{
  "schema_version": 1,
  "reviewed_on": "2026-10-15",
  "track": "agent",
  "review_window": {
    "start": "2026-10-13",
    "end": "2026-10-15"
  },
  "review_boundary": "353条摘要/身份范围逐项收口；当前窗口按实际官方正文片段核定方法范围或资源终态；历史条目不冒充新论文全文审查；无全文通读、实现、训练、付费调用或GPU运行。",
  "reviews": [
    {
      "id": "2610.20582",
      "title": "Latent Compaction with Test-Time Training",
      "published": "2026-10-14T17:57:01Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "8135f87ab79b32bb2a96064a3f4cbdfb8aa4481cb2e301b71b42f7dc4e4bf649",
      "priority": "P1",
      "reason": "Latent Compaction 冻结模型但对每份文档做attention fitting和端到端cache蒸馏，并非零训练压缩。原文建议两个阶段各1000更新，RULER每task-length有20文档；完整准备/多模型评测成本明显，单卡预算内无法廉价确认总耗时。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3 Method: Latent Compaction",
        "4.2 Evaluation benchmarks and metrics",
        "Training Duration"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20582v1",
      "source_sha256": "4f214311c15a79b2a2e4f0cbfe14b146d89a1729d700da4e34b147493a709f0a",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20582v1",
        "basis": "资源与训练时长段给出每文档优化量；没有把相对压缩收益当作绝对便宜。",
        "scale": "逐文档1000+1000更新；Qwen3-0.6B/1.7B/4B等",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.20486",
      "title": "Average-Reward Policy Identification from a Single Continuing Trajectory",
      "published": "2026-10-14T16:57:47Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "9ea0c1df4e2eb6b9f92b3504d89c115c83273d5d7f7ae1dbbaefa0170f80bb70",
      "priority": "P2",
      "reason": "研究average-reward tabular RL的单continuing trajectory样本复杂度与信息论下界，不是LLM工具Agent/外部harness RSI；按摘要明确范围分流，本轮未核全文或估造GPU成本。",
      "locations": [
        "arXiv metadata and abstract"
      ],
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.20486v1",
      "source_sha256": null,
      "source_hash_scope": "not_captured_metadata_hash_separate",
      "review_level": "abstract_scope_triage",
      "evidence_limitations": [
        "本条正文审查未完成；终态仅表示范围/证据例外归档，不表示实现或实验验证。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.20392",
      "title": "FOLD: Position-Free Far Memory with Local Order for Long-Context Speculative Decoding",
      "published": "2026-10-14T15:55:10Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "d170d335b87f5555383cfb68d84bf245fed0d174ce31a07bef91cbef48ed86bc",
      "priority": "P2",
      "reason": "FOLD 的position-free far memory服务长上下文speculative decoding，是drafter attention/推理系统优化，不是Agent可持续记忆管理或自改进。归foundation-model范围，不因Agent memory关键词纳入本track实现。",
      "locations": [
        "abstract",
        "B.4 Implementation",
        "Limitations"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20392v1",
      "source_sha256": "4ab36a093349d953cf49ee4c3fc611cb5203e96c749062c6096e55ceda055243",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.20277",
      "title": "TANGO: RNA Topology and Geometry Co-Design",
      "published": "2026-10-14T14:37:51Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "af137693eff814a6b83da67614ec6cf907169e24b4b8681d48b28bae5a5ad01e",
      "priority": "P2",
      "reason": "TANGO 的multi-agent RL用于RNA序列拓扑/几何逆设计，是生物分子优化，不是LLM agent/harness/RSI；按实际摘要范围分流，未声称全文或成本审查。",
      "locations": [
        "arXiv metadata and abstract"
      ],
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.20277v1",
      "source_sha256": null,
      "source_hash_scope": "not_captured_metadata_hash_separate",
      "review_level": "abstract_scope_triage",
      "evidence_limitations": [
        "本条正文审查未完成；终态仅表示范围/证据例外归档，不表示实现或实验验证。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.20252",
      "title": "Adversarial Reinforcement Learning for Robust Planetary Landing",
      "published": "2026-10-14T14:23:47Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "848daf465e02097b6852c18c89e3a0afcb40add1486ba03367bf01faf743c4f2",
      "priority": "P2",
      "reason": "该工作是三自由度行星着陆控制的adversarial DRL/disturbance generator，不是语言工具agent或harness自改进；按摘要范围排除，不以未取得正文假装高成本。",
      "locations": [
        "arXiv metadata and abstract"
      ],
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.20252v1",
      "source_sha256": null,
      "source_hash_scope": "not_captured_metadata_hash_separate",
      "review_level": "abstract_scope_triage",
      "evidence_limitations": [
        "本条正文审查未完成；终态仅表示范围/证据例外归档，不表示实现或实验验证。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.20009",
      "title": "Outperforming 1000 Euclidean Layers With Hyperbolic Goal-Conditioned RL",
      "published": "2026-10-14T11:23:28Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "10a0d58a5a2e8d3010e8f74383200a0756f3eb2e9d08a4495d77a43e6e7f4aa2",
      "priority": "P2",
      "reason": "Hyperbolic GCRL 研究maze目标到达的神经表示几何与contrastive RL，非LLM agent/RSI；只做真实摘要范围归档，不把12倍相对速度解释为本地低成本。",
      "locations": [
        "arXiv metadata and abstract"
      ],
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.20009v1",
      "source_sha256": null,
      "source_hash_scope": "not_captured_metadata_hash_separate",
      "review_level": "abstract_scope_triage",
      "evidence_limitations": [
        "本条正文审查未完成；终态仅表示范围/证据例外归档，不表示实现或实验验证。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.19951",
      "title": "AtomSearch: Evidence-Grounded Tree Search for Agent Skill Optimization",
      "published": "2026-10-14T10:33:30Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "8940b6e97b8ca4ae13747eb2c3397adedea3a61594865665809ab1ebb502f671",
      "priority": "P1",
      "reason": "AtomSearch 的局部规则编辑、skill-state树搜索和回归/修复记忆是实际核心。正文附录B.2报告17个搜索run、单run923–17200 item evaluations与16–45 optimizer calls；完整多目标搜索成本高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3 Method",
        "B.2 Evaluation cache and optimization cost"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19951v1",
      "source_sha256": "155ed8f9af500a4811c4930a2c849aa7a6be647e39c11188d83ac16c7d346d76",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19951v1",
        "basis": "论文逐run预算表，缓存命中不抵消已经报告的新评测开销。",
        "scale": "17 runs；923–17200 item evaluations/run；多模型",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.19868",
      "title": "Explorer-Zero: Improving Self-Play for Long-Horizon Tasks with Decomposable Verifiable Rewards",
      "published": "2026-10-14T09:29:31Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "4513fc8747cf2d301a71f101711c33ed386215413224ab6b053b24e12422a19d",
      "priority": "P1",
      "reason": "Explorer-Zero 需要proposer/solver交替自博弈并保存环境checkpoint作训练奖励。附录D是8×H100、每ALFWorld运行约15 node-hours、AppWorld约29；不能以小fixture或只实现reward函数当作自博弈复现。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3 Methodology",
        "4.1 Experimental Setup",
        "Appendix D Implementation Details"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19868v1",
      "source_sha256": "7cfeab24dc6d9fbfdb52c2fc901230d7f418e813825c5a38f77acad7dcd89e63",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19868v1",
        "basis": "正文训练配置及报告wall-clock，而非本地实测。",
        "scale": "4B/30B-A3B；8×H100；约15/29 node-hours/run",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.19810",
      "title": "ReproBAIT: Can coding agents build their own biological AI tools?",
      "published": "2026-10-14T08:50:11Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "bd5e9d6faf0e60171ed0026d0fa76b7ea1feae7b2937077a2f13a15f151c4c4d",
      "priority": "P2",
      "reason": "ReproBAIT 是在隔离scorer环境中评价coding agent重建生物工具的基准；held-out数据不可进入agent。保留沙箱/记忆检查协议参考，不把基准或只重算已存输出登记为能力复现。",
      "locations": [
        "C.1 Evaluation security and infrastructure",
        "3 Results"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19810v1",
      "source_sha256": "d3c08a55ec1f5a10455278d5612e012b30411d3634bb38bc17df361f8d4dfda8",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.19795",
      "title": "StateCraft: Augmenting Video World Models with Abstract-State Reasoning",
      "published": "2026-10-14T08:39:41Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f88781a84f263919b4a577e737605bd105b53c8acfac12b097ccbaaf68a9c861",
      "priority": "P1",
      "reason": "StateCraft 将持久抽象状态与视频diffusion共同训练；原文两个backend各约400 H200 GPU-hours、约2000条×300动作轨迹，完整链路超预算。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "4 Experiments",
        "Training schedule."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19795v1",
      "source_sha256": "29ccdcd90b64c8e72e00ac87927927a0e7f6d64fdc2b549e45ee1738ffd3afc6",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19795v1",
        "basis": "原文训练表，不因可编辑state或小参数backbone忽略训练量。",
        "scale": "DFoT/CF++；每backend约400 H200 GPU-hours",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.19712",
      "title": "New Model, Old Habits: Behavior Transfer under Model Swapping in Multi-Session LLM Agents",
      "published": "2026-10-14T07:38:54Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "d51eb60a471a46a430a9176bc6da6ad83727b7e84b60e8ce57ff7785345e91a0",
      "priority": "P1",
      "reason": "DASH 按交换方向和历史作者选择原始历史/摘要，需在多session真实任务上反复换模。已查重复交换设置和局限；闭环任务、原始推理可见性及多模型组合的总调用成本仍无法在单卡/Mac预算下廉价核定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "4 Evaluation under Repeated Swapping",
        "4.1 Experimental Settings",
        "Limitations"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19712v1",
      "source_sha256": "b812043f3e9de523fd1f2dd8211156482a663280dbd6596a45d49cf6316d2812",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19712v1",
        "basis": "只读正文设置可确认需长状态任务和多个模型，未把摘要省tokens当完整复现成本已确定。",
        "scale": "Web Shopping/Travel Planning；重复换模、多历史策略",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19679",
      "title": "MCAIR: Monitoring and Correcting Agent Unfaithfulness via Internal Representations",
      "published": "2026-10-14T07:00:02Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "cda92f7be3a3e80bd9d6180d033c4ffce047ef55d2a2b361051e986eb87507de",
      "priority": "P1",
      "reason": "MCAIR 的CPU probe单次低于0.2ms仅是在hidden state已经提取之后，不含模型生成和纠正。原文8B/12B/14B、两张RTX A6000、分层扫描和三次纠正实验；完整采样/表征提取成本仍高或不确定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "Implementation Details.",
        "Probe Monitoring Cost",
        "Correction Implementation Details"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19679v1",
      "source_sha256": "2c7b4a839bea8a8457a8785e0b22a5e15f13bd7e5374bbfb723387a53fde0ac2",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19679v1",
        "basis": "原文明示孤立probe成本排除generation、GPU转移及correction，不能据其低延迟接受全链路。",
        "scale": "三种8B–14B模型；2×RTX A6000；多seed/纠正循环",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19591",
      "title": "L3M: A Verifiable Harness and Runtime for Self-Extending and Self-Improving Agents",
      "published": "2026-10-14T05:03:06Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6c4f6e6439ee7dca0b7c526975b54dc09c813983393c8f1f4649cb1c356ed675",
      "priority": "P1",
      "reason": "L3M 的Lean权限类型和Git轨迹机制有工程价值，但原文能力比较采用37个Terminal-Bench任务、Opus 5和长预算。仅做日志/假模型测试不能替代真实自扩展与安全工具能力；完整研究运行成本无法廉价核定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "Experiments are bounded agents.",
        "Setup"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19591v1",
      "source_sha256": "315f66eef2c36deb0ea45924fe2b26857f7eaa8c2703dd68210c9ae4bffde6cf",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19591v1",
        "basis": "官方正文Setup给出长任务与frontier调用；未启动Lean安装或任务试跑。",
        "scale": "37个最长约一小时筛选的任务；Opus5、64000输出上限",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19372",
      "title": "Policy-Verified Self-Training: Learning Agent Compliance from Verifiable Policy Rewards",
      "published": "2026-10-14T01:16:15Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b5282838db7a0b62ef9ed83e16fb963f546530e56eabfa96c0cffe536a9c847c",
      "priority": "P1",
      "reason": "PVST 把确定性policy gate变为RFT/DPO监督，但正文也说明verifier受attack-goal设计、部分binding弱于原条款。8B/4B多臂训练及AgentDojo闭环rollout的总成本经已有正文配置仍不确定；不把仅编译gate登记为训练能力。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "5.1 Main result and controls",
        "7 Limitations",
        "Appendix E Common evaluation set and task-clustered confidence intervals"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19372v1",
      "source_sha256": "71a037f6b09a2f26ebcd250f06beeab768d44f725e1b6683dca5c299c9737eec",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19372v1",
        "basis": "已核真实训练与每臂约230–255条评测轨迹，未找到足以界定本地完整耗时的已取证配置；不做昂贵探测。",
        "scale": "Qwen3-8B/4B；RFT/DPO/多控制臂及三seed",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19068",
      "title": "Plan or Retrieve First? A Scale Study of Tool Discovery Architectures",
      "published": "2026-10-13T19:03:21Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6e6bfe7be70c2aac17886bc8378dbafdd148b2613b55c55fe7c147a5b557a656",
      "priority": "P2",
      "reason": "这是RtP/PtR工具发现的规模研究，核心证据是复杂度相关交叉及75–104%额外token开销。归档为路由设计/评测参考，不把尚属方向的complexity-adaptive routing声明成已实现算法。",
      "locations": [
        "3.1 Main Results",
        "4.3 Discovery as the Binding Constraint"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19068v1",
      "source_sha256": "49777dc4714ce52dd16b6e5da457494fe97d620fcd42835ba287e1c29baf3d3a",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.19046",
      "title": "What Information Buys in Active Inference",
      "published": "2026-10-13T18:40:53Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "5734792e0103322d12a345154a5789acc515248bed5969f6fa629d598a437d36",
      "priority": "P2",
      "reason": "这是tabular MDP中active inference/Bayes-UCBVI探索置信与regret理论，不是LLM agent或可执行harness evolution；按摘要范围排除，本轮未作全文算法审查。",
      "locations": [
        "arXiv metadata and abstract"
      ],
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.19046v1",
      "source_sha256": null,
      "source_hash_scope": "not_captured_metadata_hash_separate",
      "review_level": "abstract_scope_triage",
      "evidence_limitations": [
        "本条正文审查未完成；终态仅表示范围/证据例外归档，不表示实现或实验验证。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.18933",
      "title": "The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery",
      "published": "2026-10-13T11:28:03Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b192eec70e076153e1c8362075e527f6914e4eabc610688ce23d42cf443d8e65",
      "priority": "P1",
      "reason": "The Agentic Scientist 是完整研究编排/验证管线，不是仅多agent对话。附录有20–60 ideation cycles、每candidate最长3小时、250epochs和最高1B环境steps等任务配置，超本轮预算。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3.1 Research Task Setup",
        "G.2 Details of Doc-to-LoRA Inductive Bias",
        "G.4 Details of Error-Diffusion experiments"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18933v1",
      "source_sha256": "ece78fdc7f1b8e2df3b728667ee61b125921501c3fa92bb0896911bbf247d228",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18933v1",
        "basis": "正文附录任务预算；没有执行任何实验或把论文评审proxy当科研能力。",
        "scale": "20–60 ideation cycles；最长3h/candidate；最高1B steps",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.17895",
      "title": "Interaction-as-a-Judge",
      "published": "2026-10-13T07:49:50Z",
      "revision": 2,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "abb31988f03961db2204da8575b73e7d47026b1a7adcc7d4d7c2bd35b90c86f8",
      "priority": "P1",
      "reason": "Interaction-as-a-Judge 用真实电脑交互评估和修复应用，训练奖励还需交互轨迹。v2附录G为Qwen3.5-9B、384prompts×8采样=3072轨迹/batch；完整推理与RL链路成本过高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "4. Interaction-as-a-Judge for Evaluation",
        "Appendix G Reinforcement-Learning Configuration and Evaluation"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17895v2",
      "source_sha256": "3c46f051d7f48c1750f742e7a7a35622863a5a0eaa3a1cd66fa30148a4af4401",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17895v2",
        "basis": "原文v2交互30动作和RL配置，不把静态代码judge或手写轨迹替代电脑交互。",
        "scale": "6858应用对/222应用；9B、3072轨迹/batch",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.17867",
      "title": "RT-J: Large-Scale Pretraining of Relational Transformers for Context-Efficient Predictions",
      "published": "2026-10-13T07:27:44Z",
      "revision": 2,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "78f9408a89bc6eb0978332e0d17bef4c7eaab296392c444cbafcab192967390b",
      "priority": "P2",
      "reason": "RT-J 是85M relational foundation model的大规模预训练/表格预测，虽比较agent featurizers但不提供Agent/RSI算法；归foundation-model范围，Agent track不重复收作实现。",
      "locations": [
        "Evaluation.",
        "Setup.",
        "Appendix G Pretraining Details"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17867v2",
      "source_sha256": "c6fe5ce3df499ba3fbd6d8634203de7d2241582f69e49f18292a4bcbc1db6e0f",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.20563",
      "title": "Scaling Discovery Requires Forgetting",
      "published": "2026-10-14T17:49:37Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "8a2c4e41ee29e54cc8f7029359741892c2f369c28ca4b91ab78848d75dc715c4",
      "priority": "P1",
      "reason": "Forgetting 清空搜索会话/历史而保留best candidate，机制有价值；但研究包含多agent、四种可执行科研任务、30提案续跑与多replicate。PDF费用表与图证实是持续搜索预算，不能以常量缓存重置fixture代替完整发现过程。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "PDF page 1",
        "PDF page 7",
        "PDF page 8"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/pdf/2610.20563v1",
      "source_sha256": "9c14df8357e1fb46caa8349e6225d502180ee74c20c62d5af97ec29cd4285edc",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/pdf/2610.20563v1",
        "basis": "费用来自原文表，完整跨任务多臂还含kernel/model训练；本轮不执行。",
        "scale": "四任务；1/4个agent；30提案续跑与三replicates；单对照$3.44–$27.51",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.20560",
      "title": "PhySGA: A Physics-Space Grounded Agentic Framework for Generalizable and Interactive 3D Composition",
      "published": "2026-10-14T17:47:24Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "971d824fca941d63bd6f408fa74d0ebe5319babd040884f7d55cacc9a286c85a",
      "priority": "P1",
      "reason": "PhySGA 需Assembly/Critic、真实物理primitive、skill反馈和Isaac Sim完整资产链。§4.1在120 prompts上用GPT-5.5与Isaac Sim 4.2.0；硬件、资产准备及反复物理闭环总成本仍无法廉价核定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3 Method",
        "Implementation Details."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20560v1",
      "source_sha256": "7783d6d2de7af95422d5bcf86a202f75702612e02dc0cd4add3628842334cafd",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20560v1",
        "basis": "已取官方方法与设置，不能把缺资源细节当零成本或用坐标启发式替代物理执行。",
        "scale": "Isaac Sim；120 prompts；GPT-5.5；双agent闭环",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.20548",
      "title": "PerSo: Investigating Facets of LLMs' Social Intelligence in Real-World Persuasion Scenarios",
      "published": "2026-10-14T17:39:35Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "a6856d3499a7f1dc3635f9b1d1e10c44e0bd830d46530341159043feadc820c2",
      "priority": "P2",
      "reason": "PerSo 是基于真实CMV对话的社会认知/行为评测，并与人类比较；本轮保留测量参考，不创建新的Agent/RSI能力或伪称persuasion机制复现。",
      "locations": [
        "4.1 Evaluation Setup",
        "4.1.1 Model Evaluation"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20548v1",
      "source_sha256": "fd2fb11eba7850672f5b7fa0229ea3fcf78aa64d8c5004794ff40d3ff9f8cf0e",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.20457",
      "title": "Same Claim, Different Interpretations: How LLM Agents Resolve Ambiguity in Scientific Claims",
      "published": "2026-10-14T16:42:05Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "c7e9c7b7724284e04253b459ad2c104aa6c05e05c96ded645d6bf411f8725676",
      "priority": "P2",
      "reason": "Claim Cards 将科学claim的九个解释字段显式记录；900次运行等检验解释敏感性，并不证明自动科学发现能力。作为歧义/证据审计参考归档，不重跑大型多条件研究。",
      "locations": [
        "Appendix B Experimental Setup in Detail",
        "Overhead:"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20457v1",
      "source_sha256": "298717944a351f5af960f9eba62927fc27f30705a1a74adb0ab6b7920f745eeb",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.20107",
      "title": "Why Generate a Click? Plug-in Decision Heads for Web Agents",
      "published": "2026-10-14T12:39:51Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "197d90494ab7ce79bbaf9e75686761e362b4cd2f9210062d94ed252870a12cea",
      "priority": "P1",
      "reason": "Plug-in heads 需要先有213K web steps训练出的4B/8B generator，再训练heads与级联；不是仅加一个小线性层。附录B两GPU训练、4B joint约11小时及96GB卡配置，成本过高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3.4 Confidence Cascade with a Shared Prefill",
        "Appendix B Training Details"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20107v1",
      "source_sha256": "6abb5c5443c150a2b9e000606af818e8ee851e38ac79130b86c7ff5b41d92def",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20107v1",
        "basis": "正文训练时间和资源配置；不以单次推理节省token代替训练总成本。",
        "scale": "213K steps；4B/8B；2 GPUs；4B joint约11h",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.20085",
      "title": "Designing Textual Interfaces To Enhance Llm Reasoning On Audio-Visual Question Answering",
      "published": "2026-10-14T12:23:59Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "bfb2260572302965f5d782210f1526ab49f58b851edb1f957e4c9a37136f0f3a",
      "priority": "P2",
      "reason": "本研究比较AVQA三种textual interface与GEPA/COPRO优化，125train/125validation/947test明确隔离。属于接口/优化兼容性实证参考，不把组合现有APO框架登记为新的Agent能力实现。",
      "locations": [
        "3 Experiments",
        "3.1 Evaluation Dataset"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20085v1",
      "source_sha256": "2601c8491540251b122b2c51f6a4a529f5df49bd97ebfa44864e42a8a8fbdaf8",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.20067",
      "title": "Which Knowledge Should LLMs Trust? Uncovering a Generalizable Steering Direction Across Knowledge Conflicts",
      "published": "2026-10-14T12:06:32Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "90dfaa42c5e374f56a51f2dbce0c17cd8251f69db80a52f934cd8039a014c706",
      "priority": "P1",
      "reason": "知识冲突偏好方向是残差空间的线性干预，原文采用7B–9B、单卡A100、仅推理。接受真实模型上独立开发冲突对提取方向与隔离测试的核心候选；不把先按clean正确筛出的条件准确率当无条件能力，也不把评测答案送入推理。",
      "locations": [
        "3.5 Evaluation",
        "Appendix D Implementation Details"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20067v1",
      "source_sha256": "ebfb67790338d6d9a3c13c04c82b97b10fd8cfafdc7f02a8b4b3c25fdddeb818",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "accepted",
      "resolution": "affordable_core_review_candidate",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20067v1",
        "basis": "单模型少量开发/held-out样本即可执行真实方向提取和干预，无需复跑四模型全套；并非本地测量。",
        "scale": "7B–9B；1×A100 80GB；最大8192输入/200输出，无fine-tuning",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      },
      "acceptance_boundary": "只进入人工审查候选；本轮未实现，未进入自动执行/evolve，未来实现须保留定义算法并另过数据隔离、预算和适用GPU证据门槛。"
    },
    {
      "id": "2610.19906",
      "title": "Beyond Benchmark Accuracy: A Diagnostic Test for LLM Chemical Prediction",
      "published": "2026-10-14T10:03:05Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f911ebfcfafef3e4fcfcd6142cbe826f4f3192a37e22b40b4ec20bcce3744217",
      "priority": "P2",
      "reason": "该文是化学预测中的identity-blinded结构属性诊断，不是交互Agent或RSI；保留测量边界但不纳入Agent复现，不能以缺少agent关键词代替此内容判断。",
      "locations": [
        "abstract six chemical tasks",
        "4.1 Experimental Setup"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19906v1",
      "source_sha256": "f769afbd120d637b33d6f5582b913470637a3a149f8c0c4ae9c8869edabdcbda",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.19869",
      "title": "SEA-HELM Audio: Evaluating Multilingual ASR-LLM Pipeline Performance Across SEA-HELM Subsets",
      "published": "2026-10-14T09:29:36Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1d763e4861e764995ed534bb9c9bd17f1bcc1db5daad7e9bfabd1ca822550bdd",
      "priority": "P2",
      "reason": "SEA-HELM Audio 测量相同verifier下speech/text性能差，主要贡献是多语言ASR→LLM评测协议，不是新推理/自改进算法。归档为voice-agent可靠性参考。",
      "locations": [
        "2.2 Datasets and Task Selection",
        "3 SEA-HELM Audio"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19869v1",
      "source_sha256": "9675565bbd2aaaa86cbf06ed0ee8bc94149dda693e9d8ef8621170d2052fdde4",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.19801",
      "title": "Delta Tokenizer: Reference-Identical Tokenization of Session Deltas and Segment-Parallelism for Long Prompts on Multicore CPUs",
      "published": "2026-10-14T08:44:46Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6f096b6efe619d4b6ab631210d661d18805c54f6db7661751ea3be06d2dbe65a",
      "priority": "P1",
      "reason": "Delta Tokenizer 的核心是由 tokenizer 拆分规则证明安全的增量接缝、精确分段及 chunk 调度，而不是简单拼接 token。CPU 核心可在本地 Mac 以真实 tokenizer 和原始文本核验一致性；不复跑 41.8B 文档或多服务器服务实验，不声称已测 TTFT。",
      "locations": [
        "Conclusion",
        "Why the splice is exact",
        "Experimental Setup",
        "Limitations."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19801v1",
      "source_sha256": "4b9de83985cbaf9f45c8c15792d5824f589d319c47c3fa6dec9ea0aeab9f861f",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "accepted",
      "resolution": "affordable_core_review_candidate",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19801v1",
        "basis": "正文将分词与 GPU 服务实验分开；可只运行真实 tokenizer 的精确增量/分段路径，限制为少量 CPU workers。",
        "scale": "CPU tokenizer；每 worker 约320 MB，核心不需要 LLM 权重",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      },
      "acceptance_boundary": "只进入人工审查候选；本轮未实现，未进入自动执行/evolve，未来实现须保留定义算法并另过数据隔离、预算和适用GPU证据门槛。"
    },
    {
      "id": "2610.19799",
      "title": "Evolve Together, Fall into Place: How Social Scaffolding Shapes Self-Evolving Multi-Agent Systems",
      "published": "2026-10-14T08:42:05Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "511527c8d23c08623bef793601233ea410015e6f8a9ce75865a3f64394025c84",
      "priority": "P1",
      "reason": "AGORA 研究社会scaffolding如何改变自演化team，而原实验8张96GB RTX PRO 6000、900秒session、多次重复。完整进化和控制组成本过高，不把组织标签或静态角色配置当实现。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "Appendix G Infrastructure and Validation",
        "Limitations"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19799v1",
      "source_sha256": "1bdac11708642668483bf9e8a97d011a543136b82f640d2d6cd3bda0c92c2e8b",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19799v1",
        "basis": "正文硬件/session上限和重复结果；不是本地GPU运行。",
        "scale": "8×96GB RTX PRO 6000；900秒/session；avg@3",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.19762",
      "title": "PRISMATIC : Prescription Risk Inspection System for Multi-Agent Treatment Interaction in Clinical Decision-Making",
      "published": "2026-10-14T08:17:05Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "cdde586050982b5f0a6b8e60c4a6d38c2e36be0152835702ede458515a35388a",
      "priority": "P1",
      "reason": "PRISMATIC 的DSWA/DFPC需要多专科代理迭代并跨病例更新。已查MIMIC-IV病例/处方数据和blackboard方法；访问受限病历、LLM多轮总调用与医学复核成本未能廉价界定，不启动试跑，也不将金标处方作为代理输入。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "5.1. Experiment Setup",
        "4. Methodology – PRISMATIC"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19762v1",
      "source_sha256": "6aa9a8baa2092f82051ef66f2500053fe4ebe874372dc100afc050a303b5f550",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19762v1",
        "basis": "可得原文片段确认依赖病例和闭环；未声称原文报告特定GPU或本地耗时。",
        "scale": "MIMIC-IV三种复杂度病例；多专科循环与历史更新",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19659",
      "title": "Measuring and Reducing Unusable Replies of LLM Agents",
      "published": "2026-10-14T06:35:49Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "ce21af7dec25c385ae518f85a0279a6f71c52a81b21e9700bc9f4d19dd6852c8",
      "priority": "P1",
      "reason": "回复可用性把读者实际见过的信息作为边界，单次重写或追加明确回复要求无需训练。接受少量真实回复上的reader检查与实际重写机制候选；不能用金标可用回复替代重写器，480回复/8模型统计不视为本地结果。",
      "locations": [
        "abstract reply usability and interventions",
        "Reports of agents about their own work."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19659v1",
      "source_sha256": "d7b6a61af35e58573f6321a9ea11aa2a8d119c248e16ba374d321a508ec06581",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "accepted",
      "resolution": "affordable_core_review_candidate",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19659v1",
        "basis": "原文干预是一条追加要求或单次secondary-model改写，不需要多轮训练或长环境rollout；核心可用本地小模型有限调用核验。",
        "scale": "训练免费；每条回复一次reader/rewriter调用",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      },
      "acceptance_boundary": "只进入人工审查候选；本轮未实现，未进入自动执行/evolve，未来实现须保留定义算法并另过数据隔离、预算和适用GPU证据门槛。"
    },
    {
      "id": "2610.19638",
      "title": "J-tau: A Fully Localized Japanese Benchmark for Tool-Agent-User Interaction",
      "published": "2026-10-14T06:07:40Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "2a0fcc0ab8b577052b58bea2194e5524d6344b94c8b094a7124d820e248a7935",
      "priority": "P2",
      "reason": "J-tau 是日本语localization、无效task修正和aggregation修复的基准；gold actions属于scorer，不得输入agent。归档评价数据/协议，不作为能力复现队列。",
      "locations": [
        "Setup.",
        "B.1 Translation",
        "Appendix I Selection of the User Simulator"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19638v1",
      "source_sha256": "559880d3f52f10f340f8a8216efa6b8698ef534145492713257e1b6649e0fb1b",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.19587",
      "title": "Moral Combat: Exploiting Moral Foundations Theory to Jailbreak LLMs",
      "published": "2026-10-14T04:57:48Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "a8b9b62387f62b8a6100ad509f0766d82268e31cb7e112f3693e2e356a638803",
      "priority": "P2",
      "reason": "Moral Combat 是MFT重框架攻击的安全测量工作；只保留失败模式和防护评测参考，本轮不生成攻击、运行红队或将攻击成功率当正常Agent能力收益。",
      "locations": [
        "4.2 Evaluation of models’ moral biases",
        "4.4 Metrics and Evaluation"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19587v1",
      "source_sha256": "ef94e8aa453de38c4d15fa692233781f8d1148abe50f063ccb475a99a9b15821",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.19424",
      "title": "VideoHarness-SRSI: Structured Recursive Self-Improvement for Video Understanding Harness",
      "published": "2026-10-14T02:17:07Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "73c0373d37827da266a26958fdbee816c9fd1c7fc133840e2fec93055247b83a",
      "priority": "P1",
      "reason": "VideoHarness-SRSI 保留content-free反馈、probe-race、factorial组合和独立验证边界；原研究分析4000多个演化候选、四视频基准和三seed，不能廉价完整搜索。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "5 Experiments",
        "Appendix A Experimental Setup",
        "Protocol."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19424v1",
      "source_sha256": "5b8a2b4a1df10c946d6ad9fac0d4e652eb76750f94096d34131a2b4579d93ee9",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19424v1",
        "basis": "原文实验规模，候选执行/视频帧处理仍是绝对成本；不以相对少帧作低成本依据。",
        "scale": "4000+候选；4个视频基准；3 seeds",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.19420",
      "title": "From Failures to Learnable Worlds: Policy-Conditioned Environment Generation for Navigation Agents",
      "published": "2026-10-14T02:10:38Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "92f81112312e3bdbb8f3fbc61b00772df62dace70cdbdca5f65b2009f204719e",
      "priority": "P1",
      "reason": "WEDGE 用失败差异生成并验证weakness，再针对性生成导航世界和训练。正文4.4仍需约21K policy rollouts才得到4.2K候选；十次LLM弱点发现不是整个开销。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "4.1 Experimental Setup",
        "4.4 Analysis and Ablations",
        "A.2.1 Computational cost."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19420v1",
      "source_sha256": "35756bcf41e622d75254c6387da2365db4668545d87485215ee90be25ce82c6e",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19420v1",
        "basis": "论文明确主要成本是rollout验证和适配而非CPU弱点发现。",
        "scale": "3种导航agent；4.2K候选/21K rollouts；后续adaptation",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.19373",
      "title": "Evaluating Compressed Memory: Dependence, Utility, and Content Benefit",
      "published": "2026-10-14T01:17:03Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b7e1325092d22ad6e0c6fbb303a9ba4754f8aadb264d4797988f824b248953a5",
      "priority": "P2",
      "reason": "补读取到的官方HTML正文后确认：zeroing依赖、独立no-memory utility与content benefit是不同对照；72fits及gold-prefix单token故事设置均属受控诊断。只归档评测方法/负结果，不把annnotated evidence、gold prefix或异文档收益当真实Agent retrieval/记忆能力。",
      "locations": [
        "3.2 What Is Held Fixed",
        "Comparison models and training.",
        "Interventions and diagnostic scope.",
        "Limitations"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19373v1",
      "source_sha256": "592f9b43b570960a39728b0a7ba78f6136778f84bdbd9d5b1dd378bf00fbf14d",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "controlled_memory_diagnostic_only",
      "capability_evidence_eligible": false,
      "gold_isolation_requirement": "gold answers/plans、未来结果与oracle控制信号不得输入被评估agent；该诊断不得支持capability/evolve晋级。",
      "source_attempts": [
        {
          "url": "https://arxiv.org/html/2610.19373v1",
          "result": "unavailable",
          "error": "Internal Error returned by web.open; recovered later via alternate official HTTP transport"
        },
        {
          "url": "https://arxiv.org/pdf/2610.19373v1",
          "result": "unavailable",
          "error": "Internal Error returned by web.open; recovered later via alternate official HTTP transport"
        },
        {
          "url": "https://arxiv.org/html/2610.19373v1",
          "result": "retrieved",
          "method": "parent official HTTP snapshot; actual method/resource passages read by this reviewer",
          "source_sha256": "592f9b43b570960a39728b0a7ba78f6136778f84bdbd9d5b1dd378bf00fbf14d"
        }
      ]
    },
    {
      "id": "2610.19366",
      "title": "Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory",
      "published": "2026-10-14T01:11:57Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "2ac812c93d229c52ef8f34710fee0c3d1079cc64f051fb78d9219758bb6e11ce",
      "priority": "P2",
      "reason": "本文提供multimodal memory的two-axis survey/taxonomy与未来原则，没有独立可执行新算法或新能力实验；作为综述文献归档，不伪造复现任务。",
      "locations": [
        "abstract two-axis synthesis",
        "Policy Evaluation and Feedback Attribution"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19366v1",
      "source_sha256": "308796868bf9b1cb042bc6008e1e6f4ba6e776a2d4e13d40ad5684b477861563",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.19316",
      "title": "Truthful and Expected-Payment-Optimal Model Routing in Graph-Based Agentic Workflows",
      "published": "2026-10-13T23:59:09Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "88b90c6a70a1a4fc3ec920153269fbc5a6e33818a75cf7f8577e97d7f39bfdd1",
      "priority": "P1",
      "reason": "STAR 的分配目标、工作量单调性及积分支付可在小型真实合成图上用精确求解器执行，不需要调用 LLM。接受的是拍卖/路由核心机制候选，不把合成成本或激励保证当作真实 Agent 能力提升。",
      "locations": [
        "Experiments",
        ".2 Implementation Validation",
        "H.1 Cost and Token Sampling"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19316v1",
      "source_sha256": "694556e35fae9914bc3e3da28dd38637ef669db1644e03c263ab6aa2e321ba08",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "accepted",
      "resolution": "affordable_core_review_candidate",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19316v1",
        "basis": "正文是数值合成实验，159项实现检查；可在 Mac 保留精确分配与支付算法，不需要大模型训练。",
        "scale": "9个合成 providers；小实例枚举、901 bids、精确 CBC",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      },
      "acceptance_boundary": "只进入人工审查候选；本轮未实现，未进入自动执行/evolve，未来实现须保留定义算法并另过数据隔离、预算和适用GPU证据门槛。"
    },
    {
      "id": "2610.19291",
      "title": "Can LLM Agents Efficiently Identify Appliance Faults? Insights From Diagnosing a Simulated Appliance",
      "published": "2026-10-13T23:12:32Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "4364a24511246b6b44000407ae0d8f696aa0ecec95c13b452c9ae8a969fa264e",
      "priority": "P2",
      "reason": "CoffeeBench 用模拟家电诊断效率而非考试式QA评价Agent，并分析selective prediction heuristics。按基准/诊断研究归档，不把启发式当设备诊断能力实现。",
      "locations": [
        "5 Conclusion",
        "6 Limitations",
        "D.2 Text generation"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19291v1",
      "source_sha256": "e638e90b8d2b35c0c481acc6a27f5ab4bce9d778062b3d875a8c5aaceefe2e1d",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.19267",
      "title": "MemeTrace: A Multimodal LLM-Based Agent for Safety-Aware Meme Provenance Analysis",
      "published": "2026-10-13T22:38:06Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "5e2f060368a69fb5bc1cdc8fcfa50dd6d24c476b42a4f550ee4d5dfc283ccd28",
      "priority": "P1",
      "reason": "MemeTrace 每meme平均18.46步骤、11.54次多模态调用和3.92次search；原文平均$0.1259/111秒。仅100例主集已约3.1串行小时（由论文均值估算，非本地实测），另有多模型/消融/鲁棒性运行，完整复现成本高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "Appendix F MemeTrace Implementation Details",
        "O.1 Execution Profiles",
        "O.2 Cost and Latency Analysis"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19267v1",
      "source_sha256": "dffb0e8195b2bb7c7de46d94398ac65edca0b263c19bcb29210171e57098212e",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19267v1",
        "basis": "基于论文报告均值推导主集串行预算，未把Mac运行设备等同于免费或廉价全实验。",
        "scale": "100 memes；平均111s/$0.1259每例；多模型与消融",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.19131",
      "title": "Plant Memory: Technical Language Processing and a Verified Maintenance Ontology for Agent-Accessible Failure Knowledge from SAP PM Work Orders",
      "published": "2026-10-13T20:02:31Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "8871eec62451771a05ed3ab69dd6637b1876030c88c07dc73f644115e18244fd",
      "priority": "P1",
      "reason": "Plant Memory 的 rule-first TLP、带证据/审核状态的事实图及 SQLite 离线后端可低成本实现，LLM第二读者是可选项。接受规则/溯源核心候选；400条合成工单的数字只是受控机制证据，不能宣称真实工厂部署、通用Agent能力或自动evolve收益。",
      "locations": [
        "5 Implementation",
        "3 Problem setting: what an SAP PM export looks like"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19131v1",
      "source_sha256": "e784bab08a9ddde4c59dc62d8a43993a9ed53e212389ec4cd9a6ef71ac28d886",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "accepted",
      "resolution": "affordable_core_review_candidate",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19131v1",
        "basis": "正文提供SQLite离线实现，规则主路径不依赖120B可选LLM；Mac上小数据CPU路径可行。",
        "scale": "400条合成工单、28诊断问题；Python/SQLite/BM25",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      },
      "acceptance_boundary": "只进入人工审查候选；本轮未实现，未进入自动执行/evolve，未来实现须保留定义算法并另过数据隔离、预算和适用GPU证据门槛。"
    },
    {
      "id": "2610.18874",
      "title": "ChunkAct: Action Chunking for Agentic Robotics",
      "published": "2026-10-13T17:59:57Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "45309253a8eeb1107061cdacf18e9621167e57e89c04a541122fb66e5e4fe22b",
      "priority": "P1",
      "reason": "ChunkAct 必须真实执行VLA/分割/失败退出后重新规划；原文LIBERO-Pro多套件、GPT-5.6-sol high、每episode最多100 calls。完整VLA和仿真资源成本未能廉价界定，不用固定动作chunks冒充机器人能力。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "4.1 Setup",
        "Implementation."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18874v1",
      "source_sha256": "a957a5be89fbd78fff6be26e43ecf79f0d4c3e3715e67321d4052d08b2ee04d1",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18874v1",
        "basis": "正文实际机器人软件链与重复suite设置；仅相对减少LLM调用不足以核定总成本。",
        "scale": "8套件×10任务；VLA primitive、SAM3；最多100 calls/episode",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.18792",
      "title": "SEAR: Self-Evolving Agentic Robotics",
      "published": "2026-10-13T17:40:20Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "ada49e7e12ccde36fd47803b6dda36255dc99168e1b1b60a8bf9c1338f752512",
      "priority": "P1",
      "reason": "SEAR 区分保留经验与重复尝试、privileged state和预算，审查价值高；330模拟/5物理任务、四年以上累计agent runtime以及多小时practice/AR配置明显超预算。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract study scale",
        "Appendix E Interaction Budgets",
        "Hardware setup."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18792v1",
      "source_sha256": "0252917691ae11801f3226df642b3d7af5ce6d0c917c2006369f00f38ef040e2",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18792v1",
        "basis": "论文报告规模与硬件，不申请真实机器人或长实验。",
        "scale": "330模拟+5物理任务；累计4年以上runtime；真实双臂平台",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.18780",
      "title": "BLADE: Auditing Information Use for Failure Attribution in LLM Multi-Agent Systems",
      "published": "2026-10-13T17:34:34Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "de3200a1e0b0d9ca20d621bdf17e2c7bcbf80bc90d5030ce5f2fda4b3bf1ae55",
      "priority": "P1",
      "reason": "BLADE 审计来源到决策的使用关系，区分引入、传播和修复；原文明确不输入参考答案。全文片段报告1228轨迹的 Jev 推断估算约$1.9，属于廉价真实轨迹审查候选；仍需真实决策模型，不得用金标责任步骤做 fixture。",
      "locations": [
        "4 Experiments",
        "B.2 Baseline Implementations",
        "Inference cost."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18780v1",
      "source_sha256": "baaf13220c52b089b9edc533f61084e6feca98275e37b25d09b2c10529bf913f",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "accepted",
      "resolution": "affordable_core_review_candidate",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18780v1",
        "basis": "论文费用覆盖依赖重建、contract auditing、global diagnosis；是论文估计，不是本地付款或实测。",
        "scale": "1228条失败轨迹；结构化 Jev 判断；约$1.9论文费率估算",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      },
      "acceptance_boundary": "只进入人工审查候选；本轮未实现，未进入自动执行/evolve，未来实现须保留定义算法并另过数据隔离、预算和适用GPU证据门槛。"
    },
    {
      "id": "2610.18497",
      "title": "Judges Follow the Reference: The Derivation Gap in Verifying Tool-Call Arguments",
      "published": "2026-10-13T14:52:59Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f974c1475404f76a4ec693cb43938f1e9d464ced58bb69441b879711571a3e2f",
      "priority": "P1",
      "reason": "DeRef 用不见candidate的独立deriver从完整历史生成参考，而不是给judge直接喂gold。原文27B deriver、多个大bench、长历史并报告每reference调用秒级成本；完整多judge/多设置总运行成本高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "5 Experiments",
        "DeRef at the call level.",
        "Appendix F Implementation Details and Cost"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18497v1",
      "source_sha256": "497161861d9de0a3cd10435c14ef7b7d43adb89e4fe7d973ecb191987878309b",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18497v1",
        "basis": "论文规模与调用计时，单个judge低毫秒不包括derivation。",
        "scale": "27B deriver；1871/1442/7428 prompts；每reference约2.44–7.27s",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.18262",
      "title": "Encrypted Skills Can Become New Security Exploits If AI Agents Know How to Decipher",
      "published": "2026-10-13T12:25:46Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "2a00dc639fceb055970ccc3d390925963522086ab0e645f9e17eaf6f219cbeea",
      "priority": "P2",
      "reason": "CSI 是skill注入的安全诊断，区分decode与attack success且短skill结果不能外推原长skill。保留威胁评测参考，不实现攻击或把benchmark代理当通用能力。",
      "locations": [
        "4 Evaluation",
        "7 Limitations"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18262v1",
      "source_sha256": "ea9189f687b89a8f4d11a571e0218f27085cbdcf03afdea55e70f3f962693e59",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.18238",
      "title": "Heading Hierarchy Reconstruction for Complex and Long Documents with Level-Wise Multimodal Memory",
      "published": "2026-10-13T12:06:28Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "0921b770e782a950f09f8064abd6c0b7fa4b528dee76a91b189518f4b45ae57a",
      "priority": "P2",
      "reason": "本法用level-wise multimodal memory恢复长文档heading hierarchy，中心任务是文档解析而非交互agent/RSI；归文档/多模态模型范围，不在本track建能力实现。",
      "locations": [
        "5.1 Experiment Setup",
        "4.2 Annotation"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18238v1",
      "source_sha256": "8fea9671aa06fac2e967947c97ef44e4ccaf47319e4c0756a150b9909e5c78c8",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.18157",
      "title": "Fan Out or Relay? Ownership-Aware Routing for LLM Agents Querying Siloed Tables",
      "published": "2026-10-13T11:01:37Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b74d05dd97378462ddbf3bcba591cd870b699b342d1994750b3b425c2b0cdcca",
      "priority": "P1",
      "reason": "OAR 仅用公开schema生成一份草拟SQL并计算owner cover，在读任何私有行前决定fan-out或relay。核心可在小型真实数据库和实际SQL草拟模型上低成本检验；不能直接喂gold SQL/owner答案替代草拟，不声称13种设置已复现。",
      "locations": [
        "2 Problem Setup",
        "1 Introduction",
        "Appendix E Models and Runtime"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18157v1",
      "source_sha256": "33e8efa4bfbf1dd69d6c2761f310b6ab0afe4e4a679e86c26d318bfd505ea976",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "accepted",
      "resolution": "affordable_core_review_candidate",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18157v1",
        "basis": "执行的是schema-only草拟与确定性路由，8B drafter可单卡运行；小型公开表保留跨owner依赖，不需要原文全部模型组合。",
        "scale": "公开schema一次草拟；owner-cover计数；无模型训练",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      },
      "acceptance_boundary": "只进入人工审查候选；本轮未实现，未进入自动执行/evolve，未来实现须保留定义算法并另过数据隔离、预算和适用GPU证据门槛。"
    },
    {
      "id": "2610.18121",
      "title": "Narrative Lookahead in LLMs: Event-Level Decodability, Predictive Advantage, and Steering",
      "published": "2026-10-13T10:34:26Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "9b784014679f3580fae5ec08415e1db95398855996b370bad1036881d1f08508",
      "priority": "P2",
      "reason": "Narrative lookahead 用hidden-state probe/steering研究未来叙事event，非Agent长期执行或RSI；100-seed延伸也不是独立pilot复现。按研究任务范围排除。",
      "locations": [
        "abstract event-level probing",
        "Relation to the 100-seed experiment.",
        "4.4 Probe-Guided Activation Steering"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18121v1",
      "source_sha256": "72e0e47d7420c9c1be7bff626e5503902e39cc816bcf4f0fdf5632c0aec0e05a",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.17794",
      "title": "GIFT-Eval-CTX: A Benchmark For Contextual Forecasting Model Evaluation",
      "published": "2026-10-13T06:21:50Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "18a574781a3d6fe3be5bb31935f429b3883459eb78d89d583827d1c0ae409fff",
      "priority": "P2",
      "reason": "GIFT-Eval-CTX 的context由已实现未来派生，是有意测试已知future信息的使用能力。仅作oracle-context forecasting诊断；不能作为point-in-time预测、无泄漏Agent能力或evolve提升证据。",
      "locations": [
        "abstract future-derived effects",
        "Data source.",
        "Evaluated methods."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17794v1",
      "source_sha256": "365c2e913ea57a4e2ec3b825f176c2e5e1dcc866e553efa3cb724a74ac051e74",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "future_information_diagnostic_only",
      "capability_evidence_eligible": false,
      "gold_isolation_requirement": "gold answers/plans、未来结果与oracle控制信号不得输入被评估agent；该诊断不得支持capability/evolve晋级。"
    },
    {
      "id": "2610.17786",
      "title": "Agentic4PDM: Host-Governed Evidence Admissibility for Industrial Fault Diagnosis",
      "published": "2026-10-13T06:14:27Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b2a18d98cb634fe9d3f9b628c197f41db22b7a34ecc1883b450cb51df217800d",
      "priority": "P1",
      "reason": "Agentic4PDM 的host治理不等于无需检索/推理。正文附录F显示按样本的复杂调查与DTW基线超过千秒，原系统还需要预训练表征、类别参考和真实工业数据；完整成本无法廉价界定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3.1 Framework Overview",
        "3.5.2 Packet and Influence Gate: Controlling What Each Result Can Support",
        "Appendix F Empirical Time and Token Use"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17786v1",
      "source_sha256": "e23e7b7da29bceeb1d351e5bc685b88cbc692455cf737028c4e767b3eaea263a",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17786v1",
        "basis": "将基线与本法开销区分；四美分是论文相对LLM费用，不是总CPU/数据成本。",
        "scale": "类别参考、信号分析及闭环调查；对照DTW检索约1044.51s/query",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.17777",
      "title": "Beyond Belief Nesting: Rethinking Higher-Order Theory of Mind in Large Language Models",
      "published": "2026-10-13T06:03:36Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "02cad6cc6b3b47e8580689b8e08a1b804304a2cfda3c5edce2c23d688715d464",
      "priority": "P1",
      "reason": "HopTrace 一次编码透视记录并确定性回答多层ToM问题，不能用gold belief graph替代编码器。原完整比较100 narrative/500主问题、其他基准与多模型，报告4–217秒/story且开放大模型4×A800；本轮不能廉价界定忠实完整配置成本。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "4 Method",
        "Evaluation benchmarks.",
        "Efficiency.",
        "Appendix D Experimental Details"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17777v1",
      "source_sha256": "4c4d5bf110cd034c716a1a96d4f2fc0516a825e218f3dbdcb2457f7cf81da3c8",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17777v1",
        "basis": "原文保留单call相对优势，但多模型全评估的绝对预算仍高/不确定；不运行探测。",
        "scale": "100主narratives；一call/4235tokens/story；4–217s/story；多模型",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.17751",
      "title": "Degrees of Freedom in Multi-Agent Coordination: Between Centralized and Decentralized Regimes",
      "published": "2026-10-13T05:32:27Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "60c99d128acdb3fc8d069f9edaf91d947e1b27ecb729a96f494c43e53bfda123",
      "priority": "P2",
      "reason": "DoF 是固定80B/five-agent条件下扫中间协调区间的经验研究，并无统一最优配置。保留组织形态测量参考，不把参数扫描或retained-state观测直接注册为自改进算子。",
      "locations": [
        "Model and branch implementations.",
        "C.1 Evaluation collections and sweep allocation"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17751v1",
      "source_sha256": "456d00de1506883a3d5fd02a6af04d6f1ffb3ccfb9c44e905fc638f6f2ed4fc6",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.17727",
      "title": "SkillScreen: High-Recall Screening of Malicious Agent Skills with Decision Models",
      "published": "2026-10-13T05:03:43Z",
      "revision": 2,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6703f15479548dc2d5cc7897797b6b8301fde5d07fa41c4036c109aa22caa6b2",
      "priority": "P1",
      "reason": "SkillScreen 虽然是compact decision model并只占7.3GiB，v2附录明确单训练run仍需3.4–3.9小时（10GB H100 MIG）；多split/seed不是低成本完整复现。保留高召回筛查参考，不把少分钟inference当训练预算。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3.1 LoRA fine-tuning",
        "Optimization and hardware.",
        "Appendix N Judge prompt and request settings"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17727v2",
      "source_sha256": "0c8e07447c3a70d3ed6a91387b5da8e26aca3ccd4441973d94662af630588640",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17727v2",
        "basis": "该时间来自论文MIG测量，不等同于A100实测或可假设线性加速；完整比较成本过高。",
        "scale": "426–470 updates；3.4–3.9h/run；峰值7.3GiB",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.17673",
      "title": "RALI@iKAT 2026: Retrieval-aware User Profile Denoising via Personalized Conversational Dense Retrieval",
      "published": "2026-10-13T03:59:13Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "eb31a3ea7460dbc29cf06a228f53547408d5a783abb267489419cc6091ad5d03",
      "priority": "P1",
      "reason": "PCDR 需先训练会话retriever，再联合投影用户profile。原文基础conv-Qwen3使用四GPU、batch480、20epochs和跨GPU negatives；不能只用固定embedding替代denoising学习。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "4.6. Implementation Details"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17673v1",
      "source_sha256": "a34585af4f8a6241dfdf65262e79ad8d12fb8d17355bb5d51d2f38d09ba5a9d9",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17673v1",
        "basis": "正文训练配置，不因个性化注释少而忽略基础retriever成本。",
        "scale": "4 GPUs；batch480；20epochs；Qwen3-Embedding",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.17633",
      "title": "Executable Prefix Values for Locating Agent Failures",
      "published": "2026-10-13T03:00:35Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6dca0b9d172b27ebfeac8efd1fdea9e4229de8797563f898cf1e43fa81864c5f",
      "priority": "P1",
      "reason": "Executable Prefix Values 需要重建中间repo状态并多次真实续跑来估计V(h_t)，不是给judge标签重打分。150 SWE任务、47/60条失败曲线与多种restart schedule需要大量continuations，完整成本过高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract restart protocol",
        "Appendix R Cheaper schedules, in full",
        "X.9 Cheaper schedules, in full"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17633v1",
      "source_sha256": "105518dd206151d8bbabc3fe4992d1add44d68a5a594a8b958aee2b37ea49330",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17633v1",
        "basis": "原文强调更便宜schedule仍需要续跑；不把离线曲线重采样冒充真实恢复能力。",
        "scale": "150 SWE-bench任务；47/60失败轨迹；大量前缀重复continuations",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.17560",
      "title": "EdgeWise: Bringing Long-Horizon Agents to the Edge via Selective Cloud Delegation",
      "published": "2026-10-13T01:32:16Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "faa8112bd7d12c223b2c73eb57e58e3bbed245410ed3105c83577aadb22df3e1",
      "priority": "P1",
      "reason": "EdgeWise 的核心是joint solving/delegation step-level RL，而不只是路由标签。9B本地模型、frontier云模型、80turn长任务和RL采样/对照总成本仍明显不确定，不以62%相对省钱当完整复现便宜。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "4.2 Main Results of Our Framework",
        "abstract step-level RL"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17560v1",
      "source_sha256": "6589774721610c9793ba792766ea61b7158ccac82d8599316d9951f848437562",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17560v1",
        "basis": "已查完整方法和主实验组合，未用试训练或云API调用来测预算。",
        "scale": "Qwen3.5-9B + GPT-5.5/Opus；80turn；RL训练",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.17544",
      "title": "Many Steps into One: Reasoning Contraction and the Limits of Safe Memory Reuse for LLM Agents",
      "published": "2026-10-13T01:16:37Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "78e33ae04b2e564871f79f3a86e70c7ebdf287100a34a24b5d021d544b29372f",
      "priority": "P2",
      "reason": "reasoning-contraction的统计证书/likelihood-ratio检测有价值，但replay会通知策略每次改变gold答案的edit。仅作机制/安全界限诊断；不能据此宣称在未知变化中通用记忆能力已提升，更不能将oracle通知fixture进入evolve。",
      "locations": [
        "Compute and software.",
        "Ledger with restarts (restart only)."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17544v1",
      "source_sha256": "29c3431f2118fc6571da0c54334800a58add950a5085efde2de3ef5a734678b5",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "oracle_assisted_diagnostic_only",
      "capability_evidence_eligible": false,
      "gold_isolation_requirement": "gold answers/plans、未来结果与oracle控制信号不得输入被评估agent；该诊断不得支持capability/evolve晋级。"
    },
    {
      "id": "2610.17363",
      "title": "UMCR: Universal Medical Concept Recognition through Guideline Integration and Agentic Linking",
      "published": "2026-10-12T20:19:56Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "7290007d42fb34d24e72510ea20682ee8cb03cb7c98b62d533ddf04db96e781b",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（UMCR: Universal Medical Concept Recognition through Guideline Integration and Agentic Linking）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17363v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.17118",
      "title": "When LLM Agents See a Conflict but Still Proceed: Cross-Domain Dependencies in Network Slicing",
      "published": "2026-10-12T17:10:07Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1acae813ca9846844c4dc8977d7cf31ed1e24cb73d9c0d8df65875eaab4a71eb",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（When LLM Agents See a Conflict but Still Proceed: Cross-Domain Dependencies in Network Slicing）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17118v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.17039",
      "title": "MENTOR: Procedural Shielding of Tool-Using LLM Agents",
      "published": "2026-10-12T16:13:31Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "ca74d55b8929bd0684bca1b0b91bc3a75ea514cf816f543dfbd3575316a044b3",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（MENTOR: Procedural Shielding of Tool-Using LLM Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17039v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16859",
      "title": "ANCHOR: Assessing Naturalistic Constraint Hierarchies and Ordering in Repository-Grounded Agentic Coding",
      "published": "2026-10-12T14:28:38Z",
      "revision": 2,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "7586d3848388f5193e661985c3482f582a480a260455cdaf9f2487fd827248e7",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（ANCHOR: Assessing Naturalistic Constraint Hierarchies and Ordering in Repository-Grounded Agentic Coding）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16859v2",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16460",
      "title": "The interpreter as developer: An autoethnographic design study of building a computer-assisted interpreting tool through human-LLM co-creation",
      "published": "2026-10-12T09:19:55Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "de47ecd2a07d275fc7924575c19d312b65f6048205c14d6ae969bcbaea620d72",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（The interpreter as developer: An autoethnographic design study of building a computer-assisted interpreting tool through human-LLM co-creation）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16460v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16343",
      "title": "Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate",
      "published": "2026-10-12T08:02:41Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "5328c2de4daa07e9702c89ff13b5eecd237a1360292a4b40cfe4a227add02c72",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16343v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16288",
      "title": "REVEAL: Reverse Engineering Agentic Latents from Frontier-Model Traces",
      "published": "2026-10-12T07:14:47Z",
      "revision": 2,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "247f649570e72be2b86ecd9c546454a1d29e574e2041c42fc9b9e91ef3f3756a",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（REVEAL: Reverse Engineering Agentic Latents from Frontier-Model Traces）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16288v2",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16281",
      "title": "RelLedger: Relational Information Governance for LLM Agent Organizations",
      "published": "2026-10-12T07:06:54Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1f27cc9b5dc3b5068baf591e621cae6904ec3e09c049765e3536e39594ff9f61",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（RelLedger: Relational Information Governance for LLM Agent Organizations）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16281v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16165",
      "title": "Think Before You Speak: Enabling Training-Free Latent Communication Between Language Models",
      "published": "2026-10-12T04:40:48Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "7fb2d903925d9b4478c1e5502506af9b236b49e03ab3bbae5e1d8a1803826f53",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Think Before You Speak: Enabling Training-Free Latent Communication Between Language Models）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16165v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16146",
      "title": "Unravel: Recovering an Interactive 3D Scene from a Single Image",
      "published": "2026-10-12T04:08:17Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "44880f073b9159e22e49a6d706c754f5f68136f1cfed402c6cc6d23e90cefd47",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Unravel: Recovering an Interactive 3D Scene from a Single Image）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16146v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16141",
      "title": "Phantom Merge: When Your Large Language Model Agents Pick One but Tell You About Another",
      "published": "2026-10-12T04:01:45Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "33a58bbd3309f6d0bf2eb27c14441721f0ae326b75046e7fc3d2a3fc35f316a3",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Phantom Merge: When Your Large Language Model Agents Pick One but Tell You About Another）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16141v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16129",
      "title": "PID-Gated Continual Prompt Adaptation Under Noisy Feedback",
      "published": "2026-10-12T03:46:35Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b53254c7b8d63739c6be10d166918643db531615686bab185e1c173e366bfac3",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（PID-Gated Continual Prompt Adaptation Under Noisy Feedback）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16129v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16120",
      "title": "Scaling Multi-Agent Intelligence with Substrate-Centric Coordination",
      "published": "2026-10-12T03:37:39Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f5e9a7754324853ffbafad8d7e518096e54d3164cf52abe3b3e1001123a6943f",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Scaling Multi-Agent Intelligence with Substrate-Centric Coordination）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16120v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16071",
      "title": "Dream2Discern: Learning from Hard Latent Variants for Runtime Failure Prediction in Multi-Agent Systems",
      "published": "2026-10-12T02:19:44Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b7652b33c4a80f208a6d515abc89dc74188129cc030f03b370ea0b3bf4032958",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Dream2Discern: Learning from Hard Latent Variants for Runtime Failure Prediction in Multi-Agent Systems）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16071v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16039",
      "title": "When Agents Talk, Secrets Travel: Topology-Driven Leakage and Recipient-Conditioned Sanitization in Multi-Agent LLM Systems",
      "published": "2026-10-12T01:23:51Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "c3b91f5c13867b06eb76aa9b6c878fab4add3c4e57df452c3e4b33d44b9999c5",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（When Agents Talk, Secrets Travel: Topology-Driven Leakage and Recipient-Conditioned Sanitization in Multi-Agent LLM Systems）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16039v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15912",
      "title": "More Reasoning Is Not Always Better: Executing Clinical Guidelines with LLMs and Rules",
      "published": "2026-10-11T20:48:40Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "2dfbad42ea07401060c62d1b0428c8d7805c164ff8b6caacb0adf342660e5037",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（More Reasoning Is Not Always Better: Executing Clinical Guidelines with LLMs and Rules）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15912v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15788",
      "title": "Agent Sociodrama: Social-Dynamics-Guided Persona Recasting for LLM Agent Teams",
      "published": "2026-10-11T17:23:24Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "fdf0a10d270534f82603141b8cc7d006ec567e6ab9adf6f50a26db0a0c19458c",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Agent Sociodrama: Social-Dynamics-Guided Persona Recasting for LLM Agent Teams）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15788v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15666",
      "title": "A Persian Vision-Language-Action Dataset and Benchmark for Ambiguity- and Infeasibility-Aware Instruction-to-Plan Grounding",
      "published": "2026-10-11T14:46:28Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "ab04f73edad91ddfbdab54682143f9da7bbaab9108da5a6e05035bca3f9b78d9",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（A Persian Vision-Language-Action Dataset and Benchmark for Ambiguity- and Infeasibility-Aware Instruction-to-Plan Grounding）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15666v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15488",
      "title": "AlphaPilot: Multi-Turn Tool-Using LLM Agents for Adaptive Alpha Mining with Reinforcement Learning",
      "published": "2026-10-11T09:43:01Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "98de21f0fa63d855b54935251e83a64fd1be01a00ed1b87486edb31e388764a7",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（AlphaPilot: Multi-Turn Tool-Using LLM Agents for Adaptive Alpha Mining with Reinforcement Learning）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15488v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15093",
      "title": "TurnCost: Turns, Not Observations: Cache-Aware Loop Engineering for Coding Agents",
      "published": "2026-10-10T18:03:24Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "a94e510c6794c6c088ccda56de010cc664715170de54a6214aaf9f88676bb331",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（TurnCost: Turns, Not Observations: Cache-Aware Loop Engineering for Coding Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15093v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14958",
      "title": "What Survives When Activation Steering Is Used as Memory",
      "published": "2026-10-10T14:31:47Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "28c70bb87958d036d2a4678d03e8ae8b43ac83193a8cc65927ac44f8f2d22991",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（What Survives When Activation Steering Is Used as Memory）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14958v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14431",
      "title": "MemForm: Benchmarking Memory Formation Poisoning in LLM Agents",
      "published": "2026-10-10T00:52:32Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "16522d386669b3cfc862725fc2d546d83b73a2c3dcb7e2b666832927f074f5d7",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（MemForm: Benchmarking Memory Formation Poisoning in LLM Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14431v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14224",
      "title": "Harness^2: Recursive Agent Harnessing for an Open World",
      "published": "2026-10-09T18:34:25Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b9f8f061c18f0b842e94b35365740e875834dafaab3bc9531a4b9243386c70d6",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-09，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Harness^2: Recursive Agent Harnessing for an Open World）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14224v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.11270",
      "title": "Gated Memory: Admission-Controlled Memory Formation for Conversational AI",
      "published": "2026-10-08T05:27:02Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "166c634b4d13cc352a1293df49f05818ed87eeceee58b76b6495052dd17fb99e",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-08，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Gated Memory: Admission-Controlled Memory Formation for Conversational AI）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.11270v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14071",
      "title": "Loose Lips: Understanding Privacy Leakage in Agentic Dialogue",
      "published": "2026-10-07T23:25:11Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "ef76797dd129e3bc2fd75d01439b85c493d6b743abdc39d8c176e8daa96ec68f",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-07，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Loose Lips: Understanding Privacy Leakage in Agentic Dialogue）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14071v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.09661",
      "title": "Alice: A Large-Scale German Benchmark for Rubric-Based Multi-Dimensional Automatic Short Answer Scoring",
      "published": "2026-10-07T08:29:43Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "0bafa1b51b3d9057cc5d8dfce06c492a4f9fb059002b037368e414210c0dd148",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-07，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Alice: A Large-Scale German Benchmark for Rubric-Based Multi-Dimensional Automatic Short Answer Scoring）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.09661v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.08388",
      "title": "Foresight-over-Graph: Reasoning Beyond Local Horizons for Knowledge Base Question Answering",
      "published": "2026-10-06T14:07:43Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e3714cb02ac9ab0d4f09ca35d2bb926fb053c43eef9033143c5f7f672ac1d605",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-06，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Foresight-over-Graph: Reasoning Beyond Local Horizons for Knowledge Base Question Answering）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.08388v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.06516",
      "title": "Test-Time Adaptation of Reasoning Strategies with Bayesian Nonparametric Memory",
      "published": "2026-10-05T15:30:58Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b0bee0a027d1b9b3af96e3aa3acf55cbb687ed73654121ad8846baf3b4731f50",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-05，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Test-Time Adaptation of Reasoning Strategies with Bayesian Nonparametric Memory）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.06516v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.05732",
      "title": "PACMI: Provenance-Aware Cascading Memory Invalidation for Long-Term LLM Agents",
      "published": "2026-10-05T03:31:51Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "913e1467e6890750fe971188b879150db9f0a4f3b018d07d1dfa3314e3f4808b",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-05，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（PACMI: Provenance-Aware Cascading Memory Invalidation for Long-Term LLM Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.05732v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.06996",
      "title": "Mask-Guided KV Cache Eviction in Block Diffusion Language Models",
      "published": "2026-10-04T07:52:31Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "ce3ffd07393b745306c4a3ea2f98b3d104cff67ec843a8f488dd72da334eb85d",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-04，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Mask-Guided KV Cache Eviction in Block Diffusion Language Models）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.06996v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.03027",
      "title": "Tailoring the Quantization Space for 1-Bit KV Cache Compression",
      "published": "2026-10-02T09:04:25Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "867804daf201b33911a5ede1f190a3c4c12597d25665ea3c566d4b0aa52b81df",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-02，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Tailoring the Quantization Space for 1-Bit KV Cache Compression）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.03027v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.01118",
      "title": "Madeleine: Learning Involuntary Recall for Conversational Memory from Simulated Lives",
      "published": "2026-10-01T05:59:23Z",
      "revision": 2,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "11b7035b344ff962431ec84b2ddf665ff11ac2f802409b9d73fe58059e743b39",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-01，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Madeleine: Learning Involuntary Recall for Conversational Memory from Simulated Lives）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.01118v2",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14018",
      "title": "RoboHarness: A Simple Harness Outperforms VLA and World Actions Models",
      "published": "2026-09-29T11:36:07Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b9a4175d30ffa16454dc0e67e87c3bf57915ecc26f9216f3d4e19a355cffca22",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-09-29，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（RoboHarness: A Simple Harness Outperforms VLA and World Actions Models）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14018v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14006",
      "title": "Coding LLMs Can Be Harnessed to Learn Interpretable Models from Data",
      "published": "2026-09-28T17:05:42Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "4371946cdbd7de4ac09edfca25bf546668f96c2bef86b9dc0e806163985be6af",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-09-28，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Coding LLMs Can Be Harnessed to Learn Interpretable Models from Data）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14006v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14005",
      "title": "Artificiality's First Exam: An LLM-Referenced C-test for Supraintelligence",
      "published": "2026-09-27T20:41:46Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "3ccaeb92c27944b7eaeb515bf51f303b81dedb5baabd20ec5b0b028316492ab6",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-09-27，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Artificiality's First Exam: An LLM-Referenced C-test for Supraintelligence）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14005v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.20584",
      "title": "Player-1: Learning to Play from Synthetic Experience",
      "published": "2026-10-14T17:57:12Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "17a1f6ee5b180d27481933493b4d185556f9b4d4fadab24c5c0b19ca7bee6c72",
      "priority": "P1",
      "reason": "Player-1 的SFT与多环境PPO是核心；官方HTML附录D.3明确每模型SFT为32×H200约24h、RL为64×H200约36h，另有512并发环境。成本显著超出单卡预算，不复现。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "https://arxiv.org/html/2610.20584v1 §4.1",
        "https://arxiv.org/html/2610.20584v1 Appendix D.3"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20584v1",
      "source_sha256": null,
      "source_hash_scope": "not_captured_metadata_hash_separate",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。",
        "本子代理通过官方HTML补读，未保存原始网页字节/hash；source_sha256为null，metadata_sha256单列，二者不混淆。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20584v1",
        "basis": "补查官方HTML得到的明确硬件与wall-clock；未下载模型、数据或启动训练。",
        "scale": "32 H200×24h SFT；64 H200×36h RL；9B/35B-A3B",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.20573",
      "title": "Beyond LLM Judge: Mechanical Verification for Agentic Autoformalization",
      "published": "2026-10-14T17:54:48Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f300822ebb8345f35a8b79d2c9ed494fdb2ba84ce8a42aca4b6f047dae99526f",
      "priority": "P1",
      "reason": "该工作结合LLM语义复核、numerical/vacuity checks，不能只跑零费用vacuity check当完整验证器。附录F给1539项numerical约$314.18/77s每项、back-translation约$99.57，完整机械/模型联合研究成本高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract verification channels",
        "Appendix F Cost and runtime"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20573v1",
      "source_sha256": "ec0844f1459a03be883cd1fd7682b077931852932f4d4656d0cb55012ae6fcda",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20573v1",
        "basis": "原文成本表；不以局部零API操作掩盖核心语义/数值检查链路。",
        "scale": "1539数值检查$314.18；回译$99.57；多对照",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.20437",
      "title": "From Training to Escalation: Forensic Analysis and Local Recreation of RL/RLHF-Induced Privilege Escalation in AI Agents (CVE-2026-65616)",
      "published": "2026-10-14T16:28:09Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "589da714c9e4aedec1e2683207b87f8b4103d62cf0ea2309525d328f00ed27ef",
      "priority": "P2",
      "reason": "这是公开事件和隔离容器recreation的取证分析，不是新的RSI训练算法。本轮只作配置/containment证据参考，不验证其CVE叙事的所有外部事实、不执行privilege escalation。",
      "locations": [
        "Demystifying the AGI Myth: Reward Hacking vs. Autonomous Rebellion"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20437v1",
      "source_sha256": "8c867a02a2bfb2717d02c69e828cda7adfdcc521ad9b92e118b83b618faa2892",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.20394",
      "title": "What Should a Software Team Spend on Agents? Simulation-Based Cost Optimization of Mixed AI-Human Teams",
      "published": "2026-10-14T15:57:44Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1bf1e8049a488cd9cb577701556e73b9fad62cffefb34027fbe42b4b765a6fda",
      "priority": "P2",
      "reason": "混合AI–人类team费用来自离散时间Monte Carlo与进化搜索，是成本模型研究，不是真实Agent rollout或组织收益测量。保留仿真假设和成本敏感性参考，不当能力实现。",
      "locations": [
        "3.5 Configuration parameters",
        "4.2.5 Bug impact parameters"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20394v1",
      "source_sha256": "14f4e54607da348ff575334ebe40f4d3852379f87f23c78fe6ac232a3e99fc9d",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.20388",
      "title": "From Relevance to Validity: Zero-Shot Tool Selection via Functional and Contextual Validation",
      "published": "2026-10-14T15:53:25Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "61b16be70e99121905691bdc2e559179b9a749d2284e4a426b303725a5905e2b",
      "priority": "P1",
      "reason": "工具transfer-validity方法需representation extraction、训练与capability ranking，并在缓存工具环境中调用实际backbones/judge。附录I.4三张95GB卡、三seed及分离远端模型；完整预算单卡下不能廉价确认。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract progressive validation",
        "I.4 Computational Environment"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20388v1",
      "source_sha256": "bc1c74e12b89737a3af3bf4a18a2b358fa093658472a5e53c700ba1e729816dc",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20388v1",
        "basis": "正文资源明确区分本地ranking和不可观测远端硬件，不能把三卡全算本地能力或假设单卡直接等效。",
        "scale": "3×95GB RTX PRO6000；seeds42/43/44；远端agent/judge",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.20344",
      "title": "Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents",
      "published": "2026-10-14T15:24:39Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f84a3ad7b6b1789f731b1e749b02f5594538fb391d2ee4984e06af6490fe01fe",
      "priority": "P2",
      "reason": "Doc2Harness 的Stage2由scripted mock model驱动，测的是harness工程契约而非真实policy能力。适合作为确定性harness测试参考，不能把328测试或mock通过率宣传为Agent能力复现。",
      "locations": [
        "abstract two-stage protocol",
        "2.3.1 Public Issue Mining"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20344v1",
      "source_sha256": "35038fe8e2ff646ed66c544273d83752e72be6e973aa8ec09de8695ee66dfc0e",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "scripted_harness_diagnostic_only",
      "capability_evidence_eligible": false,
      "gold_isolation_requirement": "gold answers/plans、未来结果与oracle控制信号不得输入被评估agent；该诊断不得支持capability/evolve晋级。"
    },
    {
      "id": "2610.20312",
      "title": "Reflect After Failure: Brain-Inspired Self-Evolving Agents for Embodied Spatial Intelligence",
      "published": "2026-10-14T14:59:04Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "28ba65784442732720aa9c7d9ad340f3dee5e426124cad564c0670beac8572b0",
      "priority": "P1",
      "reason": "HiMeco 结合PEP诊断、策略bank和可执行skill；原文八A6000/API、208固定episode共1664 planned executions，成本高。相同episode四round适配与独立跨episode转移必须区分；诊断可看reference snapshot，不得把它输入在线agent或当隔离test改进。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3.1 Problem Setup",
        "4.1 Experimental Setup",
        "C.2 Evaluation episodes and round schedule",
        "Retrospective evaluation and diagnosis."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20312v1",
      "source_sha256": "e5c8e6c915c7f7ea3a57120853cda2c6d969fe338c0f16f5c45cd6c24fc9aafc",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20312v1",
        "basis": "原文setup/schedule与诊断信息边界；不运行仿真或重复测试。",
        "scale": "8×RTX A6000；208episodes×8exec；API VLM",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.20245",
      "title": "P2Patch: Exploit-Guided Vulnerability Repair",
      "published": "2026-10-14T14:19:21Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "27458162a54b66455dad4c28684adbaf7b8bdf66b1c54b50de314278f466cb2e",
      "priority": "P1",
      "reason": "P2Patch 在每次repair后生成独立同漏洞exploit并累计PoV，保留construction/held-out oracle隔离。101真实CVE、多模型攻修闭环与原文明确增加的end-to-end成本超过便宜候选范围。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract construction/evaluation oracle split",
        "6.4 Model Dependence (RQ4)",
        "6.5 Published Baselines (RQ5)"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20245v1",
      "source_sha256": "0537c70f8cc6ef3edac254fc62d9c8dfc567c6dbd3d037f7ca3cdd826a5b594f",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20245v1",
        "basis": "正文覆盖率–成本tradeoff和多轮真实执行；不执行任何目标或exploit。",
        "scale": "101个C/C++/Java CVE；多轮攻修与多模型",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.20200",
      "title": "Drivers of LLM Consensus Formation Under Competing Incentives: A Simulation Study",
      "published": "2026-10-14T13:46:42Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f4d8114aedf6b14a366c3bae46d963af055cc5042eb091e00c57cc72e634ad2a",
      "priority": "P2",
      "reason": "Network Consensus game主要测量模型沟通/网络结构/顽固度下的共识行为；属于多agent行为实证，不是独立训练或RSI能力算子。归档研究参考。",
      "locations": [
        "3.2. LLM Interaction Protocol"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20200v1",
      "source_sha256": "9e6b3f76235be13aa6b3159597e00551ebbb7af22c72e6d95d18e74be23721a8",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.20193",
      "title": "SWE-Stream: Scaling Executable Training Data for SWE Agents via Progressive Construction and Elastic Streaming",
      "published": "2026-10-14T13:41:49Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "0b7fce1ea042c07459639d4fb55fd2bba4ddcf56a710dd80af3e8a1dbb7abf6f",
      "priority": "P1",
      "reason": "SWE-Stream 的渐进任务构造和elastic streaming必须实际重建/验证仓库。80,874任务、118,688轨迹、30B训练；受控scheduler实验达到500任务也需23.06h，完整复现成本过高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "2.1 Candidate Selection",
        "Construction speed and yield."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20193v1",
      "source_sha256": "ef40d9a76546cd6e3e1ec894ccf1a638cd5c068c7d783388524dad916c006c7f",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20193v1",
        "basis": "原文规模及构造计时；不下载仓库大集或生成训练数据。",
        "scale": "10728 repositories；80874任务；118688轨迹；30B",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.20099",
      "title": "EvalBuilder: Can LLMs Build Efficient Benchmarks?",
      "published": "2026-10-14T12:35:02Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "a165361865b3181a38556212f36c092a6de7c61682e956119eb96fa70ac903c1",
      "priority": "P1",
      "reason": "EvalBuilder 要实际为昂贵reference benchmark生成/迭代proxy并核验held-out ranking，不能只写一个打分器。附录B.1真实运行4h33m、7h9m、13h7m和16h上限，超出本轮廉价复现预算。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract meta-evaluation protocol",
        "B.1 Timing and Stopping Behavior"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20099v1",
      "source_sha256": "b44af83e4585bbb5d4ee400af347550a26f0d0d9cec8d1c24d48ae30673d0762",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20099v1",
        "basis": "论文报告实际停止时间和上限，未执行任何benchmark-builder。",
        "scale": "4h33m–16h/agent run；SWE/Terminal reference评测",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.20068",
      "title": "The Code of Theseus: Can Agents Rebuild Software from Recovered Design Documents?",
      "published": "2026-10-14T12:06:37Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "4a43de12e67c47a388ef0a3631f20b70b509cacf458c63cde530fb383236d64d",
      "priority": "P2",
      "reason": "代码Theseus回建研究显示设计结构F1改善但held-out correctness无显著变化；不将结构相似度当能力提高。归档设计数据/评测参考，不重跑近869步骤的库重建。",
      "locations": [
        "VI-C Repository Reconstruction",
        "VI-C 3 Cost"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20068v1",
      "source_sha256": "cd17b070815363d8b29712f8d6a49afce4d95b196fcffdae34bcc1585a7dc35d",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.20038",
      "title": "SWE-Nexus: Automated Composition of Software Engineering Tasks from Test Execution",
      "published": "2026-10-14T11:45:59Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "c399d17e47c3dca8f136ed06a2e5dfd44c300f8541d3d3bf83a82c29e275e6ba",
      "priority": "P2",
      "reason": "SWE-Nexus 从test execution footprint构造/合成600真实仓库task，主要贡献是基准/数据构造。隐藏测试只在scorer侧，scope-weighted partial recovery不能当完整成功；不将基准标记为新能力实现。",
      "locations": [
        "https://arxiv.org/html/2610.20038v1 §5.2–5.4",
        "https://arxiv.org/html/2610.20038v1 §6.2",
        "https://arxiv.org/html/2610.20038v1 §7"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20038v1",
      "source_sha256": null,
      "source_hash_scope": "not_captured_metadata_hash_separate",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。",
        "本子代理通过官方HTML补读，未保存原始网页字节/hash；source_sha256为null，metadata_sha256单列，二者不混淆。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.20022",
      "title": "Memorizing Who, Knowing How: Benchmarking Personal and Procedural Memory for Persistent Assistants",
      "published": "2026-10-14T11:33:40Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "4770386354f62af798f1eeed81a6b3c38fa5a4e4531e0524860491af0da559e6",
      "priority": "P1",
      "reason": "PraxisMem 联合维护个人/程序记忆，并通过privileged joint-state监督、location-aware GRPO和OPSD训练manager。不能以读取gold state的fixture替代可部署manager；训练及20长trajectory成本仍无法廉价核定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract PraxisMem training",
        "F.7 Runtime and Token Efficiency"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.20022v1",
      "source_sha256": "e1a88a569f89a7aec4d347ac1c3a47d7f09ea9486f37adb76effa4ec450bac99",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20022v1",
        "basis": "已查成本将task API与memory GPU分账；他法0.85/2.03h不是本法训练耗时，不混用。",
        "scale": "联合memory-manager训练；20完整test trajectories；长session",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19999",
      "title": "A Strong Harness Narrows the Model Gap: A Comparative Evaluation for Legislative Document-to-XML Conversion",
      "published": "2026-10-14T11:15:25Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "4ba5e727e595ba97661c7151372aa0f41a5708512194a72a21ae56669fb2ba9c",
      "priority": "P1",
      "reason": "立法XML pipeline 的关键是实际工具、持久状态和citation authority lookup，而不是只切分文档。正文明确未记录tokens/latency/$，153 bills×3harness×3模型完整总成本不可从保留产物恢复；不廉价试跑探测。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract crossed ablation",
        "What we do not measure.",
        "5 Evaluation Framework"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19999v1",
      "source_sha256": "74af3679a43c3a4a9dc87cc2df88fb1e2ace150ce3ea224656368b47c1e4e2bc",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19999v1",
        "basis": "原文直接承认成本未记录/不能恢复；材料性预算不确定闭环，不伪造估时。",
        "scale": "153 bills；3 harnesses×3 Claude models；长文档和外部citation lookup",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19864",
      "title": "Euler: Towards Human-Centered Multi-Agent Mathematical Discovery",
      "published": "2026-10-14T09:27:36Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "08c0aeda4d185d1c6616a37d06841f78046c4560fd65ba19ffe2387f66390cef",
      "priority": "P1",
      "reason": "Euler 用research graph与saved team state组织长期数学探索，并具备预算reservation。三数学case不能证明其全流程廉价；已查resource accounting与case范围，frontier调用/证明工具/长期继续的真实复现预算仍不确定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract mathematical case studies",
        "4.5 Resource accounting and long-horizon continuation"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19864v1",
      "source_sha256": "7503f8dd346e4d614effd382739651b1da07afe66d397d3d91543d515545676c",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19864v1",
        "basis": "只读原文预算机制不是实际成本账本；不把graph-only demo当数学发现能力。",
        "scale": "研究图、多agent长期数学case、资源reservation",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19849",
      "title": "DeltaStreamer: Towards Human-Like Agentic Streaming Video Understanding",
      "published": "2026-10-14T09:19:15Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "15f264a823392e17d473775c07b251c544492b706f351d78817f45ad1cd0cc9e",
      "priority": "P1",
      "reason": "DeltaStreamer 的change-aware记忆、planner和历史视频replay均需27B多模态模型。§4.1是8×A100 BF16、1fps处理，不能以文本摘要fixture代替视频观察/回放。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "4.1 Experimental Details",
        "A.5 Look-Back Verification and Replay Details"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19849v1",
      "source_sha256": "8c22ddb99f8029b6d6eb0662cb9e15b6dfbb28a2ffe809c706145e3802cf5fcd",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19849v1",
        "basis": "正文硬件和历史重放开销，成本过高。",
        "scale": "Qwen3.6-27B；8×A100；1fps视频+query-driven replay",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.19843",
      "title": "LAGO: Language Agents with Reusable Growing Options",
      "published": "2026-10-14T09:12:47Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "2e2148e04206d8ad1b41165453b8b5d2303886eac6759a2b06aadb740cb7aebf",
      "priority": "P1",
      "reason": "LAGO 的meta-agent/agent-option library需要真实问题分解、工具执行和结果核验。35B-A3B、500 SWE任务及150任务重跑对照，长真实仓库运行总成本高，不能用预定options或gold decomposition替代。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "4.1 Experimental Setup"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19843v1",
      "source_sha256": "b4bf13ae7bec0d4f97b7037bc79b0d3c1a0eabc1a1fbd182fdeade08926e75e1",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19843v1",
        "basis": "正文明确未统一context/inference budget，单卡完整费用/耗时不能廉价确定。",
        "scale": "Qwen3.5-35B-A3B双角色；500/150 SWE任务；多个scaffolds",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19836",
      "title": "EGCForest: Structured Agent Memory Organization for Enterprise Group Chats",
      "published": "2026-10-14T09:07:53Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "940e5d3afd601eb1060f5243726744e1e1652b2ad523023b666ce1bbe8a09d75",
      "priority": "P1",
      "reason": "EGCForest 的三forest、Placement Map与结构桥需要实际summary构造和embedding。206MB/65ms是存储和在线retrieval，不包括三基准离线建库/回答；0.6B encoder虽小，完整LLM建库预算仍不确定，不因低在线延迟接受。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "Implementation.",
        "A.3 Implementation Details",
        "4.5 Q4: Efficiency"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19836v1",
      "source_sha256": "bc8c4413ae177dbcb421876c0a5ce77d3ade462e79019817bc4781aa6de19c07",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19836v1",
        "basis": "已查建库/查询分工与resource段，完整离线summary成本未从取得片段廉价核清。",
        "scale": "3 enterprise-chat基准；0.6B encoder；每query约2365输入/1300输出tokens",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19820",
      "title": "ActionOPD: Distilling Decisions, Not Tokens",
      "published": "2026-10-14T08:59:08Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "ebbe1fd79266fa63416abad50308ca22c089922c35a876af296504459a4266ec",
      "priority": "P1",
      "reason": "ActionOPD 必须在学生遇到的状态上采样学生/教师动作并用语义一致度做policy optimization；不能只做offline token imitation。4B/8B、多teacher、每状态4candidates与闭环工具评测，完整成本仍无法廉价核定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3 Experiments",
        "Evaluation.",
        "Collection budgets."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19820v1",
      "source_sha256": "c6f9d8e56da384d9f07e878e8588237fdd21626eca1ab8b999c577263070812a",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19820v1",
        "basis": "原文明确相同task starts不等于相同teacher queries/tokens/compute，故不外推便宜。",
        "scale": "4B/8B；2 teachers；每状态4 candidates；多工具环境",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19818",
      "title": "One Monitor Does Not Fit All: Formalism-Agnostic Runtime Verification for Agentic AI",
      "published": "2026-10-14T08:58:24Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e676429a2889eca80eb37d821d80c26bf9f31b413e58e55158a5c0807a5a64d8",
      "priority": "P1",
      "reason": "AgenticRV 的 canonical events、monitor-specific projection、统一 verdict 与 effect 前阻断是可执行的运行时机制。Python/受控工作流可在 Mac 保真核验；只接受该机制，不把立即返回的 monitor 微基准冒充语义监控效果或 Agent 能力。",
      "locations": [
        "4 Implementation and Instantiation",
        "5 Evaluation",
        "5.5 Framework Costs"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19818v1",
      "source_sha256": "4d53292bc97777e40f37233ee757fcdab10653f3c247f159ac2e5a34524616d8",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "accepted",
      "resolution": "affordable_core_review_candidate",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19818v1",
        "basis": "正文把受控验证与微基准分开，20进程、短计时批次可本地执行；无需模型训练。",
        "scale": "Apple M3 Pro 18 GiB；CPU运行时与受控集成检查",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      },
      "acceptance_boundary": "只进入人工审查候选；本轮未实现，未进入自动执行/evolve，未来实现须保留定义算法并另过数据隔离、预算和适用GPU证据门槛。"
    },
    {
      "id": "2610.19698",
      "title": "RuleTrek: Benchmarking Agentic Rule Reasoning in Complex Scenarios",
      "published": "2026-10-14T07:23:08Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "9d4832e5784c23779f0cc50057270371d1f8dfccd115c9794d8af4489a1f9c55",
      "priority": "P2",
      "reason": "RuleTrek 的rule graph、RI/RD/E2E分项是多文件规则推理benchmark，不是已训练的Agent改进方法。保留规则/执行隔离评价参考，不给agent读取reference rule plan。",
      "locations": [
        "abstract rule-grounded construction",
        "3.2 Stage 2: Rule-Specific Sub-Scenario Construction"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19698v1",
      "source_sha256": "cf4aab1cd5b1763fd9e0b7f2cb7cb3b7f5816d0b84623f9abf3e86c9568b1876",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.19695",
      "title": "C3: Rethinking Process Supervision via Confidence-Guided Credit Correction in Multi-Turn Reinforcement Learning",
      "published": "2026-10-14T07:18:13Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f88e4e8c7fa65349dd0eab1c8aa8382a04e1d193f35532985981acb6600bde08",
      "priority": "P1",
      "reason": "C3 在base RL advantage估计后按置信度修改坏span正信用，必须保留真实RL训练。Qwen3-8B search、4B full-parameter WebShop、critic与critic-free多臂及检索模型成本高/仍材料性不确定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "4 Methodology",
        "5.1 Experimental Setup"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19695v1",
      "source_sha256": "ed6130d4501e1584fd6ea00830e66df69edb99d5123dd5c773c0028f003b6903",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19695v1",
        "basis": "已查训练环境和真实detector/advantage接口；不以张量单测或gold-span fixture当能力复现。",
        "scale": "8B search+8B embedding；4B全参数WebShop；多优化器对照",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19596",
      "title": "JevCoder: When Does Decision Delegation Pay Off in Code Repair?",
      "published": "2026-10-14T05:11:25Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1802fe828e7b7894a6ab31262fe6c8b2cb71d6ef67eac2a7c6e0178c72b1a540",
      "priority": "P1",
      "reason": "JevCoder 的低费率judge不代表repair agent整体便宜。原文compute段明确两张L40S、约200 GPU-hours和$1500 API，完整研究超预算。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract decision interventions",
        "Compute and software."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19596v1",
      "source_sha256": "bd38957c914e8db50e2129e8356aeb51f25b67aceed890b242c40b6babb2159b",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19596v1",
        "basis": "论文报告预算，不是本地费用。",
        "scale": "2×L40S；约200 GPU-hours；约$1500 API",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.19589",
      "title": "JockeyBench: Evaluating LLM Agents on Horse-Racing Decision-Making",
      "published": "2026-10-14T04:59:45Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6da0cf0ac6e054eea3bdebf47235532eed2bf7df4d87da3c1ba446a4bce10a1a",
      "priority": "P2",
      "reason": "JockeyBench 是time-gated重复决策评价；所有配置mean profit为负，不能把个别盈利run当金融预测能力。仅归档forecast/utility/retrospective return分账与时间隔离协议，不提出交易或复现收益。",
      "locations": [
        "abstract benchmark results",
        "1 Introduction",
        "Long-Horizon Agents and Cost-Aware Evaluation."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19589v1",
      "source_sha256": "c5f6a5de22bca291d0028b18f69e2aa19cab4cf00711579fdf8d22e11c8948fd",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.19581",
      "title": "XferAgent: Cross-Agent Experience Transfer for Self-Improving LLM Agents",
      "published": "2026-10-14T04:51:53Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "dd04480339a50c6ba0f6d70275531d7bbe82bd27ec75112b34880b8bb73d746c",
      "priority": "P1",
      "reason": "XferAgent 在执行前从local/shared graphs路由经验、执行后用结果更新；不是静态共享提示。四域、异构framework、多task-pass构造和真实回归评测的总成本经现有方法/ALFWorld设置仍不能廉价确认。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3.1 Problem Setup",
        "C.2 ALFWorld: Embodied Task Completion"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19581v1",
      "source_sha256": "b556c117250da18f4ba67f7278fa908edec734440840cc94d18babb785338912",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19581v1",
        "basis": "已查先检索后执行/更新边界；未用gold经验或任务答案作代理。",
        "scale": "异构agents、local/shared graphs、多域重复任务流",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19566",
      "title": "Mission (Im)Possible: Towards Autonomous Discovery of Autopilot Logic Flaws in Open-Source Drone Frameworks",
      "published": "2026-10-14T04:30:11Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "dc74d8b23ea09001db06e730ead98cc27e15c2e44da3e70a23db4cf68b2ddff6",
      "priority": "P1",
      "reason": "IMF 需三agent协作静态slicing、真实模拟测试和log oracle验证；PX4/Ardupilot105个新发现不是可廉价重跑的单脚本。已查官方PDF执行例子，构建/仿真/长frontier调用预算仍材料性不确定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "PDF page 7",
        "PDF page 9",
        "abstract evaluation"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/pdf/2610.19566v1",
      "source_sha256": "e900706ff7457f7d70513564c38dd386e2a8f04e8e18fafe23447433da4160ed",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/pdf/2610.19566v1",
        "basis": "实际原文PDF例子可取得；未执行漏洞、飞控或外部目标，未猜测GPU数。",
        "scale": "PX4/Ardupilot；3角色；105 flaws及动态验证",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19474",
      "title": "Beyond Retrieval: Evidence Access and Event Aggregation in Agent Memory",
      "published": "2026-10-14T02:53:17Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "5cfaedd131ebe23e245ed26f6830c1236dbed4d6eb78c78cf5989fc90941eac7",
      "priority": "P1",
      "reason": "TallyMem 保留原始轨迹、生成事件规格并实际执行聚合，不能以金标事件集代替规格生成。接受原始日志→事件ledger→真实编译/执行的廉价核心候选；$1.05仅属原文primary arm，不能外推全AMA/多reader成本。AMA混入开发轨迹，正式能力比较应采用25开发/99隔离仓库协议。",
      "locations": [
        "7.1 Setup",
        "7.3 Held-Out Repository Histories: Cross-Dataset Evaluation",
        "Main evaluation configuration.",
        "Results."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19474v1",
      "source_sha256": "727439deb3a03566b4bc5de686da24e54ed85e12556552e6172c52e1862c3bd4",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "accepted",
      "resolution": "affordable_core_review_candidate",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19474v1",
        "basis": "原文区分受控probe与冻结principal system；只保留真实规格生成及执行机制，不整套重跑多模型全表。",
        "scale": "primary arm程序生成$0.10、回答/判断$0.95；776题/99隔离仓库是较大评测",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      },
      "acceptance_boundary": "只进入人工审查候选；本轮未实现，未进入自动执行/evolve，未来实现须保留定义算法并另过数据隔离、预算和适用GPU证据门槛。"
    },
    {
      "id": "2610.19472",
      "title": "MemCode: Discrete Semantic Representations for Long-Term Agent Memory",
      "published": "2026-10-14T02:51:29Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "a990a1db270799eef1e3db4381e3cc0f1f960973fc456eca98e5165c4162b9f4",
      "priority": "P1",
      "reason": "MemCode 不能只做cheap在线检索而省略ISQ学习。附录B给64,935 DailyDialog训练对、T5-base VQ-VAE、10epochs、32GB GPU/90GB RAM；离线训练与LLM原子写入总成本仍高/不确定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3.2 Stream-Based Atomic Abstraction Framework",
        "4.1 Experimental Setting",
        "Appendix B Implementation Settings"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19472v1",
      "source_sha256": "5d81a42b070cb4c28a0f2a65ab6e4f0fffc23f8aba12ba5690ecf9739ce7dc1f",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19472v1",
        "basis": "原文4GB runtime GPU不覆盖离线训练；不把随机codebook替代可学习量化。",
        "scale": "64935 pairs×10epochs；T5-base；32GB GPU/90GB RAM",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19458",
      "title": "PFedMA: Who Should Join Next as Federated Multi-Agent Orchestration Evolves?",
      "published": "2026-10-14T02:42:41Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e79d9b89a463e80753c5880e89720629d5f1e67a3b071c47e4baa0e878f782f4",
      "priority": "P1",
      "reason": "PFedMA 的orchestrator经PPO训练，privacy规则还要双向跨client演化；不能只实现agent selection打分。附录E.3为batch64、最多220步，闭环跨客户/安全评测完整预算仍材料性不确定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3.2.3 Composite Reward for Multi-Round Orchestration",
        "E.3 Training Configuration"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19458v1",
      "source_sha256": "d03afb27b1279124b6d6e96f5c44b31f1a2f154b0f17edbaf8afd173d1c65096",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19458v1",
        "basis": "原文训练与奖励链路已核；未假设多个GPU主机可同时使用。",
        "scale": "veRL PPO；batch64×最多220steps；多client规则演化",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19433",
      "title": "EvoProbe: Evaluation-Driven Benchmark Evolution for LLM Agents",
      "published": "2026-10-14T02:27:06Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "3ee5cea1f4746e6de4c6d34aac32b763456b7fbf38ef8cbf4a349f6f8bf5ee51",
      "priority": "P1",
      "reason": "EvoProbe 的新task生成、可执行subtask验证、多judge诊断和反馈再生均在核心闭环中。66个网络安全任务、八模型两轮并需人工quality review，完整成本与任务基础设施仍无法廉价界定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract two-round protocol",
        "Limitations"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19433v1",
      "source_sha256": "2e111cc4e5acf9b2715e178fec01d54b2f4cc0067532728ebf488a11330f4598",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19433v1",
        "basis": "原文明确额外生成/验证/人工成本；本轮仅审查，不执行安全任务。",
        "scale": "66 tasks×8 models；两轮任务演化；多judge+人工复核",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19425",
      "title": "Evaluating Agentic Capabilities Before Post-Training: An End-to-End Approach",
      "published": "2026-10-14T02:21:09Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "c6f4ad71725ea3fba05615357ef41c07a34adc67b2a92fb986d327173da5ac1d",
      "priority": "P2",
      "reason": "ASAP 是completion-interface base model的end-to-end评测scaffold，不是新post-training算法。只保留interactive diagnostic入口参考，不声称base checkpoint已做能力训练或接入执行器。",
      "locations": [
        "abstract scaffold adaptation",
        "1 Introduction"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19425v1",
      "source_sha256": "990ebf6bf3f8af1d2e2345bc81de7596fdb8caf7649230e95f927e92dbe9d3ec",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.19423",
      "title": "Beyond the Trial-and-Error Loop: Hybrid Projection and Automated Tuning for Distributed Training",
      "published": "2026-10-14T02:15:28Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f4c9f4250f020003e5d788d06834c75f79f8ba3073657a1c78435391e15b916e",
      "priority": "P1",
      "reason": "Primus Projection 具有CPU simulation路径，但原论文的tuning收益依赖真实GPU微基准校准与多节点验证；671B/1024GPU、235B/256GPU均不在预算。只跑未校准CPU仿真不能宣称复现其吞吐预测或agent tuning效果。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "Sub-node benchmarking: downscale, measure, upscale.",
        "abstract validation scale"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19423v1",
      "source_sha256": "456722bd9107961b40735e3288a6c48e4388c07888b66d24542779b73a27abe8",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19423v1",
        "basis": "论文生产验证规模与校准合同；本轮不做单卡昂贵probe以外推集群。",
        "scale": "1到1024 GPU的downscale/measure/upscale；大型dense/MoE",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.19422",
      "title": "AgentSafeEvo: Adversarial Co-Evolution for LLM Agent Safety",
      "published": "2026-10-14T02:13:29Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "9031b972252323ac35bba28dae4ede15bef3265efe792ea8efecb4a1788c02d2",
      "priority": "P1",
      "reason": "AgentSafeEvo 必须执行attacker/defender RL共同演化与learnability weighting。附录D.1每role4GPU、120steps、64batch和16Kresponse，成本过高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract co-evolution",
        "D.1 Training Setup"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19422v1",
      "source_sha256": "d962e45f213648c7e272ce7400b1595a5280a8d01bbd9c1573dfe05238848173",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19422v1",
        "basis": "原文训练配置；不以固定攻击词库取代可学习attacker。",
        "scale": "4 GPUs/role；120steps；batch64；response16384",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.19399",
      "title": "KoAGE: Do Large Language Models Know, Learn, and Coin Korean Neologisms?",
      "published": "2026-10-14T01:49:19Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "c804b0c1700a9ce44bc79a848200f3cd36d1e63a843e1825b8211ca32f61e254",
      "priority": "P2",
      "reason": "KoAGE/NeoTrace主要是Korean新词采集/temporal-cutoff和生成评测；multi-agent coinage仅为实验条件。归档语言/时间评测参考，不进入Agent/RSI能力复现。",
      "locations": [
        "abstract NeoTrace and KoAGE evaluation"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19399v1",
      "source_sha256": "e53bbf1c53a0176c2166713d73578f6765a3faa62270e9ac0cee25c606fa66e3",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.19397",
      "title": "SERO: Learning from Executable Oracle Continuations for Long-Horizon Agentic Reinforcement Learning",
      "published": "2026-10-14T01:46:40Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b2fb58a274e945908f3c674aae90fbcb6b58cd23c1c4d7d6a674d109a43ca8e3",
      "priority": "P1",
      "reason": "SERO 的可执行oracle续写、状态恢复和policy-reachability加权是中心，oracle应只用于training。附录E.1为每batch16×8rollouts、50步horizon、1.5B/7B及多optimizer，完整训练和oracle续写预算仍高/不确定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract executable oracle continuations",
        "E.1 ALFWorld Training Configuration"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19397v1",
      "source_sha256": "593de5ff89d56c9650db825c11f734f9c78c92eed14987557999c15925f4e21c",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19397v1",
        "basis": "已核训练-only oracle边界与真实环境恢复；不得把oracle计划输入被评估agent。",
        "scale": "16×8rollouts/batch；50turn；1.5B/7B；ALFWorld/WebShop",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19386",
      "title": "STRATA: Self-Maintaining Structured Experience Graphs for Self-Evolving Language Agents",
      "published": "2026-10-14T01:31:59Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "5b0cd313b8108f4af9852ed9d189295886c6032068fe9bd9f19fc956e7ab97ef",
      "priority": "P1",
      "reason": "STRATA 的typed graph、reflector、community induction、GC和contrast edges需真实在线任务流。4基准、2backbones、7baselines、三seed且WebArena812题，多模型完整闭环成本高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "Experimental Setup",
        "Where Single-Level Memory Loses Ground",
        "Component Ablation"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19386v1",
      "source_sha256": "9fe4d713aa14d8bd34b176c1fe82c053b4047f2c858a7f0426336d172dba5c05",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19386v1",
        "basis": "正文明确只收task outcome、无gold/evaluator内部，保留此边界；本轮不重跑多基准。",
        "scale": "AppWorld168/417、WebArena812、ALFWorld134、ScienceWorld23；3seeds",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.19382",
      "title": "ixi-sLM: On-Device sLM for AI Call Agent Service",
      "published": "2026-10-14T01:29:08Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1c42751258fcc40434b84069433ee00becf2ae1808650abfdd8e48839ad6b748",
      "priority": "P2",
      "reason": "ixi-sLM 是1.2B生产call service的域适配/词表裁剪/混合量化方案；small footprint不等于可复现。依赖真实私有call语料、私有服务质量口径及mobile NPU链，本轮作为生产系统参考归档，不把checkpoint/data不可用伪装成已测高成本。",
      "locations": [
        "abstract production service",
        "1 Introduction"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19382v1",
      "source_sha256": "05a445ee66343bc70a1ae2ffa1f11243a319925f91ff59eaca59f5d19ccbf7a5",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "private_service_dependency_archived"
    },
    {
      "id": "2610.19381",
      "title": "EvoWeave: Composing Reusable Agent Experience through Operation Level Graphs",
      "published": "2026-10-14T01:28:06Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f2e4073516c1a3cdd4187e60e0baed81c24700fd426041fd1915622bc09673dd",
      "priority": "P1",
      "reason": "EvoWeave 依靠真实operation contracts、directed connections及执行后的reflection，不是仅注册一张图。四基准/五设置/三backbones；附录F坦言不同成本项不能合并成统一总成本，完整预算仍不确定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3.2 Retrieval and Connections Guided by the Plan",
        "Appendix E Graph Statistics by Evaluation Setting",
        "Appendix F Token Usage and Runtime Metrics"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19381v1",
      "source_sha256": "acab1111cc1f95d67401ff0ba21ebf88ecd9d580e66112a94e7d8abeaa61577d",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19381v1",
        "basis": "原文说明reflection、validation、retries与offline construction未有统一口径；不将图大小当运行成本。",
        "scale": "4基准、5设置、3backbones；建图/检索/反思/执行",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19376",
      "title": "DeLoop: Repeated Actions as Self-Supervision for Long-Horizon Agents",
      "published": "2026-10-14T01:23:00Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "2eaa8e2aafc633a2ffc01293fa383b30031f9a0897e5b2c5157e7365aca61318",
      "priority": "P1",
      "reason": "DeLoop 从真实rollout重复动作挖negative、后续首个非loop动作作positive并DPO，不能用gold action冒充。8B/14B/32B、上千收集episode和跨环境对照的训练/采样总成本仍高/不确定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "5 Experiments",
        "Appendix D Setup Details"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19376v1",
      "source_sha256": "cefeb319f9e0c7250daad9591dfff9ca99f344bdac9d4698264833862fb7c6b8",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19376v1",
        "basis": "正文把gold-label/gold-action基线与无label本法分开，保留该区分。",
        "scale": "Qwen3-8B/14B/32B；主收集1500 episodes；DPO",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.19368",
      "title": "OS-Bench: Evaluating SWE Agents on Cross-OS Code Migration",
      "published": "2026-10-14T01:12:46Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "50373b035f54811c63adbecb277f8a3da8e57a61e4e3652c967927afa91c2d17",
      "priority": "P2",
      "reason": "OS-Bench 是Linux/Windows双边回归的迁移benchmark，官方测试属于scorer且须两OS验证。保留平台迁移测量协议，不把任务集或单OS通过当Agent能力实现。",
      "locations": [
        "abstract dual-OS protocol",
        "Models and harnesses."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19368v1",
      "source_sha256": "abc00d3fde44e22a60001757ace6c7818a28ca2a2f1aa6481886eec4e5466f6e",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.19345",
      "title": "When Is Smaller Greener? Right-Sizing Language Models for Agentic AI",
      "published": "2026-10-14T00:37:23Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "418b7bd2a0bb5c97bcd5117b913062ce72cacfcc41ca2972299ec0aed9cb3f32",
      "priority": "P1",
      "reason": "本文的贡献是按成功完成任务计量能耗/成本/时间的决策公式与Pareto筛选。接受CPU数值机制候选，可从公开参数算交叉长度；15模型GPU能耗pilot并非已复现，也不能将合成可靠率下的公式结果宣传为Agent能力。",
      "locations": [
        "abstract decision model",
        "2.2 How models are chosen today"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19345v1",
      "source_sha256": "ee673ad79a7688de9096e2aad087777077d77c3a6956a872e6f76a19d40e5c28",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "accepted",
      "resolution": "affordable_core_review_candidate",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19345v1",
        "basis": "公式核验无需下载模型或采集新GPU能耗；不重复原文多模型pilot。",
        "scale": "决策公式/Pareto筛选CPU；实证pilot为15个0.6B–27B模型",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      },
      "acceptance_boundary": "只进入人工审查候选；本轮未实现，未进入自动执行/evolve，未来实现须保留定义算法并另过数据隔离、预算和适用GPU证据门槛。"
    },
    {
      "id": "2610.19292",
      "title": "Whose turn is it? Context-guided model routing for stateful agents",
      "published": "2026-10-13T23:13:37Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e0313dde80c1dc5a300504518b7982753cde1f5a7246196bd39390f6729acc08",
      "priority": "P2",
      "reason": "这是stateful spreadsheet agent在user-turn边界路由、历史可见性和prompt cache开销的对照研究。保留非独立请求及switch不对称成本的分析，不把closed-form cost或routing标签作为已实现能力算子。",
      "locations": [
        "abstract routing study",
        "1 Introduction"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19292v1",
      "source_sha256": "839c6fd8911f7fbed9d4ff9e3fe12e533f311fc6d0624412472fd8b6e4e7608c",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.19223",
      "title": "Definition Grounding in Theory-Level Auto-formalization",
      "published": "2026-10-13T21:42:35Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "d9c65dc3b0c44a56a62a44bd7c2cfc61d57716c2d021e3e1339a8f022b15933b",
      "priority": "P2",
      "reason": "ReuseBench/definition grounding测形式定义对后续定理的可用性，主要是质量诊断与judging证据。保留独立downstream statement兼容性评价，不用gold后续形式化结果作为agent解题输入。",
      "locations": [
        "Model judges can help eliminate downstream incompatible candidates.",
        "Refactoring."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19223v1",
      "source_sha256": "bd4ea6f9b0aa9b1e7b21b2fa9a3f261461fa4ab427c6f6a1403343d37b1c5278",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.19222",
      "title": "REPORT-OPS: Hybrid Evidence-Guided Reporting with Human-Supervised Refinement for HPC",
      "published": "2026-10-13T21:40:24Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "cd0d86bb46178e07b92b8825fcc1d0f8bad54e1559020f8b529e64e0d0bf718d",
      "priority": "P2",
      "reason": "REPORT-OPS 是冻结HPC corpus上的68查询/人审policy改进case study；结论仅支持同环境复用，不支持不限域因果发现。归档evidence-gating/reference case，不声称已接入真实HPC运行或自动policy晋级。",
      "locations": [
        "abstract frozen-corpus evaluation"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19222v1",
      "source_sha256": "eff34633002d55f159700cbd2c5c8bf6a90b3d312f67b15677d3465b84dae303",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.19158",
      "title": "AgentCIScope: Grounded Contextual Integrity Evaluation for LLM Agents",
      "published": "2026-10-13T20:25:00Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "5c61508d56975840d29054865f2b6516efc337936f7d3c4aeea582d09a400348",
      "priority": "P2",
      "reason": "AgentCIScope 以936可执行workspace任务评估contextual integrity及信息流泄露，主要是基准。保留隐私评价参考，不将敏感属性taxonomy/gold disclosure约束当agent已学能力。",
      "locations": [
        "abstract unified platform and benchmark"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19158v1",
      "source_sha256": "c622ba6dd23121904628d00dd90ea040b0f2bfa0d8b7190cee28995dbe32b042",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.19005",
      "title": "CyberArena: Benchmarking AI Agents in Live Symmetric Attack-and-Defense Combat",
      "published": "2026-10-13T18:05:26Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "85c0d273112bad24ba03b92920bd2e752c73d4c0087fb8ed0fee76d5845bc46a",
      "priority": "P2",
      "reason": "CyberArena 是隔离sandbox的实时对称攻防benchmark，40挑战466matches。只作动态安全评测参考，不执行攻击、生成backdoor或把评测arena登记为正常能力改进。",
      "locations": [
        "3 CyberArena Benchmark",
        "abstract dynamic combat"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.19005v1",
      "source_sha256": "33b8749dbd6cb695d3f73d76e069d41d0bc8b3f83966f52e55b4d9d313319921",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.18813",
      "title": "Self-Evolving Agent Harness",
      "published": "2026-10-13T17:48:49Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "cf7aaa1e872b4f720c13ea0e67145b7172f357e3eae16366ce75dedc2c610d19",
      "priority": "P2",
      "reason": "Self-Evolving Agent Harness 是66页概念/理论/既有case综合，generalization/admission界限均有显式假设；未将一个新可执行operator及独立实验交付出来。作为理论架构参考，不能仅凭registry标签算集成。",
      "locations": [
        "abstract formalization and prior-work cases",
        "3.2 Repository and Runtime Logic",
        "3.3 Everything Is a Plugin"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18813v1",
      "source_sha256": "04985bc69dc81dc499406a1096302b8a065745f2d49d917f2fc3df970e18ad07",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.18808",
      "title": "Optimizing Team Behavior via Extremum-seeking Control in Multi-agent Reinforcement Learning",
      "published": "2026-10-13T17:46:27Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "bd12d7666c2dfba148bb4ede459da6f2ce323172b4ed742a73979b5632a99d31",
      "priority": "P2",
      "reason": "ADiCo 是cooperative MARL行为多样性与extremum-seeking双时标控制，不是LLM agent/RSI；按实际算法范围排除，不臆测GPU成本。",
      "locations": [
        "abstract two-timescale MARL controller"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18808v1",
      "source_sha256": "0587a3e302fb8fe879e4da05a3a1e2df5b125636455370e0c1d606d311446224",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.18760",
      "title": "Beyond the Verdict: Agentic Code Judging with Verifiable Feedback",
      "published": "2026-10-13T17:22:58Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "800a595822c4cb766b3b719f975542ad06fa07f4e9eeeed9806acc6a87256b40",
      "priority": "P1",
      "reason": "ARBITER 为35B MoE verifier经RL学习真正生成可执行counterexample；每candidate平均35.0K–46.9K输出tokens。训练及多次refinement成本高，不能用官方失败样例或judge标签替代反例生成。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "4.3 Efficient training through length control",
        "D.2 Output Tokens and Estimated Cost"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18760v1",
      "source_sha256": "5c29d490e653a1194d98144fa95ff316eb691aee2a6381156f5da1304846ff3f",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18760v1",
        "basis": "原文相对更短不等于绝对廉价，完整训练规模无法在单卡预算内廉价核定。",
        "scale": "35B MoE；35.0K–46.9K output tokens/candidate；RL",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.18684",
      "title": "Knowing the User Is Not Knowing What We Established: Grounding Awareness in Personal Agents",
      "published": "2026-10-13T16:39:59Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "77329a98da622898480411d362292ad840d0e21073e0f124a3592308943cd686",
      "priority": "P1",
      "reason": "Grounding Layer 明确分开用户事实与已接受、未确认、修订和限域约定；ledger、动作前gate及回查原始对话均保留。可低成本核验真实对话到状态提取及有限动作边界；不得直接喂入oracle +Status/+Policy或gold动作当实现。只接受机制候选，不声称全基准收益。",
      "locations": [
        "6 Grounding Layer",
        "H.1 Implementation",
        "9 Limitations"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18684v1",
      "source_sha256": "63264a51a5bdcfe5daaf201e97192e88df1a01326026a0d1c33cd20c44d86bc1",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "accepted",
      "resolution": "affordable_core_review_candidate",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18684v1",
        "basis": "正文的GL不改变模型或memory系统，检查只在后果动作触发；受控小规模真实模型调用保留核心，不需要RL训练。",
        "scale": "training-free ledger + 固定授权规则 + 后果动作边界检查",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      },
      "acceptance_boundary": "只进入人工审查候选；本轮未实现，未进入自动执行/evolve，未来实现须保留定义算法并另过数据隔离、预算和适用GPU证据门槛。"
    },
    {
      "id": "2610.18654",
      "title": "Kwai Summary Attention V2 Technical Report",
      "published": "2026-10-13T16:18:44Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "5f754651bc4f7caf8afb383f89c08fdf52357333aa70ac1608b6a887fb91298b",
      "priority": "P1",
      "reason": "KSA V2 以Summary→Meta压缩和summary-guided原block稀疏索引为核心，需实际continued pretraining与GPU kernel。已核原文1.9B/4B配置表和最高128K评测；完整改造训练与多臂验证成本仍无法在单卡预算下廉价核定，不沿用未在本条实际取得片段中核到的精确token阶段数。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "1 Introduction",
        "C-KSA",
        "S-KSA",
        "training stages"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18654v1",
      "source_sha256": "263d27029ae4438e464d219cfdd8d89c46efcd7bc4e2d341f81c4e93df7cb079",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18654v1",
        "basis": "官方正文§3训练/评测说明和1.9B/4B配置表已核；本条片段不足以廉价核定完整continued pretraining时间/费用，不运行简化attention或昂贵probe。",
        "scale": "1.9B/4B；最高128K评测；continued pretraining、C-KSA/S-KSA与GPU kernels",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.18652",
      "title": "AMemOps: Evaluating Memory Updates in Long Audio Streams",
      "published": "2026-10-13T16:16:19Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "a06df52aa614dc085b4e71339db4b472d981ba5ed66e464af9d2baa153e69dea",
      "priority": "P2",
      "reason": "AMemOps 评估小型recurrent audio memory在recording变化、旧属性与延迟下的更新，不是LLM Agent/RSI训练方案。原文训练与评估recording在source-spectrum统计上有共同生成信息，结论保持audio-memory诊断范围。",
      "locations": [
        "https://arxiv.org/html/2610.18652v1 §3 Data",
        "https://arxiv.org/html/2610.18652v1 §3 Architecture and training"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18652v1",
      "source_sha256": null,
      "source_hash_scope": "not_captured_metadata_hash_separate",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。",
        "本子代理通过官方HTML补读，未保存原始网页字节/hash；source_sha256为null，metadata_sha256单列，二者不混淆。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.18616",
      "title": "Knowledge Graphs as Files: An Agent Harness Where Tool Creation and Tool Use Co-Evolve",
      "published": "2026-10-13T16:00:50Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "98d4361d6f8d69e22879fd0aeedf450a61843f25b8ea51f366906b3fd22fa26d",
      "priority": "P1",
      "reason": "DKGI 让agent在完整KG文件上创建并使用真实工具，不是只给检索结果。附录C报告Full DKGI的Astra完整test约$3000且部分失败attempt缺token记录；最长300turn/1800秒与多图谱完整成本过高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3.1 Setup",
        "Appendix C Harness Details"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18616v1",
      "source_sha256": "433d8655ee643291af99d9de728f96cf4a6865194856a1e1422d41fddbbe6587",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18616v1",
        "basis": "原文token计费和执行上限；$3000还不包含缺usage失败项的完整可恢复账目，非本地付款。",
        "scale": "STaRK三图谱；最多300turn/1800秒；完整Astra test约$3000",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.18602",
      "title": "Auditing Future-Memory Use in Point-in-Time Language Agents",
      "published": "2026-10-13T15:56:58Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e0be60bcaccf8f2fd4e29fdbe5993d045dcb1b6ea5f95c9f34a50a4ce34de49a",
      "priority": "P2",
      "reason": "PIT研究将future fact accessibility、search allocation和final choice分开，主要结论是预注册方向未获支持。仅保留时间泄漏/因果识别诊断，不把提供未来事实的follow-up当无泄漏预测能力。",
      "locations": [
        "abstract independent future-memory probes",
        "Growth labels.",
        "B.2 Worked task: AMD and Amazon, FY2018"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18602v1",
      "source_sha256": "00657544ac793c9abf9d3a72842bedfe1ddfa7e6d95fe6255737c71d038ad11d",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.18522",
      "title": "Lost in Implicit Routing: How LLM Agents Fail to Load Relevant Skills at Scale",
      "published": "2026-10-13T15:12:12Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "297d59a63f759bde3bb8e242a4d244c88acabe9c4a993fc7064cccfff7863814",
      "priority": "P2",
      "reason": "本研究固定golden skills后加干扰，揭示implicit routing collapse；golden skills是评价锚，不可给被评估agent直接gold routing。15000轨迹/10B tokens是原研究规模；仅归档规模风险和显式selection设计参考。",
      "locations": [
        "Models and scale.",
        "3.3.2 Token cost of routing"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18522v1",
      "source_sha256": "b7572fddf5a906073d6cfdae26163a08fd400d07e1a4b090c5e6ad5d77faa01e",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.18515",
      "title": "A Global View of How LLMs Value Human Time",
      "published": "2026-10-13T15:08:38Z",
      "revision": 2,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e1de4583a7fc529d314ed22b804cb06bcc71708e2c5b64672fadd74873611531",
      "priority": "P2",
      "reason": "本文跨193国家测LLM对延误补偿的时间价值偏差，属于模型社会偏好实证，没有交互agent或RSI新方法；按内容范围排除，不因API费用小强行纳入复现。",
      "locations": [
        "abstract country-level valuation",
        "A.3 LLMs API Providers"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18515v2",
      "source_sha256": "9ff13d7a14542df86b36a1d90e53236a273c8d692b86a9668a0cb7a59264b656",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.18507",
      "title": "The Last Mile of Coding Agents for Long-Document Tasks: An Empirical Study on Clinical Trial Protocol Digitization",
      "published": "2026-10-13T15:03:31Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f2e40830cc17e12073492a307c18209fe16e3beebabc251f02af653eb9cb06b3",
      "priority": "P1",
      "reason": "TrialScribe 将domain skills和verification loop用于协议PDF数字化；真实27B推理/工具执行不能用gold提取字段替代。原文§4.2两A100、150长协议且多模型/metric，对单卡预算成本过高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "2.6 Main Results",
        "4.2 Results"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18507v1",
      "source_sha256": "a9103ca6ac1be0a305731e221a3d2c8f1a48c17e3de174a410c6ae2a39265359",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18507v1",
        "basis": "正文硬件及每例API费用不是本地测量；不运行PDF大集。",
        "scale": "Qwen3.8-27B；2×A100；150协议与多harness评测",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.18478",
      "title": "MemoryVLN: From Ego-Aware Memory Representation to Cross-Context Memory Reuse in Vision-Language Navigation",
      "published": "2026-10-13T14:40:04Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "0876d6eb95654aa602dab4990bd91b99a44d2fe65893192dc09e2a43f354214d",
      "priority": "P1",
      "reason": "MemoryVLN 的压缩、ego-aware多模态表示、层次memory与write-back需实际VLN训练。附录A.3给64GPU global batch64，另有VGGT/Habitat链路，成本过高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "A.3 Implementation Configuration",
        "A.4 Software Environment and Randomness"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18478v1",
      "source_sha256": "0fec804d1c6923af8aa02e4f4a0af89e427d85cd42d596830d70fdaae71f2d23",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18478v1",
        "basis": "原文仅一个training seed，不能将其报告为已做多seed稳定比较；本轮不训练。",
        "scale": "Qwen2.5-VL+VGGT；64 GPUs；Habitat",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.18427",
      "title": "Reliable Memory from Unstable Growth: Curating Self-Written Agent Playbooks with Randomized Experiments",
      "published": "2026-10-13T14:06:05Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "73dd2a230fe411970dc32fa841aa496887dfd5bd761ae818192f3f7d6fca4a8d",
      "priority": "P1",
      "reason": "RBC 用随机block inclusion估计lesson效果，并用fresh practice选择curated/whole/no-playbook，保留test隔离。原文19settings、每项3growth runs和三eval seeds，每score数百到数千真实episodes，完整成本高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "5 Method: Randomized Block Curation",
        "6 Experiments",
        "Appendix C Pseudocode"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18427v1",
      "source_sha256": "f60885c5faeec8aac9c656b0d229dbe4072e367d9da2b9d12a3c028b5951fcfb",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18427v1",
        "basis": "已查随机试验而非gold lesson utility；不以合成回归单测当agent能力提升。",
        "scale": "19settings；3growth runs；3evaluation seeds；百/千episodes每score",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.18408",
      "title": "Prompt-Injection Probes Learn the Conversation, Not Just the Injection",
      "published": "2026-10-13T13:51:06Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "793eb1d60e782f4fa2dfa53a8cd994774dd609cfc3c35235807dd4a134fa3e98",
      "priority": "P2",
      "reason": "补读取到的官方HTML显示标准probe可被位置/长度shortcut替代，matched negatives降低已见application假阳性，但跨application与threshold迁移仍有限。归档评测有效性/训练数据构造参考，不重跑六模型表征采集，也不把高AUROC或暴露标签当agent安全能力实现。",
      "locations": [
        "The Evaluation Can Be Passed Without the Model.",
        "7 Conclusion",
        "Limitations",
        "All Results."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18408v1",
      "source_sha256": "c25568c82ee5fbab1f83c03556bb1f1fa1a167e88069c0c606bca1058c7f6082",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only",
      "source_attempts": [
        {
          "url": "https://arxiv.org/html/2610.18408v1",
          "result": "unavailable",
          "error": "Internal Error returned by web.open; recovered later via alternate official HTTP transport"
        },
        {
          "url": "https://arxiv.org/pdf/2610.18408v1",
          "result": "unavailable",
          "error": "Internal Error returned by web.open; recovered later via alternate official HTTP transport"
        },
        {
          "url": "https://arxiv.org/html/2610.18408v1",
          "result": "retrieved",
          "method": "parent official HTTP snapshot; actual method/resource passages read by this reviewer",
          "source_sha256": "c25568c82ee5fbab1f83c03556bb1f1fa1a167e88069c0c606bca1058c7f6082"
        }
      ]
    },
    {
      "id": "2610.18378",
      "title": "Delegating Earth science model calibration to knowledge-guided LLM agents",
      "published": "2026-10-13T13:32:20Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "76e9896d018fb0fc948ec6c7e6569809c9ea0783f74a6bda2429b881dad65c08",
      "priority": "P1",
      "reason": "GCH 将agent声明冻结后交给确定性数值搜索，完整六模型calibration包含80–600 evaluations与三seed，并不等于只写配置。已查官方PDF表B3及附录C，模型执行/知识基础设施和完整搜索耗时仍材料性不确定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "PDF page 33 Table B3",
        "PDF page 48 Appendix C"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/pdf/2610.18378v1",
      "source_sha256": "fcc01773e07dcb464c789e1177e84674c9e351ca7b5e30a24ef6b67fa296c097",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/pdf/2610.18378v1",
        "basis": "实际PDF方法和表已取得；不把预存11向量重算当端到端agent校准。",
        "scale": "六Earth science模型；80–600evaluations；三seed retests",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.18356",
      "title": "Urbex: Agentic Spatial Grounding for City-Scale 3D Scenes",
      "published": "2026-10-13T13:23:24Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "91b6fbc59221b5296673895b9b79d2f98e1c9e96c801d81e8034bf1af1a3e300",
      "priority": "P1",
      "reason": "Urbex 的active spatial search政策要真实渲染/工具交互并RL训练。原文hardware段是16×H20训练；单H20约3.2s/query只代表在线推理，不含offline scene preprocessing。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "Hardware and timing protocol.",
        "Inference efficiency."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18356v1",
      "source_sha256": "8f81427f01ecea7941b46b9b2986ed5bd8fd50b72b343c01a4ad8f6abecc4ee9",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18356v1",
        "basis": "正文明确排除离线渲染/候选准备的timing边界，完整训练成本过高。",
        "scale": "16×H20 training；单卡inference；城市级3D scenes",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.18354",
      "title": "Choosing LLMs for Strategic Agent Interactions",
      "published": "2026-10-13T13:21:27Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "99b8da60ab20eb6860f7a248f2e7a4c67858512493fef23b7c99428f98a59ee2",
      "priority": "P2",
      "reason": "strategic selection通过五game/十LLM考查selector，轻量bandit仍未及best fixed model。作为非平稳对手下测量/负结果参考，不把选择模型的模拟曲线当新Agent能力。",
      "locations": [
        "abstract strategic-selection findings",
        "3.4 Evaluation Metrics"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18354v1",
      "source_sha256": "36f7da20798832d163ed216dfebd6ffefa894485c094715a274622ef28c378a6",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.18305",
      "title": "AffectWorld: Benchmarking Emotion-Driven Future World Prediction",
      "published": "2026-10-13T12:50:40Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "4fb2fa23f9a9ce7256d43a3d1833f7c6a72a80d1790eda1d58e7e82f8e083278",
      "priority": "P2",
      "reason": "AffectWorld 是情绪→动作→未来世界视频生成/预测评测，所称agent指视频中人类actor，不是工具LLM agent或harness RSI；归world-model评测范围。",
      "locations": [
        "abstract world-state benchmark"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18305v1",
      "source_sha256": "9ca3c3a7285208455c26ff4385fb17e86eea794bbf434c94693c86a680f5ad4d",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.18240",
      "title": "Beyond Answer Sharing: TrueMAD for Targeted Peer Engagement in Multi-Agent Debate",
      "published": "2026-10-13T12:08:38Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "bff14961b1c10b32ecd466b10dface30ab2509a6e9fae7f40c05cf95ad2f3297",
      "priority": "P1",
      "reason": "TrueMAD 的targeted peer advice必须由真实同伴模型生成，不能用gold critique。主设置30B-A3B、多agent、多round及三个reasoning基准；完整多臂预算经原文设置仍无法廉价界定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "4.1 Experimental Setup",
        "5 Conclusion"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18240v1",
      "source_sha256": "4537613d8cb50dd428c4044ca8ba9a86b0ded0a231ca199e61eda3c89a695c58",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18240v1",
        "basis": "training-free只说明无参数更新，不说明multi-agent推理全表便宜。",
        "scale": "Qwen3-30B-A3B；AIME/GPQA/TruthfulQA；MAD/DMAD对照",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.18229",
      "title": "When Does Opponent Modeling Pay Off? Spatial Inference Stabilises Reinforcement Learning under Partial Observability",
      "published": "2026-10-13T11:59:45Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "98e44e1452dd9a760493c47d20510419b6693d88418a972f07b74ee9af50d7dc",
      "priority": "P2",
      "reason": "该工作用hindsight目标标注训练spatial opponent heatmap并融入double DQN，是部分可观测foraging DRL，不是LLM agent或harness RSI；按摘要范围排除，无正文成本冒称。",
      "locations": [
        "arXiv metadata and abstract"
      ],
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.18229v1",
      "source_sha256": null,
      "source_hash_scope": "not_captured_metadata_hash_separate",
      "review_level": "abstract_scope_triage",
      "evidence_limitations": [
        "本条正文审查未完成；终态仅表示范围/证据例外归档，不表示实现或实验验证。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.18206",
      "title": "SWE-MISBEHAVIOR: Benchmarking and Mitigating Boundary Violations in Software Coding Agents",
      "published": "2026-10-13T11:40:10Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "d9b8b4f1eec2841da74ab15cf30be6f2e6efe4a8d5334b9a9edf6762949e3c30",
      "priority": "P2",
      "reason": "SWE-MISBEHAVIOR/CODE-KIT是coding边界违规taxonomy、基准和guideline/tool干预。只保留边界审查/评测参考；7.4K合成任务上的违规减少不等于已完成通用agent安全实现或部署。",
      "locations": [
        "abstract taxonomy and CODE-KIT",
        "3.1.1 Code Generation"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18206v1",
      "source_sha256": "b5fdf6115c05ff656b79155467d280b2cf461363c5e7a30c8553ef362b9b80d8",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.18186",
      "title": "Reinforcement Learning for Large-Scale Lifelong Multi-Agent Path Finding with Shared Spatial Fields",
      "published": "2026-10-13T11:25:41Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "443142f596971ba788977d0d5e859777227e8acfc5abb8f2112d110a81411823",
      "priority": "P2",
      "reason": "本法为10000-agent lifelong path finding训练共享CNN spatial fields，属于传统多agent导航RL而非语言agent/RSI；按正文执行/训练任务分流，不纳入本track。",
      "locations": [
        "abstract shared-field policy",
        "Execution cost."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18186v1",
      "source_sha256": "abf99fb0222cb5f3cf48601db4968adda36fb2279497973b8aca065f2efe9a40",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.18097",
      "title": "APT: Feedback-Driven Agentic Post-Training for Text-to-Image Generation",
      "published": "2026-10-13T10:14:23Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "5ec5d05c43308e58caa130db62c5c1cbc4787d39b8d1887ab04455e4f2fdd96d",
      "priority": "P1",
      "reason": "APT 的核心是agent生成评测/诊断/训练分布后实际RL更新图像生成模型。多base generator、多benchmark与迭代生成/训练成本高或材料性不确定；只做prompt选择不能算其复现。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract evaluation-diagnosis-training loop",
        "3.1 Experimental Setup",
        "3.3 Ablation Study"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18097v1",
      "source_sha256": "517b9ca79aeedc5577adc1487a37a2c8f73809b74cf1cdd9c03b7bb2d7d5e1f9",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18097v1",
        "basis": "正文方法和跨模型实验明确需要真实generator training；未启动图像生成或下载checkpoint。",
        "scale": "多个T2I backbone；evaluation/data agents与RL反复迭代",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.18092",
      "title": "LifeSearch: Benchmarking Deep Research Agents for Real-World Life and Productivity",
      "published": "2026-10-13T10:12:26Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1c7ec7b24a11b161362b007c25f380c97609adb9f58df92017d04c145a5d5d07",
      "priority": "P2",
      "reason": "LifeSearch 的203用户需求/4000rubrics和human ranking一致性用于评价deep research，不是新search算法。gold rubric/值/URLs须与agent环境隔离；归档公开评测参考，不冒称已实现DR能力。",
      "locations": [
        "abstract construction and grading",
        "1 Introduction"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18092v1",
      "source_sha256": "768d95f8944ddc0627a0c4ba4ff6917c7fe5eadc88cde32f9be114a2635b4cca",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.18083",
      "title": "AReaL-RSI: Adaptive Scaling and Coordination of Recursive Self-Improvement",
      "published": "2026-10-13T10:05:16Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "8109702905b1994e21ef109bb9a521e2777ab52f0836099ec77383d74c3ef5cf",
      "priority": "P1",
      "reason": "AReaL-RSI 需leader、conditional memory sharing及campaign后的证据复盘。原文KernelBench每researcher1小时、RSIBench每researcher2小时、四agent对照，多个任务/backbone成本高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "5.1 Experimental Setup"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18083v1",
      "source_sha256": "bc5e000da721888d1a31f29610de13d47a4ce6036bfb976ddadf5abc16862528",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18083v1",
        "basis": "论文资源分配是每agent而非总预算；不将单卡独占微基准推广成免费并发资源。",
        "scale": "4 researcher对照；1–2有效小时/agent/task；GPU kernel/科研训练",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.18052",
      "title": "Can Agents Forget When Asked? A Benchmark for Memory Unlearning in LLM Agents",
      "published": "2026-10-13T09:41:12Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1976bf5d017b8b1f6c8aa2fce6df0bf84af07de05f43ebf0d1a8567158d5f049",
      "priority": "P2",
      "reason": "Bemula 是用户请求删除记忆的paired-control benchmark，五suite区分存储、移除、保留、重学与攻击；结果是未解决而非新unlearning算法。归档负结果/协议，不能把删除gold字段fixture当真正遗忘。",
      "locations": [
        "abstract paired versions and five suites",
        "Appendix F Additional Results on Bemula -Lite"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18052v1",
      "source_sha256": "b77bc809cc42f47311e7a1f5585f2897f665621c35ffaa4721c7dcd773cac740",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.18007",
      "title": "Cost-Efficient Privacy Enforcement for Email-Generating LLM Agents under Strategic Information Requests",
      "published": "2026-10-13T09:06:12Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "739be3ec3dbff2fce10095e478ffcbf43972195a5e887b7e2d4c41ea698e7bbe",
      "priority": "P1",
      "reason": "本法需真实Gemma contract-conditioned重写再由PrivacyChecker检查；只保证remote-token节省，不代表total compute降低。已查defender objective/§5.1，local overhead换算和完整多策略email闭环成本仍材料性不确定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "Defender objective.",
        "5.1 Local Rewriting Reduces Remote Enforcement Cost"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18007v1",
      "source_sha256": "3c44a3f490e3a7cce9c69e11c53c1dde643d972e27e03e73488bc17eb77727f0",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18007v1",
        "basis": "论文明确total cost需另加本地token权重rho，不能把remote billing收益当本地总成本已知。",
        "scale": "local Gemma + remote action/checker；normal/strategic prompts",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.18004",
      "title": "SWEPatchTest: Evaluating Software Tests Beyond Reproduction",
      "published": "2026-10-13T09:04:57Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "642f0ad23c2c3fa5b54192a5c4f31f17e9b3b5c5dd74606e33098d7fdeba5165",
      "priority": "P2",
      "reason": "SWEPatchTest 用候选patch池区分reference-transition与真正patch discrimination，并以训练数据信号做SFT；主要保留JVR评价/数据筛选参考。本轮未复跑619任务/82579patch或SFT，不把reference patch/hidden tests给agent。",
      "locations": [
        "https://arxiv.org/html/2610.18004v1 §2",
        "https://arxiv.org/html/2610.18004v1 §3",
        "https://arxiv.org/html/2610.18004v1 Appendix E"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18004v1",
      "source_sha256": null,
      "source_hash_scope": "not_captured_metadata_hash_separate",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。",
        "本子代理通过官方HTML补读，未保存原始网页字节/hash；source_sha256为null，metadata_sha256单列，二者不混淆。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.17928",
      "title": "Learning How to Remember: Memory Modality Selection for GUI Agents",
      "published": "2026-10-13T08:15:52Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "4978138c54570429e65b46e66f7a5b760f5e3283788b7baa1642e09759e6838a",
      "priority": "P1",
      "reason": "SwitchMem 对同一retrieved demonstration实际比较text/image并学习step-level router；必须保留真实VLM hidden-state probe/两种memory。9B actor、三GUI基准、配对标注/训练和后续closed-loop预算仍高/不确定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "3.1 Study Setup",
        "6.3 Efficiency Analysis",
        "B.4 Efficiency Details"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17928v1",
      "source_sha256": "bb74e7c0ca567cf9babce074e2b006f8e0f4ec150744cbe3999f67a3cc940b54",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17928v1",
        "basis": "单步额外0.128–0.304s只衡量已训练router的online开销，不含数据与router构造。",
        "scale": "Qwen3.5-9B；三GUI基准；真实paired memory evaluation",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.17926",
      "title": "CineMem: Attention-Aligned Fine-grained Memory Retrieval for Multi-Shot Long Video Generation",
      "published": "2026-10-13T08:14:08Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "4becc53f4333a55556417d33bdd2c8da12dd23c2fee86c201880fb02909a41e2",
      "priority": "P2",
      "reason": "CineMem 的KV组织/查询级检索服务autoregressive multi-shot视频生成，不是交互Agent记忆或RSI；四H200效率实验归视频生成/基础模型范围，本track不重复实现。",
      "locations": [
        "abstract memory for video generation",
        "Appendix C Analysis of Efficiency"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17926v1",
      "source_sha256": "b0e50abc7e244bc0149fd7e1df5871635752aaca8dd70e7542c37edd6ebe9fd4",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.17869",
      "title": "ECMR: Evidence-Aware Deliberation Control for Multi-Agent LLM Reasoning",
      "published": "2026-10-13T07:30:30Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "45b3aeb5c6d0ed00fb545bf2bdb13a6a31f4e07303da977d20367fb388af9788",
      "priority": "P1",
      "reason": "ECMR 的state-conditioned controller需要真实支持独立性/矛盾/验证/role-action闭环；9000 held-out records和多策略、多domain完整成本高且总资源口径仍不确定。不能只回放gold state或把normalized audit cost当美元/运行时间。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "5.2 Main Held-Out Results",
        "7 Conclusion",
        "Interpretation."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17869v1",
      "source_sha256": "7ea6084c59949f72375274b76d35c30dae00883caa0f3895e22a68163a331aec",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17869v1",
        "basis": "正文指出ECMR使用最大token预算；审计score不同于token use，未推出本地绝对成本。",
        "scale": "9000records；MedQA/CaseHOLD/FinQA；多agent deliberation",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.17861",
      "title": "State-Aware Generation of Deployable Websites from Visual Prototypes and Requirements",
      "published": "2026-10-13T07:23:49Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "dd7240a4a055b815f7a529988960e510fe17955337f7284f65ece52cfc03986b",
      "priority": "P1",
      "reason": "VisStateCoder 需真实prototype grounding、图指导生成、render evidence和限域更新，不是静态Project State Graph。官方HTML§4.4五模型、每call最高16384输出、最多4round；baseline每task7200秒，完整网站生成/部署矩阵预算不确定且重。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "https://arxiv.org/html/2610.17861v1 §3.2–3.5",
        "https://arxiv.org/html/2610.17861v1 §4.4"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17861v1",
      "source_sha256": null,
      "source_hash_scope": "not_captured_metadata_hash_separate",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。",
        "本子代理通过官方HTML补读，未保存原始网页字节/hash；source_sha256为null，metadata_sha256单列，二者不混淆。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17861v1",
        "basis": "补查官方HTML的配置上限（不是实测统一耗时）；不启动网站生成或Docker矩阵。",
        "scale": "5模型×3任务级别；最多4update rounds；长生成/部署",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.17860",
      "title": "Socratic Skill Learning: Interleaving Teaching and Practice for Transferable Agent Skills",
      "published": "2026-10-13T07:23:34Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6fda2601d7b0c2e669ce67e2ba7ad2d4e54af4f341ae48e6d76fc95ee590eacd",
      "priority": "P1",
      "reason": "SSL 要实际Educator/Student交错教练、唯一Student操作环境并验证可迁移skill。35families/408cases、主22family/155case、66run/schedule及多模型pair，完整技能获取预算高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract SkillTransferBench",
        "5.4 RQ4: Selecting and focusing review around unresolved issues",
        "D.1 Overall and per-family comparisons"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17860v1",
      "source_sha256": "acf4a1349197d9c6893aa6603ec26aa2bd7d0fd38aad84a4927f8fb5d237fe97",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17860v1",
        "basis": "原文减少61.7% educator tokens是相对量，不等于整个技能实验可廉价复跑。",
        "scale": "35families/408cases；66runs每schedule；3轮获取",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.18927",
      "title": "Why Static Skill Inconsistencies Do Not Equate to Runtime Risks: A Behavioral Analysis of LLM Agent Execution",
      "published": "2026-10-13T07:01:53Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e4aefbb7a3290d27e9c782ececf0e1ae862b54062de6c8b7c4cd15c1fb6e7edd",
      "priority": "P2",
      "reason": "这是skill静态不一致与runtime行为差异的实证研究：静态缺陷不必然造成运行危害，修复也不总省成本。保留paired Original/Fixed/NoSkill评测参考，不将缺陷扫描等同于能力改善。",
      "locations": [
        "1. Introduction",
        "5.2. Efficiency"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.18927v1",
      "source_sha256": "7924805efbf1087e301cb2d84adcfb99bf7297d1fa7a589b2a3fa6a521ef0ceb",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.17837",
      "title": "ProtoArena: Can AI Agents Discover and Exploit Real-World Cross-Component Protocol Vulnerabilities?",
      "published": "2026-10-13T06:56:32Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "5fe8747c76e025e9e8a1c98d7859a8afa0609bb27604f0011e6c32b29e86c579",
      "priority": "P2",
      "reason": "ProtoArena 是80受控cross-component协议环境的安全基准，明确不测试production或披露未知漏洞。只归档scope、信息级别与hybrid verifier协议，不执行exploit或把最高amplification当普通Agent进步。",
      "locations": [
        "abstract authorized benchmark protocol"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17837v1",
      "source_sha256": "02a92d297ce827bf954004b336e95e65ad1911b7acb2a7f050c4832765f85d66",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.17832",
      "title": "Agentic-MatriBO: Capability-Compiled Agentic Bayesian Optimization for Materials and Chemical Design",
      "published": "2026-10-13T06:50:13Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "c05a923cf2ab1a53795aa77838f0def590d82b10e6a868523eea6d8b1b6e4b54",
      "priority": "P1",
      "reason": "Agentic-MatriBO 保留capability compiler、证据gate与提交后才可查oracle的边界；不是提前扫描全pool金标。四candidate pools×五seed×多baseline/agent调用，完整执行预算与依赖仍无法廉价核定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "4.1 Method Overview and Execution Flow",
        "A.3 Ablation and Model Comparisons"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17832v1",
      "source_sha256": "63f80400df20d9bb4ffc549dadbb0f6f0176dc2558f814cf7f032153fc2ae5fa",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17832v1",
        "basis": "已查算法只在legal ID提交后观察target；原文alias不是不可变model revision，总调用成本仍不确定。",
        "scale": "4 finite pools；5seeds；8baselines；LLM policy/reviewer",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.17816",
      "title": "Empowering Edge Agents: A Systematic Analysis of Edge-Cloud Collaboration in LLM Agent Execution",
      "published": "2026-10-13T06:40:22Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "57396df1f44536e53e0350a94f596d81a4b470d8619d0fd31ccca5802bdea6fe",
      "priority": "P2",
      "reason": "EC2-Bench比较edge/cloud长任务执行的performance/resource/privacy，贡献是benchmark和策略实证。归档协作模式与隐私测量参考，不将edge-only/cloud-only标签或静态路由注册为可执行evolve算子。",
      "locations": [
        "4 EC 2 -Bench : Benchmarking Edge–Cloud Collaboration for Long-Horizon Agent Execution",
        "A.2 Detailed Comparison with Prior Benchmarks"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17816v1",
      "source_sha256": "6835dfeb5f74e015b946c19557f41c17d524e6c7c8e3c1ac984b9abb51ff23f2",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.17763",
      "title": "Memory Echo in Single-Agent Long-Term Memory: Lifecycle Evaluation and Provenance-Guided Mitigation",
      "published": "2026-10-13T05:44:32Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "eccb60eb8df697d803c21546c98028694b923e6fb89f8abae5949b16a1625e0e",
      "priority": "P2",
      "reason": "MEchoBench §4.1明确以target gold answer提供verified correction，FPA还把authoritative evidence供给回答侧。仅接受oracle-assisted memory-lifecycle诊断结论，不作为无gold Agent能力复现、不进入evolve；0.920 accuracy不得用于能力晋级。",
      "locations": [
        "https://arxiv.org/html/2610.17763v1 §4.1 lines130–140",
        "§5.1 FPA",
        "6.7 Ablation Results: Sequential Additions"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17763v1",
      "source_sha256": "5a735cebcb551e583d31561802f6056172b4b09aa56ff7a746e9db2f3b0f1f80",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "oracle_assisted_diagnostic_only",
      "capability_evidence_eligible": false,
      "gold_isolation_requirement": "gold answers/plans、未来结果与oracle控制信号不得输入被评估agent；该诊断不得支持capability/evolve晋级。"
    },
    {
      "id": "2610.17757",
      "title": "Instructions Are Evidence, Not Objectives: Measuring Outcome Alignment of Agents",
      "published": "2026-10-13T05:39:34Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "5652c0a60c95e7a14fc9d1708ef788e5185fdc3f6247af126c88f6e4a3a56e41",
      "priority": "P2",
      "reason": "TiGanEval 比较partial/full requirements和consequence feedback，主要测outcome alignment；论文叙述不构成扩大用户授权的指令。保留需求可发现性/完整交付评价参考，不把full gold specification输入当新能力。",
      "locations": [
        "abstract benchmark protocol",
        "5.6 Reliability of the Measurement",
        "Scalability."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17757v1",
      "source_sha256": "f687b195edec94576453ed8458fb77477e7cb6485f8c4672905b88e298fdd970",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.17740",
      "title": "Coupled Harness--Model Evolution for Scientific Equation Discovery",
      "published": "2026-10-13T05:17:45Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "c850c9ba48952fa64e2d87e7604adbece34b3ec5608d9aef5732e4dd3cb4ef64",
      "priority": "P1",
      "reason": "Coupled Harness–Model Evolution 同时演化可执行工具并用GPT-5.5成功轨迹对Qwen controller做LoRA，不是只有工具库注册。完整工具搜索、教师轨迹和学生训练成本经现有配置仍材料性不确定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "Training setup.",
        "abstract coupled phases"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17740v1",
      "source_sha256": "8a805a470cb90cd3e9daf838256bcbe2ae91c875c72512f3f48efdda5d8925e7",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17740v1",
        "basis": "原文明确只筛有验证收益的真实teacher steps，不能用gold公式或固定工具选择代替policy学习。",
        "scale": "GPT-5.5教师；约3B active controller；tool evolution+LoRA",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.17724",
      "title": "Evidence-First Agentic Verification",
      "published": "2026-10-13T04:58:04Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "487a6f91f7511bd88ee24ea4e0dbed155da5f00ed4feb9be7da96ba560613a30",
      "priority": "P1",
      "reason": "Sauron 的bounded queries/faithfulness checks减少相对成本，但不是便宜全评测。原文100episodes单Sol配置$297.69，TRAIL为$841；仍有多backbone/故障注入/消融，完整复现过高。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "4 Experimental Setup",
        "5.2 The lead widens on process criteria",
        "5.4 Localizing real agent failures on TRAIL",
        "Cost and token accounting."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17724v1",
      "source_sha256": "e9f30e750cf20d66fbc428e600ea438ee04c197c77bdf9f64b5d05292d4a6ac1",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17724v1",
        "basis": "论文记录model-call charges，含clustering/gist/investigation/checking；不是本地付款。",
        "scale": "100episodes×2586criteria；8backbones；$297.69与$841示例",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.17723",
      "title": "Upstream Influence, Downstream Re-Anchoring in LLM Agent Chains",
      "published": "2026-10-13T04:57:25Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6e7b01da0cef4c5ee88a1c390548f5e3cda879021575ba8c2deb580d19e16d38",
      "priority": "P2",
      "reason": "该研究在32历史政治case上测upstream内容影响与下游re-anchoring，relay的clean accuracy/latency/token tradeoff未测。只归档信息架构风险分析，不实施操纵、传播政治内容或声称防御已验证。",
      "locations": [
        "abstract controlled chain study",
        "Limitations"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17723v1",
      "source_sha256": "e70e096e6aed50ac46e1380b0e49877000cbfd16011fec4c96ef8bb565c46007",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.17710",
      "title": "Adaptive-Group Process Rewards for GUI Agents",
      "published": "2026-10-13T04:37:18Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "8c910ac1da74f024e2479d7696f36ac2e5d9900becfce1782d999a36d14b08d5",
      "priority": "P1",
      "reason": "Adaptive-Group process reward需要VLM按动作和subgoal group评估，再用GRPO真实训练GUI agent。两process reward基准和OSWorld三个application组，完整reward scoring/rollout/训练成本仍材料性不确定。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract action/group rewards",
        "4.1 Experimental Setup"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17710v1",
      "source_sha256": "6cb46810ec1efa9d855d2ffe48e3340c7763b80c95850bbe0ce6fca2dbca9d56",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17710v1",
        "basis": "已查真实GUI训练设置，不把人工标注group/advantage tensor测试当能力实现。",
        "scale": "CUARewardBench/CUAStepBench；OSWorld Chrome/Impress/Calc；GRPO",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "materially_uncertain_after_read_only_checks"
      }
    },
    {
      "id": "2610.17694",
      "title": "HelpMeOut: A Neuropsychology-Grounded Benchmark for Error-Aware Monitoring of Household Task Execution",
      "published": "2026-10-13T04:20:13Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6d831a707cce9a76c2fd3ff5dfab98dfc7adb0de185a63591ef7d8c8c874adf6",
      "priority": "P2",
      "reason": "HelpMeOut 是20000 paired household execution episodes的error-monitoring benchmark；没有提出可证明提升的通用新Agent。保留错误类别和早诊测量，LLM原始日志不含gold错误位置给agent。",
      "locations": [
        "abstract paired benchmark",
        "API settings.",
        "Key observations."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17694v1",
      "source_sha256": "ece48e1aada3176589704f1cacec3fffc838f46d759b300e9a666e052e630f2d",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.17670",
      "title": "Cowork-Gym: Synthesizing Interactive Work Environments at Scale",
      "published": "2026-10-13T03:56:58Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "69e527df068d9de96669a0fbabce9269d2597d13ba1374c25702df64a6f759bb",
      "priority": "P1",
      "reason": "Cowork-Gym 的build-execute-revise环境构造和SFT/RL/OPD是中心。64.9K环境、14场景、9B/27B/35B-A3B训练规模明显超预算，不以少数合成环境fixture代替训练。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "abstract synthesis and post-training",
        "1 Introduction"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17670v1",
      "source_sha256": "3caca337c3fb26d96dfc8d6cf2f8be46c2c8035f81e7d657d51e93edb3818986",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17670v1",
        "basis": "论文明确大规模环境验证和多阶段训练；无需昂贵probe即可成本闭环。",
        "scale": "64.9K environments；14 scenarios；9B/27B/35B-A3B",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.17659",
      "title": "Can Agents Truly Forget? Explore Machine Unlearning for LLM Agents",
      "published": "2026-10-13T03:37:30Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "accedeeb435cd9dc75c50b14aeeb0d82e294b583287b6083a052b377015195de",
      "priority": "P2",
      "reason": "AUBench 在parameter/memory/tool三个通道检验TOFU知识去除，主要揭示system-level forgetting不足。仅保留离线web snapshot和provenance评价协议，不能用删掉gold答案字符串当agent unlearning实现。",
      "locations": [
        "abstract three-channel benchmark",
        "Combined condition and execution protocol."
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/html/2610.17659v1",
      "source_sha256": "a5dd7b2d5a42fce24b9653400e1733584162b9189881d1e4b25464e51a00461a",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_method_evaluation_scope_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "evaluation_or_analysis_reference_only"
    },
    {
      "id": "2610.17627",
      "title": "AiSAQ-in-Filtering for DRAM-free Filtered Information Retrieval",
      "published": "2026-10-13T02:50:02Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "7922aea2c5d6f446bf4a662c409a8c7f4731bfd8420ade31e37fb045097d5fba",
      "priority": "P2",
      "reason": "AiSAQ-in-Filtering 的LiPQ是SSD向量ANN过滤/标签联合量化算法，不是Agent或RSI机制；按摘要范围分流，不把“memory footprint”当Agent长期记忆。",
      "locations": [
        "arXiv metadata and abstract"
      ],
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17627v1",
      "source_sha256": null,
      "source_hash_scope": "not_captured_metadata_hash_separate",
      "review_level": "abstract_scope_triage",
      "evidence_limitations": [
        "本条正文审查未完成；终态仅表示范围/证据例外归档，不表示实现或实验验证。"
      ],
      "decision": "rejected_or_out_of_scope",
      "resolution": "out_of_scope_or_non_executable_synthesis"
    },
    {
      "id": "2610.17613",
      "title": "MAS-Gate: Unlocking Multi-Agent System Lifecycle via Node-Centric Quantification",
      "published": "2026-10-13T02:33:28Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "239063a5c3ea145f36b9cfa0999439f2c0e67090a1aba4cb8e11d8efda89b337",
      "priority": "P1",
      "reason": "MAS-Gate 用reference-answer likelihood变化提供训练utility、paired轨迹训练Editor；金标只可留在训练/评分侧。官方PDF p18是8×H100、batch256、group16、15Kprompt/8192response，完整训练过高，不能把gold量化fixture推广为inference算子。 成本过高，不复现（审查 DONE）；不进入自动重试、实现或evolve。",
      "locations": [
        "PDF page 7",
        "PDF page 18 Table 5"
      ],
      "fulltext_review_completed": true,
      "source_url": "https://arxiv.org/pdf/2610.17613v1",
      "source_sha256": "804cd525c4aa7830a1050c47a352966d4bec1268844d4652c124a31db7a00055",
      "source_hash_scope": "official_source_snapshot_sha256",
      "review_level": "official_bounded_resource_gate_review",
      "evidence_limitations": [
        "仅核查官方正文实际取得的方法/评测/资源片段，不是全文通读；未在本地实现、训练、调用付费模型或测量论文指标。"
      ],
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "cost_assessment": {
        "source_url": "https://arxiv.org/pdf/2610.17613v1",
        "basis": "官方PDF训练配置，未运行或把reference-answer打分暴露给被评估agent。",
        "scale": "8×H100；batch256；rollout group16；长上下文",
        "available_resource": "单卡 A100/A30 或本地 Mac",
        "cost_type": "reported"
      }
    },
    {
      "id": "2610.17436",
      "title": "SchemaWeaver: Accelerating On-Device Mobile Agentic AI via Schema-Guided Parallel LLM Decoding",
      "published": "2026-10-12T21:53:20Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "06ff285c134c982e0653dbf25350a490efbf1dc117f8580d1a005a3311f4da0e",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（SchemaWeaver: Accelerating On-Device Mobile Agentic AI via Schema-Guided Parallel LLM Decoding）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17436v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.17418",
      "title": "Why Agentic Workflows Fail: A Taxonomy Derived from Literature and Empirical Data",
      "published": "2026-10-12T21:32:14Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b2347625dc61c4d78b9ad099101104fef4a0c38d879442b650fefdcf8d0c881a",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Why Agentic Workflows Fail: A Taxonomy Derived from Literature and Empirical Data）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17418v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.17374",
      "title": "VulnBump: A Benchmark for Fixing Vulnerable Dependencies in Real Projects While Keeping the Build and Tests Green",
      "published": "2026-10-12T20:33:48Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "fca9295372019e789c24e23135f958c20afb3f0d26c4f6db36e9e7f037f515bd",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（VulnBump: A Benchmark for Fixing Vulnerable Dependencies in Real Projects While Keeping the Build and Tests Green）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17374v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.17365",
      "title": "When Goals Leak Across Relationships: How Cached Goal Representations Shape Cross-Client Decisions",
      "published": "2026-10-12T20:22:45Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "a1e3d793238af5b0e4e5a4754f08297e212302567217ceb560c380a48bd3ae30",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（When Goals Leak Across Relationships: How Cached Goal Representations Shape Cross-Client Decisions）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17365v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.17310",
      "title": "Beyond Successful Play: Diagnosing Opponent Modeling in Large Language Models",
      "published": "2026-10-12T19:08:43Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "37873aff1276f932411ffe19a0472ad8dad89754a87eff2484a4031a169ebc7a",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Beyond Successful Play: Diagnosing Opponent Modeling in Large Language Models）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17310v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.17224",
      "title": "CodeLift: Your Test Suite Is Secretly a Dense SWE Curriculum",
      "published": "2026-10-12T18:00:58Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1ec191e94c85ac1bfb925751b82c2fd1253056978474d585ae63fff3a3767026",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（CodeLift: Your Test Suite Is Secretly a Dense SWE Curriculum）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17224v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.17206",
      "title": "AutoMARS: Automatic and Scalable Multi-Agent Large Language Model Framework for Supply Chain Risk Assessment",
      "published": "2026-10-12T17:59:40Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "724cef03914ed7df278f99fa855323417de7c8d70dfef4ed0e5ced216532a452",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（AutoMARS: Automatic and Scalable Multi-Agent Large Language Model Framework for Supply Chain Risk Assessment）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17206v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.17197",
      "title": "Deleting Code Without Losing Knowledge: Regenerative Software for Python Libraries",
      "published": "2026-10-12T17:57:40Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "d1ef7c1ce81850639f3ebb04f6eea5b2e7b2b5a7657630f2970e179c2730e160",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Deleting Code Without Losing Knowledge: Regenerative Software for Python Libraries）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17197v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.17195",
      "title": "Welcome to the Age of Agent Ads: Using On-Policy Attacks to Steer Agentic Decisions",
      "published": "2026-10-12T17:57:12Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "37dcb3c5d4ab76f97b92c15a51e46fb88189c2c8b7abebc2b30c5334108d1098",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Welcome to the Age of Agent Ads: Using On-Policy Attacks to Steer Agentic Decisions）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17195v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.17183",
      "title": "ChemExplain: A Multimodal Benchmark for Explanatory Chemical Reasoning",
      "published": "2026-10-12T17:53:56Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "91c0b2c035ff7537f495b6a89d773f90972a36bb87096a9ff22413f6fbc06ae7",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（ChemExplain: A Multimodal Benchmark for Explanatory Chemical Reasoning）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17183v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.17130",
      "title": "EvoQuantLab: Training-Free Dual-Loop Evolution for Auditable Quantitative Research",
      "published": "2026-10-12T17:19:13Z",
      "revision": 2,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f788e9abb8c1c7d714fef7e3492a1f68bfce05ae3db5ab7f4adc31d6aad0ba80",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（EvoQuantLab: Training-Free Dual-Loop Evolution for Auditable Quantitative Research）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17130v2",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.17114",
      "title": "D-World: An Explicit World-Language-Action Model with Recursive Self-Improvement under Dual Control",
      "published": "2026-10-12T17:05:31Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "7edd41e6911c2111f9f937bc30787255422723f4a415f5c1e85735d312c9d947",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（D-World: An Explicit World-Language-Action Model with Recursive Self-Improvement under Dual Control）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17114v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.17043",
      "title": "Neos: Learning to Interleave Search and Visual Reasoning",
      "published": "2026-10-12T16:16:44Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "dfc5ea37ad00fe0b756d4ba2583e4d40b185ba58f8feba4be1f995a3877059b0",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Neos: Learning to Interleave Search and Visual Reasoning）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17043v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.17020",
      "title": "CWE-bench: Refocusing Agentic Security from Exploitation to Defense",
      "published": "2026-10-12T16:00:55Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "eaef1bf29444edc5967d451e265727703fc95f73f187bf2d0f6561ddc0954d92",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（CWE-bench: Refocusing Agentic Security from Exploitation to Defense）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.17020v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16998",
      "title": "Glasseek: A Transparent Recipe for Training and Trustworthy Evaluation of DeepSearch Agents",
      "published": "2026-10-12T15:48:45Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "ced136088448f0c96a79055802e322c3673f43bc75b0f008378447383c698c4f",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Glasseek: A Transparent Recipe for Training and Trustworthy Evaluation of DeepSearch Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16998v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16975",
      "title": "MADDIE: Multi-Agent Drug Discovery Engine",
      "published": "2026-10-12T15:37:37Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "236eee49924f4d65ff4d54187f2505ee1ddcfdf879bd7ad7d9449801d04dcf37",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（MADDIE: Multi-Agent Drug Discovery Engine）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16975v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16946",
      "title": "Social Priors: Descriptions of AI Users Shape How Language Models Trust and Cooperate",
      "published": "2026-10-12T15:20:01Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "5908de862d1e23d45c39862132d9d563b3776efe4078cef59b72fe88e8307b21",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Social Priors: Descriptions of AI Users Shape How Language Models Trust and Cooperate）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16946v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16919",
      "title": "ML Engineering Agents - A Data Management Perspective",
      "published": "2026-10-12T15:04:17Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "2c4c64cbadd85a72bd7aa9d403b11a99c9b44f12ece9900de5f5bb376c2d4229",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（ML Engineering Agents - A Data Management Perspective）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16919v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16911",
      "title": "Audit tools of the trade: Automated Auditor Elicitation Can Be Sensitive to the Context the Auditor Is Given",
      "published": "2026-10-12T14:57:55Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "5a3d16a21fcce69331f6333e0db0a50275780cb66031fe656c61fbedc0f2b0a8",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Audit tools of the trade: Automated Auditor Elicitation Can Be Sensitive to the Context the Auditor Is Given）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16911v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16895",
      "title": "Observe, Share, or Predict? Disentangling Partner Information in Embodied Multi-Agent Collaboration",
      "published": "2026-10-12T14:49:31Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "d42ec7f34faa1c5b8fe4efe1cd1ccd983f0352598d430ebf203ba46d6e61b89e",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Observe, Share, or Predict? Disentangling Partner Information in Embodied Multi-Agent Collaboration）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16895v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16870",
      "title": "SpecGuardBench: Can Coding Agents Construct Reliable Validation Harnesses from Natural Language Specifications?",
      "published": "2026-10-12T14:34:48Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "d6a758abef249a4efa965d68b7381d2151b2bf54b9f05b65476491377a931279",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（SpecGuardBench: Can Coding Agents Construct Reliable Validation Harnesses from Natural Language Specifications?）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16870v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16838",
      "title": "Same Operation, Different Permission: Evaluating Whether Guardrails Honor Authorization",
      "published": "2026-10-12T14:17:21Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "d039ae0a0c6afff302871c1466a38db6e4cd518c9b952771db9dddcecde5c822",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Same Operation, Different Permission: Evaluating Whether Guardrails Honor Authorization）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16838v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16725",
      "title": "Trace-to-Architecture: Capability Bottleneck Diagnosis for Harness Revision",
      "published": "2026-10-12T12:56:13Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "c69d82f103d19b41f35d8be1e82d1d274a54218ba7f0cc20599db6344eecc640",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Trace-to-Architecture: Capability Bottleneck Diagnosis for Harness Revision）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16725v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16661",
      "title": "Structural Baselines Matter for Agent Failure Detection and Attribution",
      "published": "2026-10-12T12:13:48Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "cb58e03ef8cf016d62adf38f88f68b498a163122ea42417c3b942b12876b4b30",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Structural Baselines Matter for Agent Failure Detection and Attribution）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16661v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16632",
      "title": "REEL: Learning to Retrieve, Evaluate, and Evolve Visual Memory for Any Long-Horizon Multi-Shot Video Generation",
      "published": "2026-10-12T11:54:14Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "5520a6d40abeb9654d635c1389edadf5720b8990d1c4ddaad5070d444831b759",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（REEL: Learning to Retrieve, Evaluate, and Evolve Visual Memory for Any Long-Horizon Multi-Shot Video Generation）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16632v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16625",
      "title": "LATCH: LLM-Assisted Timing Closure Harness for Post-Route FPGA Optimisation",
      "published": "2026-10-12T11:44:40Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "ff3c0d81fef3478f6a67416a088026b1d829f37dd06214cbbc22b7ddd530af3f",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（LATCH: LLM-Assisted Timing Closure Harness for Post-Route FPGA Optimisation）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16625v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16498",
      "title": "From Caricature to Collective Artifact: How Persona Construction Drives Polarization in Generative Social Simulations",
      "published": "2026-10-12T09:48:14Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "399477f238e501d78e816544b2c8cc13e4ed5da778b4bf9aea9f7cba9171a383",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（From Caricature to Collective Artifact: How Persona Construction Drives Polarization in Generative Social Simulations）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16498v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16450",
      "title": "CivWorldBench: Evaluating Language World Models in Structured Multi-Agent Sandbox Worlds",
      "published": "2026-10-12T09:16:29Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "9915d4210e540a1eca6ce748e2bc6255d90fb92f3f02b0ad1159cc0102610ec0",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（CivWorldBench: Evaluating Language World Models in Structured Multi-Agent Sandbox Worlds）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16450v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16426",
      "title": "AuthGateBench: Benchmarking LLMs as Authorization Gate",
      "published": "2026-10-12T08:55:01Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6264ec2d678d16bd8a7ee00da54b238f381178c72177b09fc25cfadbc3219f85",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（AuthGateBench: Benchmarking LLMs as Authorization Gate）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16426v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16412",
      "title": "SLS-Bench: A Benchmark for Incident Log Summarization with Synthetic Observability Data",
      "published": "2026-10-12T08:46:35Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "9462953d68f2fa50e1fa8c6259e2806f8b5b95d1cc7e3e80d9603a604f621464",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（SLS-Bench: A Benchmark for Incident Log Summarization with Synthetic Observability Data）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16412v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16402",
      "title": "EvoWeb: A Benchmark for Coding Agents on Evolving Web Applications",
      "published": "2026-10-12T08:42:42Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "8d87e760e1d47a46b9421175fef301f3aec01e56c4708d7b1bcae1cd35be714e",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（EvoWeb: A Benchmark for Coding Agents on Evolving Web Applications）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16402v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16333",
      "title": "TRAP: Tool-result Replay Attack for Probing Misbinding in LLM Agents",
      "published": "2026-10-12T07:51:02Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "c188335bf77bffc5c1c23e593a0c151b433496a49d348bdd255a27d3cf93ef54",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（TRAP: Tool-result Replay Attack for Probing Misbinding in LLM Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16333v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16315",
      "title": "ReTAC: Residual Test-Time Agentic Computation for Video Spatial Reasoning",
      "published": "2026-10-12T07:39:58Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f3810c5653bf0548df54b270650bbc7cf462bcba8f97c993bd4d8a3c3422b318",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（ReTAC: Residual Test-Time Agentic Computation for Video Spatial Reasoning）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16315v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16297",
      "title": "SLDBench v2: Active Scaling Law Discovery",
      "published": "2026-10-12T07:23:23Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "d28ed83d687bd9564b1e69ec4b8625fee12906ec4a20f9f44745177303ead011",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（SLDBench v2: Active Scaling Law Discovery）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16297v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16251",
      "title": "Avoiding What Your Teacher Already Knows: Entity Familiarity for Deep-Search Distillation",
      "published": "2026-10-12T06:25:07Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "bb93b78d9bfbda2bd3e16537e9e1655199966e98da96b32afdb7e7d6a188d81b",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Avoiding What Your Teacher Already Knows: Entity Familiarity for Deep-Search Distillation）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16251v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16231",
      "title": "Learning to GUIDE the Group: Self-Learning LLM Facilitation Across Multi-Party Tasks",
      "published": "2026-10-12T06:08:33Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "4d6f6a89448646b3287ecf50804b0dc46f19782ce93a187758a7b6c363ac1730",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Learning to GUIDE the Group: Self-Learning LLM Facilitation Across Multi-Party Tasks）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16231v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16224",
      "title": "Look Back to Move Forward: Learning to Revisit for Agentic Compositional Image Generation",
      "published": "2026-10-12T05:59:12Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "a0a87a375f969c6f012ec2f3cac783573783e52bd61aa90b608b59477e77617b",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Look Back to Move Forward: Learning to Revisit for Agentic Compositional Image Generation）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16224v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16175",
      "title": "ProCARE: Real-World Study Automation via Profile-Grounded Evidence Contracts",
      "published": "2026-10-12T04:50:24Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "32be13e80e4dba1f5ba82d42c14018f9906553aba9fb67a38e2cca39ba310263",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（ProCARE: Real-World Study Automation via Profile-Grounded Evidence Contracts）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16175v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16155",
      "title": "Uncovering Robustness Weaknesses in LLM Agents via Persona Evolution with Task-Integrity Constraints",
      "published": "2026-10-12T04:30:20Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "7226b2a5a1a3752a4fa3870c6a19201b0f92a372feaac6f5592ad0ca7193a549",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Uncovering Robustness Weaknesses in LLM Agents via Persona Evolution with Task-Integrity Constraints）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16155v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16140",
      "title": "Large Language Model Assisted Transportation Planning through a Differentiable User Equilibrium Layer",
      "published": "2026-10-12T03:59:40Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "4a5b2399fb533fd69c7b881d26613d33cce90506da7fe9639034d32ca70bfee7",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Large Language Model Assisted Transportation Planning through a Differentiable User Equilibrium Layer）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16140v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16137",
      "title": "Recall Begets Recall: Recurrent Memory Ranking for Personalized LLMs",
      "published": "2026-10-12T03:53:47Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "33a3518b48176df243c587291a93fbfef5267b098c7f30f82ace20003999fdc6",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Recall Begets Recall: Recurrent Memory Ranking for Personalized LLMs）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16137v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16117",
      "title": "Beyond Perfect Teammates: Multi-Agent Collaboration with Unreliable Colleagues",
      "published": "2026-10-12T03:31:54Z",
      "revision": 2,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "0884ea6a625d9cf841bc6d76b9cbda530f540451a6b1b8d12c120063c67b1253",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Beyond Perfect Teammates: Multi-Agent Collaboration with Unreliable Colleagues）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16117v2",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16078",
      "title": "Heterogeneous Agentic Coevolution",
      "published": "2026-10-12T02:34:01Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f247ac9020a56dfc5beb8ce355b3defc7a55b90157a6125b405a338222585016",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Heterogeneous Agentic Coevolution）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16078v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16070",
      "title": "Understanding and Reducing Misalignment in Persuasion Optimization",
      "published": "2026-10-12T02:19:21Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e8571e53391c6abd0941f85c1b7b46e9bc80433e27cea12f8826c0e0cbf9ed29",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Understanding and Reducing Misalignment in Persuasion Optimization）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16070v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16067",
      "title": "When Agent Evaluation Becomes Unreliable: An Empirical Study of Oracle Failures and Configuration-Dependent Measurement in an LLM Agent Benchmark",
      "published": "2026-10-12T02:11:43Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "9dfbe5c2324b8b304550135d8ddc17e430631fd68e1520193e9c0bf850658dac",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（When Agent Evaluation Becomes Unreliable: An Empirical Study of Oracle Failures and Configuration-Dependent Measurement in an LLM Agent Benchmark）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16067v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16027",
      "title": "Of Errors and Echoes: Responsibility-Aware Credit Assignment for Tool-Use Agents",
      "published": "2026-10-12T00:52:30Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "136cbc0ba677a13dc5cea0ba858f98598fbce055133b6f70c120e79227c848d0",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Of Errors and Echoes: Responsibility-Aware Credit Assignment for Tool-Use Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16027v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.16015",
      "title": "Computational Reflective Equilibrium for Agentic AI: A Rawlsian Alignment Scaffold for Governance",
      "published": "2026-10-12T00:31:30Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "8b28bac103d2dc7d91af0e443d2194d00a4c58b28cf306036dedd5e8fc8d5ba2",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-12，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Computational Reflective Equilibrium for Agentic AI: A Rawlsian Alignment Scaffold for Governance）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.16015v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15994",
      "title": "Large language model-generated patterns for epidemiological modeling: a multi-benchmark evaluation framework",
      "published": "2026-10-11T23:48:02Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "0eba7868a983f825a2e64a7b405c8bbbda7b032161de50bda88a310e88fb7746",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Large language model-generated patterns for epidemiological modeling: a multi-benchmark evaluation framework）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15994v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15991",
      "title": "Revocation Is Not Complete at Receipt: An Executable Outcome Contract for Distributed Agent-Tool Gateways",
      "published": "2026-10-11T23:41:18Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b2975d09b520a5755d4cd019287f4f13ae2dcd0587ef57c1634ec72c148dadd8",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Revocation Is Not Complete at Receipt: An Executable Outcome Contract for Distributed Agent-Tool Gateways）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15991v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15969",
      "title": "PopGen: Generating Semantically Diverse Populations for Zero-Shot Coordination",
      "published": "2026-10-11T22:53:33Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f35c68315583355a34cbf75e815568375cbf2e36e17be827452cb9d31b9753b9",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（PopGen: Generating Semantically Diverse Populations for Zero-Shot Coordination）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15969v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15964",
      "title": "Small but Mighty: Test Suites as Improvement Signals for Self-Evolving Agents",
      "published": "2026-10-11T22:36:54Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "d1b9889703e61f41b5da89d4072e57e85b0df37688d44914d3e35222d3bea5f0",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Small but Mighty: Test Suites as Improvement Signals for Self-Evolving Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15964v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15961",
      "title": "Unfair and Unsafe: Everyday Bias Triggers Emergent Misalignment",
      "published": "2026-10-11T22:31:46Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "de588adc1c74a73a262f7441deff779f704a2c55e10a1245273cbe20b5a1a469",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Unfair and Unsafe: Everyday Bias Triggers Emergent Misalignment）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15961v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15910",
      "title": "RIVO: Lightweight Adaptive Rerouting for Deep Research Agents",
      "published": "2026-10-11T20:45:18Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "ce9697caf6af192b9e32c460aec85dc2d54a13e270f4c5bedd76d8163b8a56f5",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（RIVO: Lightweight Adaptive Rerouting for Deep Research Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15910v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15882",
      "title": "Engram Injection: A Study of Temporal Validity in Parametric Language-Model Memory",
      "published": "2026-10-11T19:58:10Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "36464639e323723beec5a306a26519b2abca0aaa21d612d463e328c7e2b4024e",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Engram Injection: A Study of Temporal Validity in Parametric Language-Model Memory）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15882v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15766",
      "title": "SurgiCLAW: An Interpretable Atomic-Skill Agent for Precision Robotic Micromanipulation",
      "published": "2026-10-11T16:45:21Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "a3146676143f10ed5809595217ecae02988a8d87b45e01be71e2fc5e430b4298",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（SurgiCLAW: An Interpretable Atomic-Skill Agent for Precision Robotic Micromanipulation）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15766v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15765",
      "title": "A Decade in a Day: Synthesizing Agent-Native Interfaces from Human-Oriented Software Evolution",
      "published": "2026-10-11T16:43:43Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1a0f8fcb2b948c0c055d6ef70713577fbbfd6f6e210e922b29b5d3cc4b4b0761",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（A Decade in a Day: Synthesizing Agent-Native Interfaces from Human-Oriented Software Evolution）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15765v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15750",
      "title": "Google Earth AI Geospatial Reasoning Agent: Capabilities and a Systematic Evaluation",
      "published": "2026-10-11T16:31:29Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "370fca54a633f81fe3a2b2f0c934df9497c9db0b2a2cca44fa55549d31096854",
      "decision": "historical",
      "priority": "P1",
      "resolution": "historical_first_publication",
      "reason": "Google Earth AI GRA 的首发元数据为2026-10-11，早于本窗口。摘要是Gemini geospatial skills/前置规划及多benchmark评价；优先保留Google补发现身份，但本条仅历史日期/范围归档，不冒称10-13..15新论文全文审查、复现或实现。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15750v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15738",
      "title": "Visible but Not Avoided: State-Dependent Re-Proposal of Rejected Candidates in Catalog-Based LLM Optimization Loops",
      "published": "2026-10-11T16:18:11Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "3d1e8db759953a1ed5772c21a693eb1090f743506526207a97b8eea4a0d94cce",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Visible but Not Avoided: State-Dependent Re-Proposal of Rejected Candidates in Catalog-Based LLM Optimization Loops）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15738v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15708",
      "title": "SpecForge: Recursive Self-Improvement for Coding Agents via Behavioral Specifications",
      "published": "2026-10-11T15:46:22Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "689d50a56fa394f98aae21b8ae60d95400a103a9af311dadad0c3494713f8774",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（SpecForge: Recursive Self-Improvement for Coding Agents via Behavioral Specifications）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15708v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15691",
      "title": "FARL: A Fully Automated Pipeline of RL Interface and Solver Design",
      "published": "2026-10-11T15:22:18Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "34c6436e63a5c1107b14d275811ea75e746a3556cc048d8f6074169d99b33dde",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（FARL: A Fully Automated Pipeline of RL Interface and Solver Design）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15691v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15642",
      "title": "Recall What Is Missing: Composing Evidence from Long-Term Agent Memory",
      "published": "2026-10-11T14:18:11Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b03ac641db82a0762f396e69b0277683613da2b2c3460c8f28753fc1483a0881",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Recall What Is Missing: Composing Evidence from Long-Term Agent Memory）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15642v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15567",
      "title": "Reviewing Beyond the Change: Repository Size and Existing Code Across Agent and Human Pull Requests",
      "published": "2026-10-11T12:24:53Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "c65cba83056c1c5ddf4f40be5c386c49fbc1e4991765875bbd2d288534069392",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Reviewing Beyond the Change: Repository Size and Existing Code Across Agent and Human Pull Requests）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15567v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15529",
      "title": "Agent-Canary-Mas: End-to-End Evaluation of Protocol Risk in Long-Horizon Multi-Agent Systems",
      "published": "2026-10-11T11:02:00Z",
      "revision": 2,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "c2b68242be1f7b2216c655fe328eaec95bdebd31bf2176b28d0b700547fa8830",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Agent-Canary-Mas: End-to-End Evaluation of Protocol Risk in Long-Horizon Multi-Agent Systems）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15529v2",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15492",
      "title": "When Skills Remain But Access Fails: Factorized Skill Memory for Continual Skill Internalization",
      "published": "2026-10-11T09:46:32Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "8a36e9aef3739c6c0812fe0fec7846d3164d75d0d352791cbd330e459399f622",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（When Skills Remain But Access Fails: Factorized Skill Memory for Continual Skill Internalization）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15492v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15481",
      "title": "SkillFusion: Towards Reliable Multi-Skill Composition for LLM Agents",
      "published": "2026-10-11T09:30:07Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "7f720c857f5ea517d8111d34acd734a36c8480168364bc0f2d3ac20cdbcf5a04",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（SkillFusion: Towards Reliable Multi-Skill Composition for LLM Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15481v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15459",
      "title": "When Messages Go Wrong: Understanding Communication Faults in LLM-Based Multi-Agent Systems",
      "published": "2026-10-11T08:59:27Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e18be481d9e74d704a5bfe9d9d43370fa96df5163b8dbc83ed19ebff09951d37",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（When Messages Go Wrong: Understanding Communication Faults in LLM-Based Multi-Agent Systems）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15459v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15415",
      "title": "Sensitivity or Criterion? Measuring and Moving When Language Models Act",
      "published": "2026-10-11T07:36:38Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "a560da131cf2c66971d9cf7b2d2574b1f2b00f371c5ed4ace6388e0eeb45677d",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Sensitivity or Criterion? Measuring and Moving When Language Models Act）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15415v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15298",
      "title": "CHEAR: Confusion-Hierarchy Evidence-Aware Reasoning for Cost-Efficient Human Activity Recognition",
      "published": "2026-10-11T03:06:40Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "0031012ce5bb7be6ed92deb68bb951c5b56956fa57d92ae6d29768e7489af5ca",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（CHEAR: Confusion-Hierarchy Evidence-Aware Reasoning for Cost-Efficient Human Activity Recognition）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15298v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15292",
      "title": "OntoAR-HAR: Ontology-Guided Agentic RAG with Personalized Memory for Human Activity Recognition",
      "published": "2026-10-11T02:57:12Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1106e2d76883132c4cf191f5b23ed196aa6e42b77112caa41bbb410dce30b725",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（OntoAR-HAR: Ontology-Guided Agentic RAG with Personalized Memory for Human Activity Recognition）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15292v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15246",
      "title": "ObsSpec: Accelerating Agentic Reinforcement Learning with Observation Speculation",
      "published": "2026-10-11T00:40:33Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "9a2f97230631e14a0303734d329c8b00e0aae3eaf00a82efef26434adb97ecf0",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-11，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（ObsSpec: Accelerating Agentic Reinforcement Learning with Observation Speculation）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15246v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15228",
      "title": "Words to World: Benchmarking LLMs in Interactive Physical Environments",
      "published": "2026-10-10T23:40:54Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "07642aab93c64c0492bb90a3d74742a9a9ab4080c5f70b3747db4a1e77a7abe0",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Words to World: Benchmarking LLMs in Interactive Physical Environments）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15228v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15208",
      "title": "PANINI-DR: Controlled Deep Research through Generative Semantic Workspaces",
      "published": "2026-10-10T22:42:08Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "d717f7cef779757b2ad7cc49adbf6b2f8fb07e92d15bcd615983ed707d67a914",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（PANINI-DR: Controlled Deep Research through Generative Semantic Workspaces）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15208v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15150",
      "title": "Harmony Harness: Multi-Objective Self-Evolution of Agent Harnesses",
      "published": "2026-10-10T20:06:59Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1c80986b96f6602eb3b561ea5e7ff4b803b13fe41ae01100719752a94d866828",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Harmony Harness: Multi-Objective Self-Evolution of Agent Harnesses）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15150v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15137",
      "title": "When the Approved File Is Unchanged: Authority-Set Integrity for Deployment and Coding-Agent Approvals",
      "published": "2026-10-10T19:30:20Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "9da5d88d94ced002af8c38cb2a70ad4345f809274faf6ceca71b3903aca84f95",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（When the Approved File Is Unchanged: Authority-Set Integrity for Deployment and Coding-Agent Approvals）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15137v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.15084",
      "title": "Self-Evolving Visual Agent via Margin Reward",
      "published": "2026-10-10T17:33:56Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "211374ba59216b8fc7c20466802ccb6226463a3cb084a315be9b198bce1bdf7c",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Self-Evolving Visual Agent via Margin Reward）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.15084v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14966",
      "title": "SkillHorizon: Constructing Long-Horizon Agent Tasks through Skill Composition",
      "published": "2026-10-10T14:36:39Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "5a555a77746248f9d341279e81ad4cc8a6ad17be41c3774ecd45f56b7c7ee334",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（SkillHorizon: Constructing Long-Horizon Agent Tasks through Skill Composition）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14966v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14960",
      "title": "Humans vs. Machines: A Comparison With AI Agents in Software Reverse Engineering",
      "published": "2026-10-10T14:32:31Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "3429cc0e3d2d4c41c25b9c236f56070c5297436a9234fd81843173708bd9c0b0",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Humans vs. Machines: A Comparison With AI Agents in Software Reverse Engineering）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14960v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14956",
      "title": "MCPDisCo: Seesaw-Mitigated Tuning for MCP-Native Agents",
      "published": "2026-10-10T14:29:23Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "654332004834de309878971c5b018390cade071472e90288dcc27507b3ea53b3",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（MCPDisCo: Seesaw-Mitigated Tuning for MCP-Native Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14956v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14918",
      "title": "SkillForgeBench: Evaluating Agent Reliability Grounded in Agent Skill Specifications",
      "published": "2026-10-10T13:40:38Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "18d6b64ebe8e45979191dd2b2a38e2f07ecbcdaa307d9645204038184ca89b00",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（SkillForgeBench: Evaluating Agent Reliability Grounded in Agent Skill Specifications）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14918v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14915",
      "title": "WorkBorne: Discovering Agentic Tasks from Real-World Artifact Transitions",
      "published": "2026-10-10T13:37:47Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "40179751c0ebc2d130d8f912470d5148163bc172d65fe15dae15803f9c16a8a6",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（WorkBorne: Discovering Agentic Tasks from Real-World Artifact Transitions）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14915v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14900",
      "title": "TeachGrade: Training-Free Short Answer Scoring with Teacher Evidence",
      "published": "2026-10-10T13:22:16Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "d086a158952db2c51a2a7a832ec77c4882e9c9807241925ce2e5762430c74881",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（TeachGrade: Training-Free Short Answer Scoring with Teacher Evidence）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14900v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14877",
      "title": "Beyond Outcome Feedback: Trajectory-Quality-Guided Experience-based Learning",
      "published": "2026-10-10T12:50:09Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "60372c612495a2e2549f82ffbbe78373b494e9f4f9b2f5ab7925419ef453986b",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Beyond Outcome Feedback: Trajectory-Quality-Guided Experience-based Learning）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14877v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14855",
      "title": "PEAM: On-device Communication Agent for Personalized, Efficient and Autonomous Mobile Network Configuration",
      "published": "2026-10-10T12:21:05Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "7ac6ad9500731d22eeff25592fdebb21a514e7ace1dab3f02e794dab5f838089",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（PEAM: On-device Communication Agent for Personalized, Efficient and Autonomous Mobile Network Configuration）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14855v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14852",
      "title": "Getting Space Right: Agentic Rewards for Image Generation",
      "published": "2026-10-10T12:13:18Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "348a477bba9ac225eeee9147757ae5b4d194abd59a94b276964c824fa81988e8",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Getting Space Right: Agentic Rewards for Image Generation）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14852v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14823",
      "title": "AI-Native Product Management: A Vision",
      "published": "2026-10-10T11:39:13Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "ce51659eab880078f2b72fbe7ec0114ae224bcd3e9b5d99a690f826481a3b5cb",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（AI-Native Product Management: A Vision）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14823v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14808",
      "title": "Testuary: Agentic Test Generation with Zero-Shot Parallelism",
      "published": "2026-10-10T11:00:27Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e78093ecece5e07c41c356a3a509c80f98bb6b5e50a0ced1be5dd256d0cd894d",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Testuary: Agentic Test Generation with Zero-Shot Parallelism）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14808v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14790",
      "title": "Counterfactual Regret for Reflective Agents with Lean-Verified Guarantees",
      "published": "2026-10-10T10:22:39Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e93b193ef307cc3b9ebf4c28da4bbab9be3d172108f6683920e1779d57561687",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Counterfactual Regret for Reflective Agents with Lean-Verified Guarantees）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14790v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14777",
      "title": "LYNX: Analytical Program Synthesis via Grounding and Closed-Loop Refinement",
      "published": "2026-10-10T10:01:24Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "86a61f272744a0caa29df7cd01346f74e2a60fe2e113c863c1d27431eadbcf87",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（LYNX: Analytical Program Synthesis via Grounding and Closed-Loop Refinement）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14777v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14774",
      "title": "ProgAgent: Learning Tool Orchestration through Programmatic Training",
      "published": "2026-10-10T09:57:26Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b0daf39ac3195d23ad054b565611e1c9fa62b2da301386d9a00647cbd18e3ecd",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（ProgAgent: Learning Tool Orchestration through Programmatic Training）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14774v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14772",
      "title": "How Should a Camera-First AI Agent Work?",
      "published": "2026-10-10T09:53:25Z",
      "revision": 2,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f2c8727a0379fa7177025b027891efd61f7cedf252f7b47b427877fc5e6f01a1",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（How Should a Camera-First AI Agent Work?）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14772v2",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14768",
      "title": "Do Modifications Made by Coding Agents Make Subsequent Software Maintenance Harder?",
      "published": "2026-10-10T09:45:17Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1db4fc8a9da837d6842da8f52b94c8988b901a8253613f5c0899c1a6f7e757ae",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Do Modifications Made by Coding Agents Make Subsequent Software Maintenance Harder?）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14768v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14740",
      "title": "TileCas: Cascading Kernel Development Across Programming Abstractions",
      "published": "2026-10-10T09:00:51Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "64414812d5a2c393e163e211ecc37f78ad7c3018214fe4ac36daeaf471692949",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（TileCas: Cascading Kernel Development Across Programming Abstractions）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14740v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14736",
      "title": "Think in Strategies, Act with Tools: Traceable Agentic Optimization for Combinatorial Problems",
      "published": "2026-10-10T08:57:19Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "cd6825790dfcef6dac0223a6b4291ace9bb335d0b38d37d3ca4be935768820a9",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Think in Strategies, Act with Tools: Traceable Agentic Optimization for Combinatorial Problems）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14736v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14662",
      "title": "Does Failure Feedback Direct Agent Self-Modification?",
      "published": "2026-10-10T07:19:43Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "deb4c8487d2f9c7f2282bbaf17dd660f380f0573c528fa635e49001aabd0783d",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Does Failure Feedback Direct Agent Self-Modification?）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14662v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14636",
      "title": "PEARL: Process-Experience Aware Reinforcement Learning for Multi-Turn Agents",
      "published": "2026-10-10T06:35:58Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e6f84310a3301896c0c78b7801f7476c27471303eac2680c428aa22ad3377f11",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（PEARL: Process-Experience Aware Reinforcement Learning for Multi-Turn Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14636v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14629",
      "title": "Executable but Unfaithful: Mock Object, Stub, and Assertion Generation by LLMs and Coding Agents",
      "published": "2026-10-10T06:24:10Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "61f141a57ff568d914ff8082dbf6733d9b5411a9be32537940bb808384dccfbf",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Executable but Unfaithful: Mock Object, Stub, and Assertion Generation by LLMs and Coding Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14629v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14615",
      "title": "Limitations on Autoregressive Token Generation via Sequential Sampling",
      "published": "2026-10-10T05:59:16Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f9f824e789ed7fa5e77e17a92ff708c859b2993d2f1fc9ad9bffcece8012d67d",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Limitations on Autoregressive Token Generation via Sequential Sampling）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14615v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14576",
      "title": "Iterate on an LLM, Serve on an SLM",
      "published": "2026-10-10T04:51:24Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f4fb3886ca4a9d5ef195c64dd95b8f58378e7635ca2a6061ffecfc20d1bad069",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Iterate on an LLM, Serve on an SLM）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14576v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14570",
      "title": "PsyMemBehav: Psychology-Informed and Memory-Grounded Two-Level Identity-to-Action Reasoning for Social-Media Behavior Simulation",
      "published": "2026-10-10T04:38:44Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b59c9407518d458ee583b33a7c74161ebd2bc3a0e5450777d496c96caaa64ef5",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（PsyMemBehav: Psychology-Informed and Memory-Grounded Two-Level Identity-to-Action Reasoning for Social-Media Behavior Simulation）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14570v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14547",
      "title": "A Self-Defining Operator for Distributed Applications",
      "published": "2026-10-10T04:08:13Z",
      "revision": 2,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "c5db67cfa1e5639be9ebde4e5ac3ae6e33ed021a00cfa571f50c3ff13bf0ee65",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（A Self-Defining Operator for Distributed Applications）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14547v2",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14538",
      "title": "A Stateful Research Information Infrastructure for Scientific Discovery",
      "published": "2026-10-10T03:50:03Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b06f3af14f908f7a702c9d8ade4d19d851c52bad82fdb6776ec2da0d664be80a",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（A Stateful Research Information Infrastructure for Scientific Discovery）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14538v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14523",
      "title": "ScienceBuddy-PV: A Verifier-Governed Recursive Improvement Protocol for Panvascular Intervention Research",
      "published": "2026-10-10T03:35:28Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6b7ec8980753179f0cd182f04a9b46be6ac3345ad1a9564602370e66ff94dd0c",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（ScienceBuddy-PV: A Verifier-Governed Recursive Improvement Protocol for Panvascular Intervention Research）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14523v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14495",
      "title": "Reels-RSI: Program-Level Self-Improvement for Code-Rendered Explainer Videos",
      "published": "2026-10-10T02:46:44Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "aac122f3f741a18135c2375546b7942b7138514cf6c1b6b95c46ab353ad31719",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Reels-RSI: Program-Level Self-Improvement for Code-Rendered Explainer Videos）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14495v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14477",
      "title": "What Counts as the Same Task? Executable Contracts for Robustness Evaluation of Tool-Using Agents",
      "published": "2026-10-10T02:18:56Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1a18e22be5b829b420a93e12c3b0d0ebfa642dfd2ec0347b319ac67e5ed17e0e",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（What Counts as the Same Task? Executable Contracts for Robustness Evaluation of Tool-Using Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14477v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14465",
      "title": "Co-adaptive Rule Routing and Constraining for Structured Knowledge Retrieval",
      "published": "2026-10-10T01:45:48Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f0c80cebcf8721e074bd18141010b6ab5269b2fdb17525ccfc0e3e62b9f97fa2",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Co-adaptive Rule Routing and Constraining for Structured Knowledge Retrieval）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14465v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14462",
      "title": "Rethinking SZZ Search Strategies: Lessons from Claude Code",
      "published": "2026-10-10T01:40:04Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "334b7a0e9c564f794adad92e0662c7ac4091e396d0d780846de44e9598ed36d4",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Rethinking SZZ Search Strategies: Lessons from Claude Code）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14462v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14428",
      "title": "Frustratingly Easy Agentic RAG via Portable Skills: A Case Study in Customer Support",
      "published": "2026-10-10T00:49:01Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "d4700ad4d0853f0ffaf3b1fd7635e9aded8e32c7fee6fb032f0aa8c475a23832",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-10，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Frustratingly Easy Agentic RAG via Portable Skills: A Case Study in Customer Support）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14428v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14382",
      "title": "ANASSA: From High-Level Framework to Low-Level Architecture for Spatial Agentic AI",
      "published": "2026-10-09T23:03:33Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "550e375cd7047276f88f60c06e1c7ca198c5af1853bb954a42237eead189d4aa",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-09，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（ANASSA: From High-Level Framework to Low-Level Architecture for Spatial Agentic AI）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14382v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14381",
      "title": "ACED-Bench: Evaluating How LLM Agents Acquire and Act on Causal Evidence",
      "published": "2026-10-09T23:01:57Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "885461c31170525a9c949a4d107aab1c6b4ff6dd5f40b8df680f283653225941",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-09，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（ACED-Bench: Evaluating How LLM Agents Acquire and Act on Causal Evidence）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14381v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14216",
      "title": "Cortex: Measurement-Gated Evolution and Pruning of Coding-Agent Skills and Rules",
      "published": "2026-10-09T18:29:48Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e1f41193f17e36b6d07b1c00f7340af7689448a1ca659a78eefd8a981c8021e3",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-09，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Cortex: Measurement-Gated Evolution and Pruning of Coding-Agent Skills and Rules）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14216v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14186",
      "title": "Visualizing the World Model in an Agent's Value Function",
      "published": "2026-10-09T18:01:55Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "d0faddcb552e47ea5bbe27df3b3f04d5b54f4db144908978af23e204b5227099",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-09，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Visualizing the World Model in an Agent's Value Function）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14186v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14161",
      "title": "Agentic Evolution of SQL Test Generators",
      "published": "2026-10-09T17:42:26Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "52039fca70aa95e16bb65881f933cb6860b265c380e541dffab6bedda1e2471f",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-09，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Agentic Evolution of SQL Test Generators）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14161v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14157",
      "title": "OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools",
      "published": "2026-10-09T17:26:38Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6124af28be0ac3ddfe20dbe161c73fb88ea3cf2154088271e79a87efcf10a337",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-09，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14157v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14130",
      "title": "Conserved amount, negotiable placement. Prompting perturbs one architecture, an AI reviewer both",
      "published": "2026-10-09T12:08:23Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "c0afd12659d145919cabc9ca04a5b5884beace30d0334771f5c206d06007a143",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-09，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Conserved amount, negotiable placement. Prompting perturbs one architecture, an AI reviewer both）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14130v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.13399",
      "title": "RegulaTutor: A Theory-Informed Framework for Long-Term Personalized Learning",
      "published": "2026-10-09T11:11:44Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f70cedc44aef5913986827e550d295fc6693e82eb91d458c1c619240866224e8",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-09，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（RegulaTutor: A Theory-Informed Framework for Long-Term Personalized Learning）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.13399v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14096",
      "title": "SWEeper-Bench: Can Agents Discover Bugs in Interactive Software?",
      "published": "2026-10-08T19:05:11Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "db1ba08687fa6a5ab9fafcd48f1948727d3f3fbad44c8c5c245124238e62d1bf",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-08，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（SWEeper-Bench: Can Agents Discover Bugs in Interactive Software?）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14096v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.12300",
      "title": "Compact and Efficient Indexes for Learned Sparse Retrieval",
      "published": "2026-10-08T16:50:18Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e3418db06045f95a0abf2b4e9cb6334f0539c5f64e8ad36bd0264defaad93531",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-08，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Compact and Efficient Indexes for Learned Sparse Retrieval）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.12300v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.12243",
      "title": "NativeScope: Relation-Localized Retrieval over Native Topology with a Correct Anchor",
      "published": "2026-10-08T16:21:51Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1839d24274df29062b56fe99eae8f7013f63b8c13b2f94afb89c1a79b4e96ab5",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-08，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（NativeScope: Relation-Localized Retrieval over Native Topology with a Correct Anchor）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.12243v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.12090",
      "title": "Recompose and Refine Latent Reasoning Flows for Vision-Language-Action Models",
      "published": "2026-10-08T15:00:34Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f3738be6997230f2188617fa3b15edd8de99beb13b1cfd4361ea49f350582b03",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-08，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Recompose and Refine Latent Reasoning Flows for Vision-Language-Action Models）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.12090v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.11707",
      "title": "DeltaReplay: Task-Relative Memory Reuse for Mobile GUI Agents",
      "published": "2026-10-08T11:12:57Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e799a68864116517187ffeafebfae38194fca4beac94b67e17a8b3f13b5a963c",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-08，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（DeltaReplay: Task-Relative Memory Reuse for Mobile GUI Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.11707v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.11650",
      "title": "SkillContrast: Difference-Guided Text Selection for Agent Skill Reranking",
      "published": "2026-10-08T10:25:33Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "9c9a61cf93ff501320eae5847f0fd834ba0eb96db1b7139039f3fc1a8c1b1fb2",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-08，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（SkillContrast: Difference-Guided Text Selection for Agent Skill Reranking）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.11650v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.11617",
      "title": "TAM: Task-Aware Memory Distillation for Efficient Spatiotemporal Prediction",
      "published": "2026-10-08T09:58:23Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "83b908498c89e639b5e5993fa22a5584afa5f28409cd486edc4cf9e272ce7e1f",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-08，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（TAM: Task-Aware Memory Distillation for Efficient Spatiotemporal Prediction）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.11617v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.11171",
      "title": "VAMR: Multi-Question Agentic Reasoning for Efficient Long-Form Video Understanding",
      "published": "2026-10-08T03:26:54Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "47ef4bc223740823de333add0740b43e6641904f4f123967924319ff0e013ee6",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-08，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（VAMR: Multi-Question Agentic Reasoning for Efficient Long-Form Video Understanding）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.11171v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.10538",
      "title": "Never Look Back: Understanding Persistence in 3D Object Memory from Egocentric Videos",
      "published": "2026-10-07T17:59:41Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f1c642215e57a20a1a9263a94d14ee022628e8e4ae67c7568d4d6852421431a7",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-07，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Never Look Back: Understanding Persistence in 3D Object Memory from Egocentric Videos）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.10538v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.10164",
      "title": "UniSkill: Learning Actor-Aligned Skill Proposals for an Evolving Policy",
      "published": "2026-10-07T14:41:51Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f5396ad41bf1d08abc83b8fbe180bd55c76ee8b0b14790907a309ccd8039857c",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-07，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（UniSkill: Learning Actor-Aligned Skill Proposals for an Evolving Policy）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.10164v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.09146",
      "title": "Frozen Models, Evolving Expertise: Model-Agnostic Learning from Deployment Experience for Multimodal Medical AI",
      "published": "2026-10-06T21:41:27Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "8181fa22bc64984d094644b0f8acd041f1a7a074398831161b6437c347518959",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-06，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Frozen Models, Evolving Expertise: Model-Agnostic Learning from Deployment Experience for Multimodal Medical AI）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.09146v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.08553",
      "title": "DeltaTTT: Layerwise Optimization for Nonlinear Recurrent Memory",
      "published": "2026-10-06T15:40:06Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "7a765fbf0b5c9814f191d3acbfc84e30deeb2702cf5b472403226cc371bfc501",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-06，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（DeltaTTT: Layerwise Optimization for Nonlinear Recurrent Memory）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.08553v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.08138",
      "title": "Test-Time Agent Evolution for Long-Horizon Legal Reasoning",
      "published": "2026-10-06T10:53:25Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "597c7503861f8d9a1e74139f09cbd04c8482f86f7f9fa948dc2e627d249dcdcf",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-06，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Test-Time Agent Evolution for Long-Horizon Legal Reasoning）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.08138v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.07984",
      "title": "Decide Before You Look: Learning Which Retrieved Memories Deserve Pixels",
      "published": "2026-10-06T08:47:45Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "bf6170bf2e53287a88a0d00c2fbeecfd5f35a61ea2a469b34a4d09babe71b78f",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-06，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Decide Before You Look: Learning Which Retrieved Memories Deserve Pixels）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.07984v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.07725",
      "title": "PERSIST: Who-What-When Memory Across Sessions for Full-Duplex Spoken Dialogue",
      "published": "2026-10-06T04:21:11Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "2a4864575668b1df3926892256cb94db2e88056f84f9d9155d6e67c70a3c38a9",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-06，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（PERSIST: Who-What-When Memory Across Sessions for Full-Duplex Spoken Dialogue）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.07725v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.07403",
      "title": "Defense-in-Depth for LLMs: Evaluating Memory Gates Against Activation-Induced and Memory-Induced Sycophancy",
      "published": "2026-10-05T21:16:06Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "1434da42faf0e0f897c3e03b105fcdea78b7b6bab6c981f5d7e25c1dc5dcd601",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-05，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Defense-in-Depth for LLMs: Evaluating Memory Gates Against Activation-Induced and Memory-Induced Sycophancy）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.07403v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.07350",
      "title": "Trajectory-Retrieval Speculative Decoding: When Does a Model's Own History Help?",
      "published": "2026-10-05T20:19:28Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "72f33ef101af1572ab1d4539ee7ed796e3a279cdec5e87759766d7b763ac14b1",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-05，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Trajectory-Retrieval Speculative Decoding: When Does a Model's Own History Help?）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.07350v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.07309",
      "title": "The Right Memory in the Wrong Context: Verifying Retrieval Admissibility in Long-Term Agent Memory",
      "published": "2026-10-05T19:47:43Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "cdb3c4e3f1a285c00d9ebdaa1d76bb777e88d7fa94cc541576ed1ce462b8e611",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-05，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（The Right Memory in the Wrong Context: Verifying Retrieval Admissibility in Long-Term Agent Memory）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.07309v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.07184",
      "title": "Learning Scientific Exploration from Human Research Decision Trajectories",
      "published": "2026-10-05T18:05:12Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "3c56c8ac148bd926204afec20506e15e1dbe9dcd0f54b079c1bce51010f25cdd",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-05，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Learning Scientific Exploration from Human Research Decision Trajectories）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.07184v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.06672",
      "title": "VideoTapestry: Query-Adaptive Memory Refinement for Multi-Agent Long-Video Understanding",
      "published": "2026-10-05T16:44:25Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6a4f7a9830fdcf29949f922feb5f1500b61ee0e8bdfb0fe05fa73c876a18d98c",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-05，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（VideoTapestry: Query-Adaptive Memory Refinement for Multi-Agent Long-Video Understanding）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.06672v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.06361",
      "title": "Capability-Driven Self-Evolution of Agent Memory",
      "published": "2026-10-05T14:00:21Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "09a5c9f674da958ad5852c1d4207f514ca596616a2a53f617e2dd9c70f32b393",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-05，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Capability-Driven Self-Evolution of Agent Memory）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.06361v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.07075",
      "title": "CuratorMAS: Automating Dataset Curation via Multi-Agent Orchestration",
      "published": "2026-10-05T09:19:13Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "904dd1ee3b124d3a95c13d45c1632b827e0b2a8f97b4678122e6727ec914c2a4",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-05，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（CuratorMAS: Automating Dataset Curation via Multi-Agent Orchestration）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.07075v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.05739",
      "title": "HLA-WM: Hybrid Linear Attention for Long-Horizon Video World Models",
      "published": "2026-10-05T03:44:00Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "a1f6ccafe579274f8f8b76fc276df52905420ef4243a951d13c01010be338466",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-05，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（HLA-WM: Hybrid Linear Attention for Long-Horizon Video World Models）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.05739v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.05700",
      "title": "Voltic: Distinguishing Volatility from Stochasticity in Recurrent Memory",
      "published": "2026-10-05T02:23:28Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "427a5a4964feaa141c43bf84319fe537c10d58a1660fc6d63f30b656bc83c441",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-05，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Voltic: Distinguishing Volatility from Stochasticity in Recurrent Memory）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.05700v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.05526",
      "title": "Have I Scene This Before? Spatially Grounded Conversational Memory for Complex Queries in Egocentric Assistants",
      "published": "2026-10-04T20:42:29Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "df9ae1eb75399e4273012c8da0f649518c9b3a95712f70342a5fbab7f3ec49ad",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-04，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Have I Scene This Before? Spatially Grounded Conversational Memory for Complex Queries in Egocentric Assistants）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.05526v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.05489",
      "title": "SkillGATE: Gate-Aware Monte Carlo Tree Search for Skill Retrieval",
      "published": "2026-10-04T19:57:50Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "0a14ad5956dc4c774aa1e7e287a556c715b4547c0ce6dfa3a082667e29ca1bdc",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-04，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（SkillGATE: Gate-Aware Monte Carlo Tree Search for Skill Retrieval）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.05489v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.05437",
      "title": "TeleTune: Evolving Agent Skills From Offline Telemetry",
      "published": "2026-10-04T18:19:33Z",
      "revision": 2,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "ec5fc123e56a5a976324f3bd339bba24f2a8fc06d62be59e8762f11489c55501",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-04，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（TeleTune: Evolving Agent Skills From Offline Telemetry）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.05437v2",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.05382",
      "title": "Harness-Search: Guiding Long-Horizon Search through Multi-Agent Coordination",
      "published": "2026-10-04T17:14:24Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "a8520dddbb0c43d6ad37955fc40c02e825807264d1ec0148a68a5b04196c8705",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-04，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Harness-Search: Guiding Long-Horizon Search through Multi-Agent Coordination）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.05382v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.05348",
      "title": "Search Engines Never Say No: How Frozen Agents React When the Retrieval Tool Refuses",
      "published": "2026-10-04T16:14:41Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6697a000acf12dccc4e0f14a9020e7e066e2e9922aea345d1f5940ad0fcafda3",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-04，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Search Engines Never Say No: How Frozen Agents React When the Retrieval Tool Refuses）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.05348v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.04864",
      "title": "Enhancing Long-Video VLM Embeddings with Query-Aware Streaming Latent Reasoning",
      "published": "2026-10-04T02:05:48Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "e1639d2827a5e535389482af983c1a8834b4253920699212ef2fc8246b50f9e2",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-04，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Enhancing Long-Video VLM Embeddings with Query-Aware Streaming Latent Reasoning）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.04864v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.04838",
      "title": "MemTrace: State-Consistent Memory for Long-Horizon Coding Agents",
      "published": "2026-10-04T00:45:37Z",
      "revision": 3,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "558db6568d0cb2ec2b9ff9e49e2aad4e2f96b73de40d6a3fb10628a8fea4786e",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-04，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（MemTrace: State-Consistent Memory for Long-Horizon Coding Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.04838v3",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.04438",
      "title": "RAGrasp: Geometry-Semantic Template Retrieval and Grasp Transfer",
      "published": "2026-10-03T10:59:13Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "2e689222e24416e780b2de1dff0de704ce8bbca3a984883c67ee4c5f70f10b22",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-03，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（RAGrasp: Geometry-Semantic Template Retrieval and Grasp Transfer）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.04438v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.04262",
      "title": "CADForge: Agentic Single-View CAD Reconstruction with Explicit Geometry Reasoning",
      "published": "2026-10-03T03:39:54Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "a2a58d1ad35c390b4019c706f10e6c1735b1e224276132615899dc08a8670f37",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-03，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（CADForge: Agentic Single-View CAD Reconstruction with Explicit Geometry Reasoning）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.04262v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.04195",
      "title": "MemLeak: Cross-User Semantic Leakage in Multi-Tenant AI Agent Memory",
      "published": "2026-10-03T01:16:05Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "dc4f652dbb003137aa7cb80ea75e96ae6d33894be362c35d066bcacc7a08c08e",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-03，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（MemLeak: Cross-User Semantic Leakage in Multi-Tenant AI Agent Memory）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.04195v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.04109",
      "title": "SEER: Self-Evolving Event Reasoning and Retrieval for Time Series Forecasting",
      "published": "2026-10-02T22:30:10Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6df297919a6fdc6d86c6ec1e9895962ceee7d091380b175d2179dc279660a9fc",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-02，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（SEER: Self-Evolving Event Reasoning and Retrieval for Time Series Forecasting）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.04109v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.04012",
      "title": "Beyond the Parameter Monolith: Reconstructive Memories, Executable Skills, and Residual Assembly for Language Models",
      "published": "2026-10-02T20:18:49Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "238c91f41ba35cf416bfcb9c5380a0f5816ce831a0d185293bb7328bbd904d24",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-02，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Beyond the Parameter Monolith: Reconstructive Memories, Executable Skills, and Residual Assembly for Language Models）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.04012v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.03934",
      "title": "SGAnalog: An End-to-End Circuit Benchmark from Open-Source Silicon Tapeouts",
      "published": "2026-10-02T18:46:53Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "df188b18a349dbdb0bc0e89b92aac6e4a4058acc60889df2377605e8732bfad1",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-02，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（SGAnalog: An End-to-End Circuit Benchmark from Open-Source Silicon Tapeouts）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.03934v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.03651",
      "title": "MRVQ: One Resident Index for Dimension- and Rate-Elastic Vector Search",
      "published": "2026-10-02T17:33:13Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "4e71692a55f1a46e8675be1f5832d2397f8cf2c61b1feafe444ed4e7759797c5",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-02，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（MRVQ: One Resident Index for Dimension- and Rate-Elastic Vector Search）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.03651v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.03564",
      "title": "Knowledge or Calculator? Decomposing the Skill Premium in Verifiable Financial Agent Workflows",
      "published": "2026-10-02T16:44:04Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "b3ca2bc780f5e23320329d3760cb02ab34f2e7e07921d71d1b68f5515848a0f8",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-02，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Knowledge or Calculator? Decomposing the Skill Premium in Verifiable Financial Agent Workflows）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.03564v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.03510",
      "title": "Weave Forcing: Compositional Memory Routing for Interactive Long Video Generation",
      "published": "2026-10-02T16:04:26Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f44cd9d463d3091ac6fc8eacf34cdd08c20032095cea7658e095bf32bae2a528",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-02，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Weave Forcing: Compositional Memory Routing for Interactive Long Video Generation）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.03510v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.02945",
      "title": "Continual Graph Memory for Mathematical Research Agents",
      "published": "2026-10-02T07:38:09Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "6f298332ad6fabf7e0c36f2060b645767c79bb1cdc6c4de028ded9924e6207d7",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-02，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Continual Graph Memory for Mathematical Research Agents）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.02945v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.02867",
      "title": "TACD: Distilling Efficient Text-to-Motion Models via Terminal Amplification Control",
      "published": "2026-10-02T06:06:36Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "7ffa7c2d1a5dc785f32fc2712355e77f77b48be62e57bb4bdeb833b1f34317d9",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-02，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（TACD: Distilling Efficient Text-to-Motion Models via Terminal Amplification Control）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.02867v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.02387",
      "title": "SOLO: Certified-Recall Metric Similarity Search with Scan-Only Sampled Inverted Lists",
      "published": "2026-10-01T19:09:03Z",
      "revision": 2,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "381e81ff261560d015106fc59b95d4c13eadda3ffdd6848c026a0419fa4fdd7a",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-01，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（SOLO: Certified-Recall Metric Similarity Search with Scan-Only Sampled Inverted Lists）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.02387v2",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.02200",
      "title": "VISTA: A Visual Harness for Reasoning in an Interactive World",
      "published": "2026-10-01T17:59:45Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "f4f507acc95a73ec70c94fc46aae94ad13f87fef8bc9ed56bf2a69c48b904de1",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-01，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（VISTA: A Visual Harness for Reasoning in an Interactive World）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.02200v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.01058",
      "title": "MOMAT: Mixture of Multiple Atlases for Low-Power Jailbreak Defense of Quantized LLMs",
      "published": "2026-10-01T04:55:41Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "32903d3a9ea7362624d5b781ffba2f0598e7ce606cc16a288c6b8227b8cf5a7b",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-10-01，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（MOMAT: Mixture of Multiple Atlases for Low-Power Jailbreak Defense of Quantized LLMs）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.01058v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.00366",
      "title": "What Should an Agent Remember? Disentangling Retention from Retrieval in Bounded-Memory Evaluation",
      "published": "2026-09-30T06:38:14Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "c267bca1b12ced8731c9587aa911f8aab015a3a6800aff7ac2907d66bd3ecffb",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-09-30，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（What Should an Agent Remember? Disentangling Retention from Retrieval in Bounded-Memory Evaluation）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.00366v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.14021",
      "title": "CoEvoke: Evoking Multi-Agent Collaboration for Trajectory Collection",
      "published": "2026-09-29T13:54:06Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "d7dfc8b57f94f933979f9c362f89c2f50ee243f081d3e67a62f73f651bdc590d",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-09-29，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（CoEvoke: Evoking Multi-Agent Collaboration for Trajectory Collection）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.14021v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.03758",
      "title": "Global Evaluation of AI and NWP Precipitation Forecasts During Atmospheric River Events",
      "published": "2026-09-26T22:16:27Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "88d20e6eb24501fdb9b15e8117b7e7aeba9dde771f368331746669192e29bc08",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-09-26，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（Global Evaluation of AI and NWP Precipitation Forecasts During Atmospheric River Events）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.03758v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    },
    {
      "id": "2610.12496",
      "title": "An Input-Dependent Readout Phase for Resonant Sequence Models",
      "published": "2026-09-26T15:11:18Z",
      "revision": 1,
      "abstract_review_completed": true,
      "full_article_read": false,
      "review_done": true,
      "scope_triage_completed": true,
      "automatic_retry": false,
      "implementation_status": "not_implemented_by_review",
      "metadata_sha256": "4c99d7118de47ad54d4ed47e8b207d11de5030bc0ab656468f814f5ef6b9d0d7",
      "decision": "historical",
      "priority": "P2",
      "resolution": "historical_first_publication",
      "reason": "输入元数据所载首发为2026-09-26，早于2026-10-13..15窗口；已核对该ID、题名、版本及摘要范围（An Input-Dependent Readout Phase for Resonant Sequence Models）。本轮只做身份/日期/范围补发现归档，未进行新论文全文审查，不启动复现或自动重试。",
      "review_level": "historical_metadata_abstract_scope_triage",
      "fulltext_review_completed": false,
      "source_url": "https://arxiv.org/abs/2610.12496v1",
      "source_sha256": null,
      "source_hash_scope": "metadata_snapshot_only_no_fulltext_hash",
      "locations": [
        ".cache/oct15-review-closure/agent-review-input.json: arxiv_id/title/published/revision/abstract"
      ],
      "evidence_limitations": [
        "本轮沿用检索输入的metadata snapshot；未重新抓取该历史论文正文，首发归档不表示全文核验。"
      ]
    }
  ]
}