e7d6d6d356405d9fb45097501db7b6b02d", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《Do correct answers have stable latent knowledge support?》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15277v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15154", "title": "Where Did We Go Wrong? Localizing Errors to Repair Multi-Turn Conversations", "published": "2026-10-10T20:10:59Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "dcf0a3b6d7e07efc7efb039bb8c15860c9ad0786c90c3c9bf3ec9dd3d0886a90", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《Where Did We Go Wrong? Localizing Errors to Repair Multi-Turn Conversations》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15154v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15097", "title": "Anchors Enhance Controllability and Reasoning of Diffusion Language Models", "published": "2026-10-10T18:14:59Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "83537a9f0b67c5e1f6fabc56f0191e879a6d0ce20d847daf52651b2793e93e93", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《Anchors Enhance Controllability and Reasoning of Diffusion Language Models》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15097v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15058", "title": "The Optimal Decoder is Tilted Temperature", "published": "2026-10-10T16:49:38Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "37568ece999c12eca67d48c207d73f09c951111f5fbc76557a6be15c54d9da90", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《The Optimal Decoder is Tilted Temperature》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15058v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15056", "title": "Select or Rewrite: Oracle-Guided Fine-Tuning for Generative ASR Post-Editing from N-Best Hypotheses", "published": "2026-10-10T16:48:27Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "c6047dc42b3cf0261b42c84dcf804d55f9a930ce6225d27a59c407033843d0cf", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《Select or Rewrite: Oracle-Guided Fine-Tuning for Generative ASR Post-Editing from N-Best Hypotheses》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15056v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15044", "title": "PAT: A Personality-Driven Augmentation and Transfer Learning Framework for Depression Detection on Social Media", "published": "2026-10-10T16:30:22Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "7ee013dbc8bbcfabcb41525369338614a625fb84d581d0cfafe9c4bbb0dc07b8", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《PAT: A Personality-Driven Augmentation and Transfer Learning Framework for Depression Detection on Social Media》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15044v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14976", "title": "Metric-Driven Reinforcement Learning for Object Hallucination Reduction in LVLM-Based Image Captioning", "published": "2026-10-10T14:47:47Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "11b0f932e49c9f38171c39d9b84ed13a7981598830c9c67b571a7a1660ad5d7e", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《Metric-Driven Reinforcement Learning for Object Hallucination Reduction in LVLM-Based Image Captioning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14976v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14937", "title": "Predicting how model propensities generalize out of distribution", "published": "2026-10-10T14:10:07Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b464859b2435b167a94737d4cbf779aba8fd532c9b7d9cc8eb48b17268397d07", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《Predicting how model propensities generalize out of distribution》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14937v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14936", "title": "CFG-Policy: Unified Diffusion Policy Learning Framework via Classifier-free Guidance", "published": "2026-10-10T14:09:24Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "3115169c8f01e96900f3d1b327c49f285192f4d1054a4ca8184bfd14fd3fb2f2", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《CFG-Policy: Unified Diffusion Policy Learning Framework via Classifier-free Guidance》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14936v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14913", "title": "Bridging Risk Approximation Gaps in Model Predictive Task Sampling via In-Context Modeling", "published": "2026-10-10T13:35:02Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "fbf4291ca50eff0247fcf74062e6cb1c603954c9e0a8cf5e1155c805a7a03fcb", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《Bridging Risk Approximation Gaps in Model Predictive Task Sampling via In-Context Modeling》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14913v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14883", "title": "MARL-BattleGrounds: A Novel JAX-native Benchmark for Heterogeneous and Competitive Multi-agent RL", "published": "2026-10-10T12:57:40Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "580f1ef530954896a630d1bf32cbf61c4220040a801dffa3444b6010137dae7b", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《MARL-BattleGrounds: A Novel JAX-native Benchmark for Heterogeneous and Competitive Multi-agent RL》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14883v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14860", "title": "Terrarium: Growing Personal Assistants via Scalable Synthesis of App Worlds", "published": "2026-10-10T12:30:31Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "c230d28f812e3b905dd14267842639b546485f24a7eb672093808649747d698b", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《Terrarium: Growing Personal Assistants via Scalable Synthesis of App Worlds》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14860v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14763", "title": "CUSE: Fine-Grained Multimodal Unlearning with Counterfactual Guidance and On-Policy Self-Distillation", "published": "2026-10-10T09:37:30Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b3941f7b81410ab076df8625d0dc2536f22342c314f3724d6c655089a049745d", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《CUSE: Fine-Grained Multimodal Unlearning with Counterfactual Guidance and On-Policy Self-Distillation》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14763v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14612", "title": "LLM-Augmented Bayesian Optimization with Adaptive Policies for Chiplet Design Space Exploration", "published": "2026-10-10T05:55:09Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "3c96735d0eec2594c0f047b1e01dd39a1f820ce3394d7ebf1f2a6e1786d84a3d", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《LLM-Augmented Bayesian Optimization with Adaptive Policies for Chiplet Design Space Exploration》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14612v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14578", "title": "Uniform Information Density-Based Preference Optimization", "published": "2026-10-10T04:55:56Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "15d79268950abe0c4dcb729bfab3d902b20b0a1320e75df3d5ce726a9ac19e59", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《Uniform Information Density-Based Preference Optimization》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14578v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14571", "title": "Dissecting the Carbon-SLO Tension in Geo-Distributed LLM Serving", "published": "2026-10-10T04:38:49Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "1fd75cfdab3910be99b1f303737f94c14cfd2fbd467da49ea59fefd4f9d952f2", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《Dissecting the Carbon-SLO Tension in Geo-Distributed LLM Serving》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14571v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14519", "title": "Query-Adaptive Neuron Steering for LLMs", "published": "2026-10-10T03:30:09Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "d22234d7fa8fb9bbd7657eb49b5a597e49705ee1efa324e80493874dcbbb85e2", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《Query-Adaptive Neuron Steering for LLMs》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14519v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14460", "title": "Re-executing Coding-Agent Evaluations: Unchanged Decisions and Unresolved Policy Risk", "published": "2026-10-10T01:38:03Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "82151053e93dab0d948eb831397825d200bbc352f79a7a0160aeb94e0ce89018", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《Re-executing Coding-Agent Evaluations: Unchanged Decisions and Unresolved Policy Risk》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14460v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14405", "title": "SPROUT: Scaling Agentic Reinforcement Learning for Long-Horizon Software Engineering", "published": "2026-10-09T23:58:44Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "25811d8d500519850bfe527a2f1be5c10f620f06406f948460c4ad80f5212ff3", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-09，早于2026-10-13窗口起点；《SPROUT: Scaling Agentic Reinforcement Learning for Long-Horizon Software Engineering》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14405v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14400", "title": "Sparse-TD-JEPA: Improving Zero-Shot Reinforcement Learning With Sparse Latent Task Representations", "published": "2026-10-09T23:42:51Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "a637c4b43e98e82f33502485e63f860d547440c07c6eaeebd6c997f09d2c648e", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-09，早于2026-10-13窗口起点；《Sparse-TD-JEPA: Improving Zero-Shot Reinforcement Learning With Sparse Latent Task Representations》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14400v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14393", "title": "SciCurateBench: Curating Experimental Data from Heterogeneous Scientific Literature with Verbal-Action Agents", "published": "2026-10-09T23:32:35Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "acda66cb23b085feda72412485437c4f1975c5793a971f69915decbb6973ff38", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-09，早于2026-10-13窗口起点；《SciCurateBench: Curating Experimental Data from Heterogeneous Scientific Literature with Verbal-Action Agents》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14393v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14346", "title": "Evaluating LLMs as Agentic Workflow Executors in Real-World Customer Support Skills", "published": "2026-10-09T21:45:51Z", "revision": 2, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "c36ed56462fe0a356868e2da1237737b5e61036e8c2fb9000f0e99f7e1111598", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-09，早于2026-10-13窗口起点；《Evaluating LLMs as Agentic Workflow Executors in Real-World Customer Support Skills》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14346v2", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14309", "title": "ASSET: Adaptive Speech Segmentation via Reinforcement Learning", "published": "2026-10-09T20:36:45Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "0d8bb76b833ae6c4426dfcd9ba1c0d73cd515d84d6cc4b2cd8ccdc6173835ce2", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-09，早于2026-10-13窗口起点；《ASSET: Adaptive Speech Segmentation via Reinforcement Learning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14309v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14270", "title": "The Thinking Leakage in Reinforcement Learning for LLMs", "published": "2026-10-09T19:43:51Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "4067794d60a28d532a4131b736188952c73e414c4bb73b1202b30cfad8231bb5", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-09，早于2026-10-13窗口起点；《The Thinking Leakage in Reinforcement Learning for LLMs》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14270v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14238", "title": "Agentic Data Engineering for LLMs: System Design and a Controlled Empirical Study", "published": "2026-10-09T18:56:21Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "2eb8f128e17d640d46b964dac8e9413790c680d84c1bfcdfcde4f101b2295a62", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-09，早于2026-10-13窗口起点；《Agentic Data Engineering for LLMs: System Design and a Controlled Empirical Study》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14238v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14180", "title": "Unconstrained CTI-REALM with Open-Weights", "published": "2026-10-09T18:00:33Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "6f5c0dfb1a5b01c0b217a48411f7d35a65831b1766336a9b9ed0e159ff92d00e", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-09，早于2026-10-13窗口起点；《Unconstrained CTI-REALM with Open-Weights》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14180v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14118", "title": "ClawLoop: Self-Improving Agents via Harness-Native Training Loop", "published": "2026-10-09T08:52:34Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "f513590cbdd50f35571369a4d0b9155ba6726075ed478c4d7b044860ebb87fea", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-09，早于2026-10-13窗口起点；《ClawLoop: Self-Improving Agents via Harness-Native Training Loop》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14118v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.12828", "title": "TreeHCA: Hindsight-Guided Exploration and Credit Assignment for Multi-Turn LLM Agents", "published": "2026-10-09T00:23:39Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "475218baf92c7bc74ffdfdab6341ba803b4f96472d4a14a3bc467ee276a70964", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-09，早于2026-10-13窗口起点；《TreeHCA: Hindsight-Guided Exploration and Credit Assignment for Multi-Turn LLM Agents》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.12828v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.12313", "title": "Verdict Without the Rule: Diagnosing and Auditing Regulatory Rule Sensitivity in LLM Compliance Systems", "published": "2026-10-08T16:59:40Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "bbfceb305a239f380baece88abc38a6875c05f206903098957b82d5ab312f9dc", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-08，早于2026-10-13窗口起点；《Verdict Without the Rule: Diagnosing and Auditing Regulatory Rule Sensitivity in LLM Compliance Systems》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.12313v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.11599", "title": "Large Language Model Turnover Undermines Screening for Artificial Intelligence-Assisted Scientific Writing", "published": "2026-10-08T09:44:08Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "e89d2985fc2a39cd60ec448ea75e3a050ae5020b2a5cc96e6b6104fdeacd1b07", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-08，早于2026-10-13窗口起点；《Large Language Model Turnover Undermines Screening for Artificial Intelligence-Assisted Scientific Writing》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.11599v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.10848", "title": "Amortized Off-Policy Evaluation for LLMs", "published": "2026-10-07T19:51:59Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "7fef2683183dbda156cf542b1ea2d5d844982ea1b55b1a2a05abe6393b91c185", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-07，早于2026-10-13窗口起点；《Amortized Off-Policy Evaluation for LLMs》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.10848v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.10506", "title": "Validity Without Ground Truth: What Stated-Preference Economics Offers the Evaluation of Language Models", "published": "2026-10-07T17:51:07Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "6246640a25686ab175f894ffbbb86b1cb7037305c1ed79caa16f3b8841118360", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-07，早于2026-10-13窗口起点；《Validity Without Ground Truth: What Stated-Preference Economics Offers the Evaluation of Language Models》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.10506v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14063", "title": "Auditing LLM Program Search for Bermudan Stopping with Primal--Dual Certificates", "published": "2026-10-07T09:00:25Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "d6fb702df423e5eb557646ba174a74f6696d5d7633d38add2520562fe2b7059c", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-07，早于2026-10-13窗口起点；《Auditing LLM Program Search for Bermudan Stopping with Primal--Dual Certificates》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14063v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.09458", "title": "Right Number, Wrong State? Measuring Cross-Jurisdiction Substitution in LLM Recall of State Policy", "published": "2026-10-07T05:12:42Z", "revision": 2, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "1ef14aaa5b000385a8e75552ae593a94cd689459e52904c36e27c1837596aec5", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-07，早于2026-10-13窗口起点；《Right Number, Wrong State? Measuring Cross-Jurisdiction Substitution in LLM Recall of State Policy》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.09458v2", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.08402", "title": "VETTA: Coordinating Turn- and Token-Level Credit Assignment for Multi-Turn LLM Agents", "published": "2026-10-06T14:15:11Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "ee2e66fa0d63011885a350a439c706726b08530157828a349e86be9bfda445bc", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-06，早于2026-10-13窗口起点；《VETTA: Coordinating Turn- and Token-Level Credit Assignment for Multi-Turn LLM Agents》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.08402v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.07580", "title": "LOGIC: An LLM Benchmark for Intent-Grounded Change Impact in Aerospace Electrical Systems", "published": "2026-10-06T01:14:09Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "8488700b659b2c24c608aef9326c1caaba0205ca9227db30a05282193982706d", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-06，早于2026-10-13窗口起点；《LOGIC: An LLM Benchmark for Intent-Grounded Change Impact in Aerospace Electrical Systems》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.07580v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.07563", "title": "HouseholdBench: Evaluating Large Language Models as Predictors of Household Economic Behavior", "published": "2026-10-06T00:50:28Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "ada30b732191e177879dbf607bc39b0cfe37e2948c7b595bcf638b72edb75bb5", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-06，早于2026-10-13窗口起点；《HouseholdBench: Evaluating Large Language Models as Predictors of Household Economic Behavior》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.07563v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.05966", "title": "HuatuoGPT-3: RL-Only Domain Adaptation from Base Models", "published": "2026-10-05T08:16:36Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "dead27b376beaf049e2c00dbbc00f03c4f2b5c93a3d78da3fafa982ef48abd81", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-05，早于2026-10-13窗口起点；《HuatuoGPT-3: RL-Only Domain Adaptation from Base Models》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.05966v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.05450", "title": "Groupwise Distortion Guarantees for Preference-Based Alignment", "published": "2026-10-04T18:45:36Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "591b46c97c5bfc42ce7a5b5da69dfd72e9b5fa213e7c76fa27b9c36d70fee2bd", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-04，早于2026-10-13窗口起点；《Groupwise Distortion Guarantees for Preference-Based Alignment》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.05450v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.05077", "title": "How Execution Assumptions Change Short-Horizon Sharpe Rankings: Evidence from a Synthetic Trading Benchmark", "published": "2026-10-04T09:22:10Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "2f7e7dc2e38579f6af052a080ad4919c3749f3dd593758b6df2a8e6a4b11781b", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-04，早于2026-10-13窗口起点；《How Execution Assumptions Change Short-Horizon Sharpe Rankings: Evidence from a Synthetic Trading Benchmark》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.05077v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.04928", "title": "Prompt Dominance and Asymmetric Verifier Costs: Empirical Ablations of GRPO at 1B Scale on GSM8K", "published": "2026-10-04T04:13:20Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b1d06a9f6bbf1f44caaceb46cc5bdd7daef086a13f6d9802f51664584dc3f7ff", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-04，早于2026-10-13窗口起点；《Prompt Dominance and Asymmetric Verifier Costs: Empirical Ablations of GRPO at 1B Scale on GSM8K》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.04928v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.04620", "title": "Stance Drift: How AI-mediated Communication Distorts Our Message", "published": "2026-10-03T16:09:37Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "e63ae3daa8fdef9268c811fadc7a960a316a80ac3be5052a1fb205ac03ff317c", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-03，早于2026-10-13窗口起点；《Stance Drift: How AI-mediated Communication Distorts Our Message》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.04620v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.04537", "title": "PhaseGate: Phase-Aware CPU Retrieval Scheduling for On-Device LLMs on Unified Memory", "published": "2026-10-03T14:07:04Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "441c77179ba41b4b80997b3444f83b6017b7873e07c25f9f9dd69d25e2bdb6c8", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-03，早于2026-10-13窗口起点；《PhaseGate: Phase-Aware CPU Retrieval Scheduling for On-Device LLMs on Unified Memory》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.04537v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.04364", "title": "A Bird's-Eye View of Iterative Reward Design", "published": "2026-10-03T08:02:22Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "6f82ebfdfcff023f9961325411a5ea7a1f99f2b38be00176b08d1dcaa7777f58", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-03，早于2026-10-13窗口起点；《A Bird's-Eye View of Iterative Reward Design》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.04364v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.04137", "title": "Dynamic Harness Search: Building Multi-Agent Systems Per-Query via Prediction", "published": "2026-10-02T23:18:55Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "ba938c7fa3e41849bfeb5384608776bd92f188a9e6a17e9a77506ff0c27f8f50", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-02，早于2026-10-13窗口起点；《Dynamic Harness Search: Building Multi-Agent Systems Per-Query via Prediction》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.04137v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.04089", "title": "Pareto-Dominant Clarification: Post-Training Coding LLMs via PPO-Lagrangian Budget Constraints", "published": "2026-10-02T21:54:46Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "6a2c71c780b4ddc08e4f43ffa48936cf391f04cc5cc056949c037d4147a3676d", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-02，早于2026-10-13窗口起点；《Pareto-Dominant Clarification: Post-Training Coding LLMs via PPO-Lagrangian Budget Constraints》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.04089v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.03667", "title": "Planning to Learn", "published": "2026-10-02T17:38:28Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b200fc0eac42b12f466070c86a812d97dd39284081a33e0eada4cbdc998b01c1", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-02，早于2026-10-13窗口起点；《Planning to Learn》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.03667v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.03458", "title": "A Near-Zero Monitor Readout Is Not Evidence of Behavioral Control", "published": "2026-10-02T15:33:56Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "57941c5a2e6d8a3415f5f33d5fc321438e5774f6657d238c95b925e7c5aedf70", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-02，早于2026-10-13窗口起点；《A Near-Zero Monitor Readout Is Not Evidence of Behavioral Control》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.03458v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.02488", "title": "Harnessing LLMs as Agents: What Does It Cost?", "published": "2026-10-01T21:05:49Z", "revision": 2, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "c1dadee4e4513751f6b855f9efd497280b403d24ca00dd3796422dddec3058de", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-01，早于2026-10-13窗口起点；《Harnessing LLMs as Agents: What Does It Cost?》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.02488v2", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.02438", "title": "Are you Synthesizing or Recalling? Evaluating LLMs on Algorithmic Code Retrieval", "published": "2026-10-01T20:04:14Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "c10100a0a693b7d0f1892733f3614bbee4bfb9dd57f09588677c3fcb1de5b884", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-01，早于2026-10-13窗口起点；《Are you Synthesizing or Recalling? Evaluating LLMs on Algorithmic Code Retrieval》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.02438v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.02381", "title": "Latent-MOPD: Latent Multi-Teacher On-Policy Distillation", "published": "2026-10-01T19:01:56Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "6bee0595955acb40bb3cf513bec6be6015086377d116d0a621d0db03a4a53393", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-01，早于2026-10-13窗口起点；《Latent-MOPD: Latent Multi-Teacher On-Policy Distillation》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.02381v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.01984", "title": "Universal Byte-Level Encoding: UTF-8/UTF-16 Routing to Reduce Cross-Script Token-Budget Disparities", "published": "2026-10-01T16:26:10Z", "revision": 2, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b634ed894573d0816f7ee4a2086be9fbd0976ad80ad6f3347f022258475e1430", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-01，早于2026-10-13窗口起点；《Universal Byte-Level Encoding: UTF-8/UTF-16 Routing to Reduce Cross-Script Token-Budget Disparities》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.01984v2", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.01652", "title": "Iterative Policy Refinement through Semantic Rollout Analysis", "published": "2026-10-01T13:15:42Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "015ca4b7b35d6dcc76c76d1bd7a289ca87ca8da22ad19cded4aa98e24a969e8d", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-01，早于2026-10-13窗口起点；《Iterative Policy Refinement through Semantic Rollout Analysis》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.01652v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.03799", "title": "StepCAD: Mesh-to-CAD Code Generation via LLM Policy and Geometry-Guided Search", "published": "2026-10-01T05:57:05Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "0fe8d91f7e2969f0d9363987d7f25749bfc93da94b61b284c2712339a7f28b17", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-01，早于2026-10-13窗口起点；《StepCAD: Mesh-to-CAD Code Generation via LLM Policy and Geometry-Guided Search》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.03799v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.00309", "title": "Tokenized Key-Gated Adapter Routing: A Secure Access Control Mechanism Against Private Data Leakage in LLMs", "published": "2026-09-28T19:52:42Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "f9d69a23df53783d22221b3db16f122c51cccba6961cbf49a3ca040c2fc075ac", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-09-28，早于2026-10-13窗口起点；《Tokenized Key-Gated Adapter Routing: A Secure Access Control Mechanism Against Private Data Leakage in LLMs》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.00309v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14000", "title": "VH-GPO: Verified Handoff for LLM Agent Reinforcement Learning", "published": "2026-09-27T05:25:37Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "196138722cffd867f7cf6b3d77fbcdaaec808a21b7ab8fbddad1cd5003df5b0b", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-09-27，早于2026-10-13窗口起点；《VH-GPO: Verified Handoff for LLM Agent Reinforcement Learning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14000v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.13995", "title": "GameSmith: Scaling Zero-Sum Self-Play with Synthesised Games", "published": "2026-09-26T23:07:02Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "5e6487372c502af6bca81ef58afc6de6cf5d84a48f42cee2e3680ff082deb9fa", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-09-26，早于2026-10-13窗口起点；《GameSmith: Scaling Zero-Sum Self-Play with Synthesised Games》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.13995v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.20334", "title": "TreeFlex: Agents Navigating Their Own Search Trees", "published": "2026-10-14T15:16:50Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "a406192e99f95a01775ccce5870e6e40a41539702e48989290d5ff21e9e4e4de", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "TreeFlex 的Reflection/Execution共享policy导航自身search tree，并以tree-aware GRPO训练，真实工具/状态恢复不可换成预设搜索树。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§2 TreeFlex", "§3 Experimental Setup", "Appendix D.1 Training and Evaluation Tasks" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.20334v1", "source_sha256": "f84d203aa4df85e3709d6526d6b33307ae0e24cadd13da3b37f9b387f4efbc6b", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.20334v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。4B deep-research、9B SWE-Gym/SWE-bench，600任务子集（500train/100val）和多基线RL；完整容器/rollout/训练总时未核定", "scale": "4B deep-research、9B SWE-Gym/SWE-bench，600任务子集（500train/100val）和多基线RL；完整容器/rollout/训练总时未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.20253", "title": "Repairing Speech Access After Text Updates with Frozen Affine Fields", "published": "2026-10-14T14:24:38Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "31ed1e312d85bb75b54e9258e5443074a8780422ce91d5df1d3ac102d46bb9bb", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "AFFIRM 以冻结base修正加零均值query-conditioned affine correction修复speech access；不能把operator参数小误认整个获取/训练链路便宜。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 AFFIRM", "Appendix A.2 acquired checkpoints", "Appendix A.3 operator optimization", "Appendix A.6 hardware" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.20253v1", "source_sha256": "8382c63c0bdb7e320b535e4bfb2888cdf9944a49902d3135c2c41cf702ba6ff0", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "source_attempts": [ { "url": "https://arxiv.org/html/2610.20253v1", "result": "retrieved", "sha256": "8382c63c0bdb7e320b535e4bfb2888cdf9944a49902d3135c2c41cf702ba6ff0" } ], "cost_assessment": { "source_url": "https://arxiv.org/html/2610.20253v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。Qwen2.5-Omni-7B全文本参数获取10epochs、14,267 PopQA或4,000 TOFU，加CosyVoice合成语音/多operator训练；单A800执行，完整总时未核定", "scale": "Qwen2.5-Omni-7B全文本参数获取10epochs、14,267 PopQA或4,000 TOFU，加CosyVoice合成语音/多operator训练；单A800执行，完整总时未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.19796", "title": "Learning Perception to Action Rules through Reinforcement Learning for Theory-of-Mind Reasoning", "published": "2026-10-14T08:40:06Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "4bfa168f9b397254e736a284c39f43a584602ec14133f57a0a843a161b00f353", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "ToM rule-GRPO 显式连接perception/belief/action并训练规则生成；须在development选择checkpoint且test隔离，不能用手写if-then取代模型学习。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§2 Methods", "§2.5 Training" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.19796v1", "source_sha256": "703bc7b7db7b0dc41e040332304cab62b2cbe0c697141fdb8adb4c939864aa71", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.19796v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。Qwen2.5/Llama3多规模、四ToM benchmarks，SFT/GRPO均2epochs，H100/A100/A40；多模型多格式总运行成本未核定", "scale": "Qwen2.5/Llama3多规模、四ToM benchmarks，SFT/GRPO均2epochs，H100/A100/A40；多模型多格式总运行成本未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.19476", "title": "What Does LALM Post-training Learn? Rethinking Audio Understanding and Reasoning Capabilities", "published": "2026-10-14T02:54:57Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "6fbcdfc1aae140e4d9a86ceca5e88be15962d3a341cf7685d4c189f2b711cdde", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "evaluation_analysis_reference_only", "reason": "LALM研究用informative audio/silence与题型变换的controlled experiments分离音频能力和文本捷径，没有提出新的训练算法；不能把选择题增益当音频理解提升；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "§3 Controlled Evaluation Protocol", "§4 Experimental Setting", "§5–§6 controlled analyses" ], "fulltext_review_completed": true, "review_level": "primary_scope_or_evaluation_passages_review", "source_url": "https://arxiv.org/html/2610.19476v1", "source_sha256": "18de70deee6621b708f7255cc8206362296d399c3ce51e1f0442935911f7aee7", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。" }, { "id": "2610.18853", "title": "Two Clocks: Why Larger Rollout Batches Learn Slower but Fail on Schedule in Reinforcement Learning with Verifiable Rewards", "published": "2026-10-13T17:58:57Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "27d94cdc485d8b5869519163ee2769e2f903b3745af8c795ce58da6edef53736", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Two Clocks区分RLVR采样进度与collapse时钟，检验batch/sqrt-LR缩放及switch schedule；训练实证不能由二次近似数值图代替。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 Two-Clock Model", "§4 Experimental Setup", "Hardware and compute." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18853v1", "source_sha256": "ea7627bdc5b397728953646b0ddff3f218aba2c793ad246796335e04844919c8", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18853v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。1.5B每run4 RTX A6000、7B每run4 H100；作者总约2,000 GPU-hours", "scale": "1.5B每run4 RTX A6000、7B每run4 H100；作者总约2,000 GPU-hours", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18766", "title": "Rubrics from the Task, Verdicts from Evidence: Pointwise Rubric Rewards with Tool-Augmented Judges", "published": "2026-10-13T17:27:23Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "67bd58328a9c4a2424193584f9592a6fe909fbd6c639b2148a915754aa5272f0", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "ReVeR 以instruction-only rubric和tool-augmented pointwise judge做label-blind蒸馏，标签只过滤teacher outputs，不能泄漏到生成。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 ReVeR", "Appendix B Training and Evaluation Details" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18766v1", "source_sha256": "cfc56f08d0e279937c36d72f23696faa52130187ee26080d78ec86092473bb91", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18766v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。4B generator/judge、7B policy，多阶段30k/60k traces；训练列给出8/8/5/4/5 GPUs，最长16,384tokens", "scale": "4B generator/judge、7B policy，多阶段30k/60k traces；训练列给出8/8/5/4/5 GPUs，最长16,384tokens", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18755", "title": "PreciseCoder: Faithful and Generalizable Debugging Beyond Unit-Test Rewards", "published": "2026-10-13T17:19:10Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "943a8ea2a180109ef8b721e1ab66ee3e09936c69a722f6b029b200512e077e16", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "PreciseCoder 在unit-test reward外增加语义核验的局部edit信用，并评估仓库级迁移；不能只以减少diff行数作为正确性奖励。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 PreciseCoder", "Appendix C.1 Hyperparameters" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18755v1", "source_sha256": "61c78d718d81f699c6c52ccc8f0b89edc0002205598557bb65feafbcbbe70d6c", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "source_attempts": [ { "url": "https://arxiv.org/html/2610.18755v1", "result": "retrieved", "sha256": "61c78d718d81f699c6c52ccc8f0b89edc0002205598557bb65feafbcbbe70d6c" } ], "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18755v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。4B用4 GH200、9B用4 H200、27B用8 GH200，2epochs约325–328steps，4B/9B各三seeds，group6", "scale": "4B用4 GH200、9B用4 H200、27B用8 GH200，2epochs约325–328steps，4B/9B各三seeds，group6", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18738", "title": "Enabling Token-wise Importance Sampling in Off-Policy RLVR with Proximal KL Regularization", "published": "2026-10-13T17:09:23Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "d0ea3f5c41fd220a9d8026d998da8e084260e1824175ce56e64e96e63ca128c2", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Proximal KL 将proximal reference与behavior policy分离，用token-wise IS及future KL credit替代clip；tabular保证的coverage/expressiveness条件不自动适用于Transformer。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 proximal update", "§4 Algorithm", "§5 guarantees", "§6 Empirical Experiments", "Appendix C.1 Training and Evaluation Protocols" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18738v1", "source_sha256": "ec605429a2d566040c7bb82e4e0bf57e2c38e349f5ddeed2b04256ffb47e40ca", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18738v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。1.7B/4B，七matched off-policy比较、两LR、staleness/reuse和八math benchmarks，另on-policy及horizon消融；完整训练/扫评总时未核定", "scale": "1.7B/4B，七matched off-policy比较、两LR、staleness/reuse和八math benchmarks，另on-policy及horizon消融；完整训练/扫评总时未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.18695", "title": "A Refresh Interval Does Not Fully Specify a Refresh Schedule: Learning-Rate-Dependent Placement in Kronecker-Preconditioned Language Model Training", "published": "2026-10-13T16:46:34Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "e0ec0377be6a6cc134fc780ece029bd420153ef2cdd8087b60bc77b4c840a1ea", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Kronecker refresh placement研究固定refresh数下的LR相关排序反转，是训练调度实证，不以单次数值测试冒充完整跨LR结论。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 Refresh Placement", "§4 Experimental setup", "Models, evaluation and compute." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18695v1", "source_sha256": "ffef5974ad8df24d716838b19367ecd8e1bf04f69e90eaa3fab26f6c11d0fe43", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18695v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。64M–929M GPT decoders，1,973完成runs，总约920 A100-hours；单run6.2–55.4min不等于整项研究便宜", "scale": "64M–929M GPT decoders，1,973完成runs，总约920 A100-hours；单run6.2–55.4min不等于整项研究便宜", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18613", "title": "Does self-distillation improve reasoning?", "published": "2026-10-13T15:59:36Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "63d82e15e00a70e47d560155648ab8047786befade196847f53b0ee05a5ad078", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Self-distillation reasoning复核区分OOD泛化、thinking-mode迁移和梯度对齐；保留其负结果，不能重新标成普适收益。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4 Experimental results", "Appendix E Implementation details" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18613v1", "source_sha256": "8ffaedee230645c077fc2c6ecad71bb6f5f15bb2e2c57ea0f1dc5958a76a6971", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18613v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。1.7B–14B/Qwen与7B OLMo，8 H100节点；SDPO代码run约40h，其他约2–6h，评测/梯度分析另计", "scale": "1.7B–14B/Qwen与7B OLMo，8 H100节点；SDPO代码run约40h，其他约2–6h，评测/梯度分析另计", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18517", "title": "Imprinting: Knowledge Distillation Can Shift Student Behavior", "published": "2026-10-13T15:09:59Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "8581ee21bc4b8f552bfacecaf6d5521ca887ccc1e0739c775a1f1156e69c542f", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Imprinting研究教师安全倾向经普通任务distillation迁移，并比较divergence-token mask/rewriting与后续偏好优化；是多教师-学生因果比较，不是单步loss检查。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§2 Experimental Design", "§4 divergence tokens", "Appendix D Training and evaluation details" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18517v1", "source_sha256": "973cb741e485c6cb781e8c7a7718ca5d581ae81d99d559f56c46ccae88533e06", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18517v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。H200 141GB单卡jobs，单student fine-tune不足2h但还需BC-LAT教师、生成、多配对和安全评测；完整活动总成本未核定", "scale": "H200 141GB单卡jobs，单student fine-tune不足2h但还需BC-LAT教师、生成、多配对和安全评测；完整活动总成本未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.18484", "title": "ReTerm: Recovery Training for Self-Improving Terminal Agents", "published": "2026-10-13T14:44:24Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "47012071adeb87cd0421e7c37631c32db2c57189c94f1497484c7c341490c823", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "ReTerm 先以恢复demonstration SFT，再从自身failure states fork多条恢复rollout做ReTerm-GRPO。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 ReTerm", "Training Details." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18484v1", "source_sha256": "c788a3752546ced326fce4668895df99744fa8392f35a8552a4d405876528640", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18484v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。Qwen3.5-9B，ReTerm-10K：6K SFT/4K RL；16 A100 80GB，global batch160、65k context、32turn", "scale": "Qwen3.5-9B，ReTerm-10K：6K SFT/4K RL；16 A100 80GB，global batch160、65k context、32turn", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.18472", "title": "Simplifying and Improving On-Policy Distillation for Few-Step Diffusion", "published": "2026-10-13T14:35:46Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "510ef758ec7dea01c3a6a1d95b8c741871429518bcf9ad89d239b36970e3529e", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "source_exception_archived", "reason": "摘要描述图像/3D/视频/音视频few-step diffusion的SMD/Pivot蒸馏。官方HTML返回404，官方PDF先后超出8MB及24MB有界读取上限；并非声称源站没有PDF。未读方法/资源正文，不能把文档预算例外伪装为高成本算法结论。本次明确非阻塞归档，不自动重试或进入实现/evolve。", "locations": [ "official arXiv abstract", "official HTML retrieval attempt", "official PDF bounded retrieval attempt" ], "fulltext_review_completed": false, "review_level": "abstract_scope_review_with_document_exception", "source_url": "https://arxiv.org/html/2610.18472v1", "source_sha256": null, "evidence_limitations": "决定仅基于官方元数据/摘要范围或明确的来源例外；未核验正文算法或资源，不以metadata hash冒充文档hash。", "source_attempts": [ { "url": "https://arxiv.org/html/2610.18472v1", "result": "unavailable", "error": "HTTP Error 404: Not Found" }, { "url": "https://arxiv.org/pdf/2610.18472v1", "result": "unavailable", "error": "bounded PDF document limit exceeded" }, { "url": "https://arxiv.org/pdf/2610.18472v1", "result": "unavailable", "error": "bounded PDF document limit exceeded (24000000 bytes)", "read_limit_bytes": 24000000 } ] }, { "id": "2610.18459", "title": "Teacher-Guided Tool Learning for Small Language Models via Bidirectional KL Constraints", "published": "2026-10-13T14:26:41Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "be7175451cda096271b2d80d6415837cfe3a3efe4bc572f5f25ed637ac351bfc", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "SeTPO 以低reward轨迹reverse-KL advantage加持续forward-KL示范监督协调探索；0.5B能放入单卡不等于完整对照预算已核定。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4 SeTPO", "Datasets and Models.", "Appendix A.1 Models, Data, and Hardware", "Appendix A.2 Training Configuration" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18459v1", "source_sha256": "930d9b25014d7c50c422e97e99f4f029baf5745661d061ebecde01d2db49c175", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18459v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。0.5B/1.5B/3B学生配1.5B/7B/14B教师；最小用1 L40S，4,000训练例×3epochs、4096上下文，含BFCL/API-Bank/通用评测；总时未核定", "scale": "0.5B/1.5B/3B学生配1.5B/7B/14B教师；最小用1 L40S，4,000训练例×3epochs、4096上下文，含BFCL/API-Bank/通用评测；总时未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.18082", "title": "Failure Recovery via Belief Updates and World Modeling for Long-Horizon Agents", "published": "2026-10-13T10:05:06Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "abb69b39db6823800f9d2e807d821f2de6dad3b0e76004936b1a9f89d5e3db02", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "BELIEFWEAVER 以跨尝试relational belief和IWM预测修复方向，且将IWM组件SFT/OPD到9B学生，非简单backtrack状态机。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4 BELIEFWEAVER", "§5.1 Experimental Setup", "§5.3 Distillation Results" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18082v1", "source_sha256": "e834846d3e1a1194aa52b53a1ac421d55925ddee3f4b4f3b1f900b898989019d", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18082v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。DeepSeek V4 Pro teacher、9B学生、三个terminal suites，多尝试/rollout/teacher API和训练；54.8%相对API节约未核定绝对全链路成本", "scale": "DeepSeek V4 Pro teacher、9B学生、三个terminal suites，多尝试/rollout/teacher API和训练；54.8%相对API节约未核定绝对全链路成本", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.18067", "title": "WorkRover: Scaling Long-Horizon Cowork Agent Training Via Recursive Environment Evolution", "published": "2026-10-13T09:52:02Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "d2a91ec0f7caecdc29fc8a2a6b8c150ed552eb6042e89860ffe7e7cd1f6096f6", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "WorkRover 用dependency-closed factorial replay选择最小validated环境修订，并以VTCA给真实workspace进度分配credit；不能把静态registry标签当可执行自进化。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 WorkRover", "§4.1 Experimental Setup", "Appendix D.2 Matched VTCA Comparison" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18067v1", "source_sha256": "f1fbdd2ecb2a54b7723372118e2b0ac818420f3a3dd280c2486c8e064cf486d6", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18067v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。Qwen3.5-35B-A3B、长时程cowork环境合成/递归evolution/GRPO，JobBench/OfficeQA/GDPval；训练与构造总成本未核定", "scale": "Qwen3.5-35B-A3B、长时程cowork环境合成/递归evolution/GRPO，JobBench/OfficeQA/GDPval；训练与构造总成本未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.18011", "title": "Correcting or Continuing? What On-Policy Distillation Learns from Student Errors", "published": "2026-10-13T09:10:58Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "9ddfac95e2ff03c7dc1b36a25b030c61cddbcbd9a1bbc1725235eae78f27c8b9", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "错误集中OPD研究比较五机制、八配置与非失败集中positive control；保留accuracy下降和训练KL仍下降的负结果，不把diagnostic当改进。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 Experiment and Results", "Appendix F Reproducibility" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18011v1", "source_sha256": "ec4296e77b54b64d04e3bb5015ec5a64800f233d65a3c98c1ee25513a2ea52f1", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18011v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。单H200 NVL 141GB共享teacher/student，Qwen3/Phi4、八配置、多数学/迁移评测；较单A100/A30的内存/全活动总成本未能廉价核定", "scale": "单H200 NVL 141GB共享teacher/student，Qwen3/Phi4、八配置、多数学/迁移评测；较单A100/A30的内存/全活动总成本未能廉价核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.18010", "title": "Fish Audio Drama 3 Technical Report", "published": "2026-10-13T09:06:28Z", "revision": 2, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "f5d190af92894fa4ed261b98341529ca4a1942e02e00c50434ea64cce5bf5bb0", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "Fish Audio Drama 3 v2是35B-A3B语音生成模型的三阶段pretraining、SFT和human-preference RL报告，推理单H200不代表训练可单卡复制。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "v2 §3 Training", "v2 §3.1 Pre-training", "v2 §3.2 Instruction SFT", "v2 §3.3 RL" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.18010v2", "source_sha256": "2f45cfcc6c9e0eb8806c565e9cd03075a840f456c92301078e8eeb0e31086c71", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "source_attempts": [ { "url": "https://arxiv.org/html/2610.18010v2", "result": "retrieved", "sha256": "2f45cfcc6c9e0eb8806c565e9cd03075a840f456c92301078e8eeb0e31086c71" } ], "cost_assessment": { "source_url": "https://arxiv.org/html/2610.18010v2", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。35B MoE，数百billion pretraining tokens，8K/16K sequence并pack到128K rows，随后SFT/large-batch RL", "scale": "35B MoE，数百billion pretraining tokens，8K/16K sequence并pack到128K rows，随后SFT/large-batch RL", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.17994", "title": "Improving Neural Symbolic Regression with a Diverse Prior and Post-training Alignment", "published": "2026-10-13T08:57:21Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "985817ab32568531c7d0e41985423099da53a14ba756f93eb304aafe9f24868d", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "non_llm_symbolic_regression_out_of_scope", "reason": "RASR 是symbolic regression表达式模型的合成prior与RL对齐，不是语言模型偏好/OPD/agentic后训练；本次仅摘要分流，未评判正文或成本；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "official arXiv abstract: scope triage" ], "fulltext_review_completed": false, "review_level": "official_abstract_scope_triage", "source_url": "https://arxiv.org/abs/2610.17994v1", "source_sha256": null, "evidence_limitations": "决定仅基于官方元数据/摘要范围或明确的来源例外；未核验正文算法或资源，不以metadata hash冒充文档hash。" }, { "id": "2610.17900", "title": "PaceReasoner: Label-Free Learning to Answer Briefly from Long Reasoning", "published": "2026-10-13T07:55:23Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "8640382c206332e9f7f9f756794900f86cf233bb47144e4768aaf9a2bdbf5754", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "PaceReasoner 用LONG consensus监督SHORT，配行为reward与ratio controller；label-free不等于低生成成本，也须保持adaptation/test边界。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§4 PaceReasoner", "Implementation." ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.17900v1", "source_sha256": "552a1e81c473d5b5180a6728c3334d55fe3626431d626855afd30c5f51033f97", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.17900v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。4B/8B三backbones，四bench，15epochs×每题15rollouts，8 H20，最大batch256题", "scale": "4B/8B三backbones，四bench，15epochs×每题15rollouts，8 H20，最大batch256题", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.17767", "title": "Efficiently Learning To Reason or Not to Reason: Root-token Policy Optimization for Adaptive Thinking", "published": "2026-10-13T05:53:34Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "d23abe934601094635614380115eb197f2dea1e4dca2bdf06b4b1dbfad5d00ed", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "RPO 只训练是否think的root token，但必须采样两mode真实长轨迹形成group-relative advantage，不能把人工difficulty gate当复现。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 root-token policy optimization", "Training configuration" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.17767v1", "source_sha256": "1299d6c1614f36670f84fcf3c15eef3d1d26c01921a2c1169e62435bc714da26", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.17767v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。默认3 H100，每迭代8queries×16rollouts，context16384，多model-family/scale", "scale": "默认3 H100，每迭代8queries×16rollouts，context16384，多model-family/scale", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.17666", "title": "DeltaTTG: Training Multi-Turn Agents with a Token-to-Go Progress Meter", "published": "2026-10-13T03:49:32Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "452cfa65515251906e317915945e30fd4861f5e00487c50adf6dd929c3857608", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "DeltaTTG 以真实多策略成功轨迹训练TTG progress meter，再冻结meter重分配actor outcome advantage；meter不能以已知剩余金标计划长度替代。原文资源配置超出本轮单卡低预算范围；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 DeltaTTG", "§4.4 Computational Cost" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.17666v1", "source_sha256": "ed39bc28eba6f9f6bce4fefd996ecd92b87149cc0e11a4c0a559c48c791f2894", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.17666v1", "basis": "依据官方原文披露的配置或作者运行统计作预算判定；不是本地测量，不把相对加速当作绝对低成本。Qwen3-4B/Llama3.1-8B三bench；ALFWorld meter单独RL已10.45 GPU-hours，150step actor成本/评测另计", "scale": "Qwen3-4B/Llama3.1-8B三bench；ALFWorld meter单独RL已10.45 GPU-hours，150step actor成本/评测另计", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "reported" } }, { "id": "2610.17654", "title": "Decouple to Forget: Semantic Mediation for Sensitive Unlearning in Large Language Models", "published": "2026-10-13T03:32:37Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "5a7e5e2d7db2b4e1a959250b1b215906d63dcb917643e054c00ad5197de51600", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "reviewed_deferred", "priority": "P1", "resolution": "not_reproduced_high_cost", "reason": "D2F 先拆解敏感fact mediator、生成保持非敏感内容的反事实，再做preference-based unlearning；单A10040GB和2epochs路径可容纳，但不能忽略GPT-4o构造/审计和双benchmark全链路。核对官方方法/资源片段后，完整忠实复现的总成本仍有实质不确定性，按单卡预算门槛收口；成本过高，不复现（审查 DONE），不进入自动重试或 evolve。", "locations": [ "§3 D2F", "§4.1 Experimental Settings", "Appendix B.2 Training Settings and Hyperparameters" ], "fulltext_review_completed": true, "review_level": "primary_method_resource_gate_review_bounded_passages", "source_url": "https://arxiv.org/html/2610.17654v1", "source_sha256": "7921d7a01894655cd40b1f621782ea65d82b39c6fb878ef958b25722e43c6fc7", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。", "source_attempts": [ { "url": "https://arxiv.org/html/2610.17654v1", "result": "retrieved", "sha256": "7921d7a01894655cd40b1f621782ea65d82b39c6fb878ef958b25722e43c6fc7" } ], "cost_assessment": { "source_url": "https://arxiv.org/html/2610.17654v1", "basis": "官方原文已披露下列配置，但未能通过便宜只读核对核定完整训练、教师/环境调用和对照总时；不以试跑探测成本。Llama2-7B LoRA、TOFU-forget10及KnowUnDo Privacy，D2F各2epochs，单A100 40GB；GPT-4o多步骤生成/审计和完整对照总时未核定", "scale": "Llama2-7B LoRA、TOFU-forget10及KnowUnDo Privacy，D2F各2epochs，单A100 40GB；GPT-4o多步骤生成/审计和完整对照总时未核定", "available_resource": "单卡 A100/A30 或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2610.17616", "title": "StateHarness: Maintaining Evolving Project State for Long-Horizon Coding Agents", "published": "2026-10-13T02:39:03Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "417c302cb662362d1f109b1e43ddbce327b9d16397c4138d97cd6f38fee43c48", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "rejected_or_out_of_scope", "priority": "P2", "resolution": "cross_track_runtime_state_reference", "reason": "StateHarness 明确不更新模型参数，以project state、诊断会话和runtime支持长期编码；仅作为Agent harness参考，不能把任务完成率增加写成后训练模型能力收益，且原文有pass-to-pass回归增加；本轮仅保留参考，不登记为已实现能力，也不自动重试或接入 evolve。", "locations": [ "§3 StateHarness", "§4.4 Execution Cost and Performance Trade-offs", "Appendix B.3 Models and Implementation Versions" ], "fulltext_review_completed": true, "review_level": "primary_scope_or_evaluation_passages_review", "source_url": "https://arxiv.org/html/2610.17616v1", "source_sha256": "e5b367faaea3b5db76988b4e4df767c12a54f1a4ed875a54339a20114353b040", "evidence_limitations": "仅核阅列出的官方正文方法、评估或资源片段；未阅读全文，不是完整算法/全部结果复核，不是实现、训练或GPU验证。" }, { "id": "2610.17471", "title": "Scaling Inference-Time Compute with Teams of Coding Agents", "published": "2026-10-12T22:35:08Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b0dd11ba530f42812abd4bd4f380c52ee20f6d09f6966aa9d6c5ffde42d01f73", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Scaling Inference-Time Compute with Teams of Coding Agents》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.17471v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.17444", "title": "The belief-action gap in language model agents", "published": "2026-10-12T22:04:36Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "80a9be9ae288a3eb51a022fe1511a10ee1b96b9356d9e3a219ccdaa399bd95ec", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《The belief-action gap in language model agents》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.17444v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.17409", "title": "Position: Beyond Training Time, Test-Time Coordination is Essential for Cooperative MARL", "published": "2026-10-12T21:17:48Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "f21e32f646b930469e19c434e1e4f33e057f267906f58bd00c43835f89a54702", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Position: Beyond Training Time, Test-Time Coordination is Essential for Cooperative MARL》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.17409v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.17370", "title": "HarnessTTT: Learn to Optimize Harness at Test-Time for Scientific Discovery", "published": "2026-10-12T20:27:32Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "e8e37e725ea1d0267bdf5d8a41284737f194f2153d66e5d1a91e3280d36da0e7", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《HarnessTTT: Learn to Optimize Harness at Test-Time for Scientific Discovery》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.17370v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.17309", "title": "StartLux-Decision: AutoThink Decision Models with Reinforcement Learning for Calibrated Deliberative Decisions", "published": "2026-10-12T19:05:56Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "fc06eef76558b0226e4c4400878917b02960062fb142ba25a0fcde6589b43532", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《StartLux-Decision: AutoThink Decision Models with Reinforcement Learning for Calibrated Deliberative Decisions》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.17309v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.17113", "title": "Calibrated Typed Decisions for Entity, Relation, and Event Extraction", "published": "2026-10-12T17:05:20Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "7050b259ffc2e9cf663c963b832a7846ee2c3a9b9917b7cadf4d785cbef08f46", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Calibrated Typed Decisions for Entity, Relation, and Event Extraction》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.17113v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16867", "title": "Davinci-Tool: Scaling Agentic Tool-Use Supervision with Verifiable Executable Environments", "published": "2026-10-12T14:33:16Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "eeb02509fcc332494c83e5005f7c7cab43bd03252b3da00a7da7da46b23239e0", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Davinci-Tool: Scaling Agentic Tool-Use Supervision with Verifiable Executable Environments》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16867v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16794", "title": "TPPO: Tracking-Value PPO For Language Model Reasoning", "published": "2026-10-12T13:47:36Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "9240f9d8d1fd9f431828394eb45253b4c2ed0a9bd24ccdc549cde94ec36aa1ea", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《TPPO: Tracking-Value PPO For Language Model Reasoning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16794v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16354", "title": "Self-Distillation with Verbal Regularization", "published": "2026-10-12T08:11:35Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "cf90b714a8a67f74b1a78505033d9dc9cf5d7a03360b28403fb8ce991f4a46dc", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《Self-Distillation with Verbal Regularization》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16354v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.16104", "title": "SelectTTRL: Representation-Guided Data Selection for Efficient Test-Time Reinforcement Learning", "published": "2026-10-12T03:12:53Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "508f81875599cdb0e4147a4e6a1823c5d0f8138c4fc5b9a89544a6113261b578", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-12，早于2026-10-13窗口起点；《SelectTTRL: Representation-Guided Data Selection for Efficient Test-Time Reinforcement Learning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.16104v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15888", "title": "Variance-Optimal State Resampling for Reinforcement Learning with Verifiable Rewards", "published": "2026-10-11T20:09:24Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "36d77a336bace1243eb6c0aa21f1167b42b9688ec73afb6f2869c70bc11d876f", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《Variance-Optimal State Resampling for Reinforcement Learning with Verifiable Rewards》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15888v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15368", "title": "ANTE: Anticipate, Then Execute -- Learning Internal World Models for Coding Agents from Verified Beliefs", "published": "2026-10-11T05:46:07Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "fb221894dc1e1423288528eb74c027b91c3db0a2f8b8b83f8db8450bbc4c1c50", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《ANTE: Anticipate, Then Execute -- Learning Internal World Models for Coding Agents from Verified Beliefs》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15368v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15258", "title": "Sequent-Prover: Training Agents for Formal, Checkable SMT-based Reasoning", "published": "2026-10-11T01:10:19Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "0bc72ed0e4ba571b91fb0887a8c2add9de7a62ee5fc46bdadf6d42e1359ae955", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《Sequent-Prover: Training Agents for Formal, Checkable SMT-based Reasoning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15258v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15244", "title": "EvolvAttack: Backdooring self Evolving VLM Agents", "published": "2026-10-11T00:33:53Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "6bcca235511c3800d7fbe18f23458ca7fb0a98d90bb18bf034c73fe81364700d", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-11，早于2026-10-13窗口起点；《EvolvAttack: Backdooring self Evolving VLM Agents》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15244v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15173", "title": "Physics Foundation Models for Efficient Control of Fluid Flows", "published": "2026-10-10T20:51:19Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b420a0cef7932bf3eb36c99ccec941f81350571745b94aac33642040619be145", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《Physics Foundation Models for Efficient Control of Fluid Flows》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15173v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.15033", "title": "What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO)", "published": "2026-10-10T16:18:41Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "f0fd60ef2af2a9aaa5d88d48a2ed102b33b43193d2104b96cd22c0318cfeb567", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO)》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.15033v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14705", "title": "Lingdan-V2: Large language models for clinically grounded reasoning in traditional Chinese medicine", "published": "2026-10-10T08:22:12Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "9ee1274e577cb3dd98335c590c5d705c01becc8562b4ed5ec4c4f6e30bb33644", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《Lingdan-V2: Large language models for clinically grounded reasoning in traditional Chinese medicine》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14705v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14572", "title": "Reward-Aligned Distillation", "published": "2026-10-10T04:43:04Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "8f32d3d739ce86e65df9c552011fbcce14c8ef5c0277517ebb84145852fea4c3", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《Reward-Aligned Distillation》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14572v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14479", "title": "AphaVoice: Text-to-Speech Model for Aphasia Patient Simulation", "published": "2026-10-10T02:22:51Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "8cecf9aa057e083bb09c74e83d5695e572a5e44c1e6454cd68b4106dcaf01537", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《AphaVoice: Text-to-Speech Model for Aphasia Patient Simulation》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14479v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14448", "title": "KV4-RL: Unlocking Faster Asynchronous Rollouts for RLVR with INT4 KV Cache Quantization", "published": "2026-10-10T01:24:03Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "95fe6bcb88a312ff20547af1fc728549680ccdde12a264a6a025961d6d34eb61", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《KV4-RL: Unlocking Faster Asynchronous Rollouts for RLVR with INT4 KV Cache Quantization》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14448v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14440", "title": "Target-ESS Tempered Policy Optimization (TETPO): Smooth Variance Control via Bridge Importance Sampling", "published": "2026-10-10T01:15:04Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "5a8285654722c9a03b17872d341e3c0e2f1585b4a78648fe84c097b5eddc3dac", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-10，早于2026-10-13窗口起点；《Target-ESS Tempered Policy Optimization (TETPO): Smooth Variance Control via Bridge Importance Sampling》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14440v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.13923", "title": "Goal-Oriented Bundle Management and Flow Control for Deep-Space Communications via Reinforcement Learning", "published": "2026-10-09T17:53:46Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "23a0a6e2d8d9f0ff2eda315ab929dabf130a15171597b763eaabbe0347474747", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-09，早于2026-10-13窗口起点；《Goal-Oriented Bundle Management and Flow Control for Deep-Space Communications via Reinforcement Learning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.13923v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14163", "title": "AutoEnvScaling: Automating the Data Flywheel with Terminal Agents", "published": "2026-10-09T17:51:50Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "ab4561db9676712e2789242e69d7f2e80730ad532f1c9dbd442623d0bebd4ecc", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-09，早于2026-10-13窗口起点；《AutoEnvScaling: Automating the Data Flywheel with Terminal Agents》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14163v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.13353", "title": "EmoGavel: Grounding-Aware Reinforcement Learning with Verifiable Rewards for Multimodal Emotion Preference Learning", "published": "2026-10-09T10:28:54Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "cd6641640ccfc6e886a994f05dd892b8bf022d3c9cf6c3adb05b574ffa1075bc", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-09，早于2026-10-13窗口起点；《EmoGavel: Grounding-Aware Reinforcement Learning with Verifiable Rewards for Multimodal Emotion Preference Learning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.13353v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.13306", "title": "Structure-Aware Numerical Stabilization for Half-Precision Proximal Policy Optimization", "published": "2026-10-09T09:52:15Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "9b16db76078e09f6ff38ae8ee20469a8afc4a54657b8560e18d9f66127f564f1", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-09，早于2026-10-13窗口起点；《Structure-Aware Numerical Stabilization for Half-Precision Proximal Policy Optimization》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.13306v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.13205", "title": "Rethinking Relative Policy Differences Transfer of On-Policy Distillation", "published": "2026-10-09T08:24:48Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "4825ac580a3a30d40bb79ad00ce0f78b7ce3db0373e60b5f10390377444fa0ab", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-09，早于2026-10-13窗口起点；《Rethinking Relative Policy Differences Transfer of On-Policy Distillation》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.13205v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.12572", "title": "DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models", "published": "2026-10-08T18:01:44Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "e5310c442dbdc5eda161f59f0eab6de819ac49ceedd7f85f386684f648084123", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-08，早于2026-10-13窗口起点；《DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.12572v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.12403", "title": "ViSkill: Reinforcing VLM Agents with Evolving Visual-Native Skills", "published": "2026-10-08T17:44:19Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "50ae2b0133f9219582e8ec823135d60a09e5c17b36859f58c4f46a9ea832108e", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-08，早于2026-10-13窗口起点；《ViSkill: Reinforcing VLM Agents with Evolving Visual-Native Skills》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.12403v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.11423", "title": "Policy Alignment: New Signals for Membership Auditing in On-Policy Distillation", "published": "2026-10-08T07:48:55Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "9ca9214a68a2687707abc355e9a16ac23606de1d866bad5cb8f53602381772c3", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-08，早于2026-10-13窗口起点；《Policy Alignment: New Signals for Membership Auditing in On-Policy Distillation》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.11423v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.10447", "title": "A Good Self-Teacher Meets the Student Where They Are: Joint On-Policy Learning and Teaching", "published": "2026-10-07T17:19:19Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "a5f44f049359e35156c56e874137dbb3d3524c51738c5a3ef76bc17e21d33856", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-07，早于2026-10-13窗口起点；《A Good Self-Teacher Meets the Student Where They Are: Joint On-Policy Learning and Teaching》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.10447v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.10273", "title": "A Closed-Loop Non-Asymptotic Convergence Analysis of PPO with Learned Critics and Clipping", "published": "2026-10-07T15:40:47Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "58cce7ac462799f04b685c6bca95ef4bf70aa970ba573c7d0605f80420d93cee", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-07，早于2026-10-13窗口起点；《A Closed-Loop Non-Asymptotic Convergence Analysis of PPO with Learned Critics and Clipping》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.10273v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.09108", "title": "Convex-Concave Reinforcement Learning", "published": "2026-10-06T20:59:21Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "7e8c2d8a49e58ab3e53a4b010c31fe245ff20975c8f4e714e9e651492427859c", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-06，早于2026-10-13窗口起点；《Convex-Concave Reinforcement Learning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.09108v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.08448", "title": "Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability", "published": "2026-10-06T14:37:22Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "36807fd3783cd6b4a08173318342d778759db196d0c21d5beb1f911b188660a5", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-06，早于2026-10-13窗口起点；《Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.08448v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.08398", "title": "UP-MOPD: Update Projection in Multi-Teacher On-Policy Distillation", "published": "2026-10-06T14:13:31Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "946f85c5fb0f33a874b969da72d212def7cb3f8b046fe52b96e7d6aea6271b0d", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-06，早于2026-10-13窗口起点；《UP-MOPD: Update Projection in Multi-Teacher On-Policy Distillation》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.08398v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.08112", "title": "Energy-Aware Path Following: Comparative Analysis of Reinforcement Learning and NMPC for Electric Vehicles", "published": "2026-10-06T10:35:13Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "1e85b324bfcafbb1c38a5ae8baf1f240b8e664293b6ca8c2a960dd843b5ce39a", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-06，早于2026-10-13窗口起点；《Energy-Aware Path Following: Comparative Analysis of Reinforcement Learning and NMPC for Electric Vehicles》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.08112v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.07874", "title": "On-Policy Distillation with Negative-Policy Rollouts", "published": "2026-10-06T07:25:03Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "b235470a5b7c571982990958cdc8f158be6f2c0b5b44fc47b407945a5755497a", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-06，早于2026-10-13窗口起点；《On-Policy Distillation with Negative-Policy Rollouts》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.07874v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.07405", "title": "What pass@k Cannot Measure: Evaluating Diversity and Capability Retention after Post-Training", "published": "2026-10-05T21:16:32Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "a74cc22bdfdf3998c608db74b8d3b13a243181e9138ead6ade46fd69a26e87b4", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-05，早于2026-10-13窗口起点；《What pass@k Cannot Measure: Evaluating Diversity and Capability Retention after Post-Training》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.07405v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.06105", "title": "Flash-OPD: Fast On-Policy Distillation", "published": "2026-10-05T10:36:57Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "1f9924ffe8615dcb6fde63d2e3accab5e53dab072f4c4e80e5b49239d3d85fef", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-05，早于2026-10-13窗口起点；《Flash-OPD: Fast On-Policy Distillation》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.06105v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.05974", "title": "Transfer-Stratified On-Policy Distillation for RL-Improved Reasoning Teachers", "published": "2026-10-05T08:26:20Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "2e4fefb482e6c0e9319996128afb535fd8efc0edb4c4771d47b3e137e4bdc76b", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-05，早于2026-10-13窗口起点；《Transfer-Stratified On-Policy Distillation for RL-Improved Reasoning Teachers》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.05974v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.05954", "title": "MEND: RL For Flow Models via Proximal Velocity Matching", "published": "2026-10-05T08:07:18Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "a1b3a0d4bb5e9b37e9882293806ea83c47c6110100a108c0f1fdd790569845ff", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-05，早于2026-10-13窗口起点；《MEND: RL For Flow Models via Proximal Velocity Matching》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.05954v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.05745", "title": "Beyond In-Distribution Preservation: Recovering Generalization in Quantized VLAs via Vulnerability-Oriented Tuning", "published": "2026-10-05T03:51:21Z", "revision": 2, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "50ff2334684bfb3ba6a3bcbd8c802f238d9d08a571fa873f139b79e2116f0d60", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-05，早于2026-10-13窗口起点；《Beyond In-Distribution Preservation: Recovering Generalization in Quantized VLAs via Vulnerability-Oriented Tuning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.05745v2", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.05070", "title": "Outcome-Guided On-Policy Self-Distillation", "published": "2026-10-04T09:13:27Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "4149fac094d3fc6389e917597b45737ddcdde5f4be3f937d9df92c154c182755", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-04，早于2026-10-13窗口起点；《Outcome-Guided On-Policy Self-Distillation》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.05070v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.04754", "title": "FoSeRL: Formal Sequential Robustness Certification for Reinforcement Learning Policies", "published": "2026-10-03T20:43:47Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "758923081bf7d082e6beff4023c75199554a881f3eb7f97eed78c64fedb3bb20", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-03，早于2026-10-13窗口起点；《FoSeRL: Formal Sequential Robustness Certification for Reinforcement Learning Policies》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.04754v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.04638", "title": "Policy as Data: Replay-Based Policy Dual Averaging via Advantage Regression", "published": "2026-10-03T16:29:39Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "4a66db34eea6c134409df8a4460af252ba26e3dd0497fb05a4b2ca8ad4b2f37a", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-03，早于2026-10-13窗口起点；《Policy as Data: Replay-Based Policy Dual Averaging via Advantage Regression》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.04638v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.03880", "title": "Reinforcement Learning on the Discrete Composition Channel of a Crystal Generator: Validated Gains and Reward Hacking", "published": "2026-10-02T18:02:02Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "86c8e835f2b37ff5218e719943dc19b99cfc30f167e680f18900c2d8d23bfe4a", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-02，早于2026-10-13窗口起点；《Reinforcement Learning on the Discrete Composition Channel of a Crystal Generator: Validated Gains and Reward Hacking》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.03880v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.02911", "title": "Probe the Harness: Setup Checks for Stale-Data RL Comparisons in Language Models", "published": "2026-10-02T07:02:13Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "cf4f2d2f038b05ac0294a955f3e5262970a7b9240a81ca69aa0edcb3b9e9d5b9", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-02，早于2026-10-13窗口起点；《Probe the Harness: Setup Checks for Stale-Data RL Comparisons in Language Models》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.02911v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.02848", "title": "Permutation Robustness Is Not Enough: Action Collapse in Multi-Agent Transformer Policies", "published": "2026-10-02T05:35:53Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "5e549d58597e298e7f396f89374adc9eea690ab5f3aa25056cec8df72c9b0897", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-02，早于2026-10-13窗口起点；《Permutation Robustness Is Not Enough: Action Collapse in Multi-Agent Transformer Policies》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.02848v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.02846", "title": "Understanding Enrichment in Reinforcement Learning", "published": "2026-10-02T05:34:41Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "897bf385e55db09bfff788d0f1f44ddbed1ddb780752e9d389814bd03016a669", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-02，早于2026-10-13窗口起点；《Understanding Enrichment in Reinforcement Learning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.02846v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.02808", "title": "ROUTEAUDIT: Interaction-Aware Identification for Budgeted Multi-Verifier Routing", "published": "2026-10-02T04:59:24Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "41e08ab5a79d9abd7cc27047b045574f6f9e6a465659d63931a69e85cb771799", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-02，早于2026-10-13窗口起点；《ROUTEAUDIT: Interaction-Aware Identification for Budgeted Multi-Verifier Routing》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.02808v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.02505", "title": "Multi-Fidelity Policy Gradients Stabilize Data-Scarce Reinforcement Learning", "published": "2026-10-01T21:27:02Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "05aef7e42ac7f99b6785a7014c676767847eb9b23f76e790716916cb962b2a7c", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-01，早于2026-10-13窗口起点；《Multi-Fidelity Policy Gradients Stabilize Data-Scarce Reinforcement Learning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.02505v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.02098", "title": "Are We Recovering Mechanisms? Objective-Level Recovery Gaps in Mechanistic Interpretability", "published": "2026-10-01T17:25:23Z", "revision": 2, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "107dd25479d6f7c0db662ef6d7b1b67cc82180ae7026f9e225e55685ad42173b", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-01，早于2026-10-13窗口起点；《Are We Recovering Mechanisms? Objective-Level Recovery Gaps in Mechanistic Interpretability》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.02098v2", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.01729", "title": "Function-Structured Reinforcement Learning with Executable Verifiers for Mathematical Reasoning", "published": "2026-10-01T14:02:24Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "ac497e112a8a98a5c8ea0cb6b7d33d59c00cd36b196233365d829f54acee5bb8", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-01，早于2026-10-13窗口起点；《Function-Structured Reinforcement Learning with Executable Verifiers for Mathematical Reasoning》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.01729v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.01399", "title": "Reusing Past Samples in Proximal Policy Optimization: When and How Does It Help?", "published": "2026-10-01T10:05:35Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "4aaff7be72763a1cf4428b921660a36739918b5b60815af0a82338b2b35183e6", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-10-01，早于2026-10-13窗口起点；《Reusing Past Samples in Proximal Policy Optimization: When and How Does It Help?》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.01399v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14030", "title": "Rotate the Loop: Blockwise Conformality for Looped Transformers", "published": "2026-09-30T23:32:34Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "877c4c8df9099d05e05f32f2e654b023a5d3d0e3333c166e6ea1c2b44f8caee9", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-09-30，早于2026-10-13窗口起点；《Rotate the Loop: Blockwise Conformality for Looped Transformers》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14030v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14022", "title": "OPMD: On-Policy Mechanism Distillation Beyond Output Supervision", "published": "2026-09-29T14:07:18Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "51e0c2d9f0649476168ca0656d1158564684a7f36870552b36175abbb5c94704", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-09-29，早于2026-10-13窗口起点；《OPMD: On-Policy Mechanism Distillation Beyond Output Supervision》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14022v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.14001", "title": "PreSO: Privileged Representation Isolation for On-Policy Self-Distillation", "published": "2026-09-27T08:48:42Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "55e794f1414688f8b800865fe9c619879aaaf499b529437b16f4820636d785c4", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-09-27，早于2026-10-13窗口起点；《PreSO: Privileged Representation Isolation for On-Policy Self-Distillation》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.14001v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.13998", "title": "Enhanced Trading Systems with Intelligent Market Analysis", "published": "2026-09-27T03:53:05Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "6b8e188dd0060442bd7c8df33e3bda69a5860f884589fc85016157d6a732ecdd", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-09-27，早于2026-10-13窗口起点；《Enhanced Trading Systems with Intelligent Market Analysis》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.13998v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.08819", "title": "HydroSphere: A Framework for Governed, Self-Healing Wastewater Infrastructure", "published": "2026-09-24T17:18:48Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "285cbe5bad72aa257265c8accc8a56473d74c771278df24a9b3ee2fa8261e511", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-09-24，早于2026-10-13窗口起点；《HydroSphere: A Framework for Governed, Self-Healing Wastewater Infrastructure》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.08819v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" }, { "id": "2610.00202", "title": "When a Data Artifact Isn't a Shortcut: Causal Auditing of Synthetic RLVR Corpora", "published": "2026-09-20T11:31:26Z", "revision": 1, "abstract_review_completed": true, "full_article_read": false, "metadata_sha256": "eeac2d9d67b68fd8223d7a7a9d0ab8f0e088bbcd1cbdc40123d939ac2499e917", "scope_triage_completed": true, "review_done": true, "automatic_retry": false, "implementation_status": "not_implemented_by_review", "decision": "historical", "priority": "P2", "resolution": "historical_pre_window_identity_date_scope_triage", "reason": "官方元数据首发为2026-09-20，早于2026-10-13窗口起点；《When a Data Artifact Isn't a Shortcut: Causal Auditing of Synthetic RLVR Corpora》本次仅作身份、首发日期和摘要范围分流。即使公告或修订落在本窗口也不算新首发；未审正文/成本，不计入本窗口新论文全文覆盖，明确历史归档且不自动重试。", "review_level": "official_metadata_and_abstract_date_scope_triage_only", "abstract_review_scope": "identity_publication_date_and_scope_triage_only", "fulltext_review_completed": false, "source_url": "https://arxiv.org/abs/2610.00202v1", "source_sha256": null, "locations": [ "official arXiv metadata: identity, original published date and revision", "abstract: date/scope triage only" ], "evidence_limitations": "历史补发现分流，不是2026-10-13..15新论文正文审查；metadata_sha256为元数据hash，不是HTML/PDF文档hash。" } ] }