{ "schema_version": 1, "reviewed_on": "2026-10-15", "new_papers": false, "read_scope": "Latest documented versions, not inferred chronological algorithm diffs.", "reviews": [ { "id": "2610.13935", "title": "Hide and Seek: HIDE Concepts to SEEK explanations for black box models", "published": "2026-10-09T17:56:59Z", "revision": 3, "priority": "P1", "decision": "reviewed_deferred", "resolution": "revision_review_complete_no_implementation_upgrade", "reason": "已读v3摘要、定义方法/评测及Appendix A成本分解：VLM检索concept真实图像、定位遮挡后查询黑盒输出差；查询成本随每概念图像数变化，encoder可复用不等于全流程免费。只完成当前版本审查，不虚构与缺失旧基线的逐项语义diff，不升级实现。", "review_level": "known_paper_latest_revision_bounded_review", "abstract_review_completed": true, "fulltext_review_completed": true, "full_article_read": false, "source_url": "https://arxiv.org/html/2610.13935v3", "source_sha256": "a627ecf4f343ee0236fd2b4352d6f665485c7261c056329646f7653de4e45579", "metadata_sha256": "390ada5e44f100c4a50fc6f007aef8a2658e1db23e81ccdc9e3614b33a18cb50", "locations": [ "method", "evaluation", "Appendix A Computational Cost Comparison" ], "review_done": true, "automatic_retry": false, "automatic_implementation_queue": false, "implementation_status": "existing_evidence_not_upgraded_by_review", "semantic_diff_claimed": false, "new_paper": false, "review_scope": "Current version bounded defining-method/resource/correction check; no baseline equivalence or whole-article reading claim." }, { "id": "2610.11896", "title": "Automated Assembly Instruction Generation from CAD Models Using Grounded Large Language Models: A Human-in-the-Loop Framework", "published": "2026-10-08T13:04:03Z", "revision": 2, "priority": "P2", "decision": "reviewed_deferred", "resolution": "source_exception_archived", "reason": "v2身份/日期及摘要已核，是CAD装配说明的人类在环LLM方案；官方HTML与PDF均不可取得，不能核定修订方法。不虚称阅读全文，保留真实失败并非阻塞归档；仅新原文证据时重开。", "review_level": "known_paper_latest_revision_bounded_review", "abstract_review_completed": true, "fulltext_review_completed": false, "full_article_read": false, "source_url": "https://arxiv.org/html/2610.11896v2", "source_sha256": null, "metadata_sha256": "677b616c4a06adb7772eccdb0dcaef07335acc779b763d3a1e38115d1a20b0c7", "locations": [ "arXiv API revision/published/updated and abstract", "failed official HTML and PDF retrieval" ], "review_done": true, "automatic_retry": false, "automatic_implementation_queue": false, "implementation_status": "existing_evidence_not_upgraded_by_review", "semantic_diff_claimed": false, "new_paper": false, "review_scope": "Current version bounded defining-method/resource/correction check; no baseline equivalence or whole-article reading claim.", "source_attempts": [ { "url": "https://arxiv.org/html/2610.11896v2", "result": "unavailable", "error": "HTTP Error 404: Not Found" }, { "url": "https://arxiv.org/pdf/2610.11896v2", "result": "unavailable", "error": "HTTP Error 404: Not Found" } ], "scope_triage_completed": true }, { "id": "2610.11152", "title": "Do LLMs Learn from Rewards in Context? : Rethinking the role of reward in In-Context RL", "published": "2026-10-08T03:15:00Z", "revision": 3, "priority": "P1", "decision": "reviewed_deferred", "resolution": "not_reproduced_high_cost", "reason": "v3用奖励翻转/随机/移除及轨迹打乱等对照，发现直接ICRL收益多由ICL上下文驱动；不能把论文结论推广到所有推理时RL。原文Computational Cost实报单A100上ScienceWorld十episode约5h、代码约3h，整套多模型对照成本高，不复现、审查 DONE。", "review_level": "known_paper_latest_revision_bounded_review", "abstract_review_completed": true, "fulltext_review_completed": true, "full_article_read": false, "source_url": "https://arxiv.org/html/2610.11152v3", "source_sha256": "7fa3955b02ecc5bc76723600316900a6f7c83e93435c046825443a002161ce11", "metadata_sha256": "fded8ddc8afa9e5026e08591ea42b287aa8bb0b799b18dbbf57b60a3f61b6f89", "locations": [ "abstract", "controlled experiments", "Computational Cost" ], "review_done": true, "automatic_retry": false, "automatic_implementation_queue": false, "implementation_status": "existing_evidence_not_upgraded_by_review", "semantic_diff_claimed": false, "new_paper": false, "review_scope": "Current version bounded defining-method/resource/correction check; no baseline equivalence or whole-article reading claim.", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.11152v3", "basis": "abstract; controlled experiments; Computational Cost", "scale": "单A100 80GB；十episode ScienceWorld约5h，HumanEval/MBPP约3h；六模型四benchmark", "available_resource": "单卡 A100 或 A30，或本地 Mac", "cost_type": "reported" } }, { "id": "2610.11030", "title": "NOMOS: Compiling Written Policies into Statically Verified Tool-Call Gates for LLM Agents", "published": "2026-10-08T00:23:22Z", "revision": 2, "priority": "P1", "decision": "reviewed_deferred", "resolution": "revision_review_complete_no_implementation_upgrade", "reason": "NOMOS v2以四阶段编译器生成静态可验证的工具调用gate，并与宽松、policy prompt、动态LLM gate对照；不把schema checker冒称完整policy编译器。正文明确banking benign utility被用于policy修订，不是held-out utility；原稿长agent/26B/70B judge全套不在本轮执行。只做修订审查，无新版能力验收。", "review_level": "known_paper_latest_revision_bounded_review", "abstract_review_completed": true, "fulltext_review_completed": true, "full_article_read": false, "source_url": "https://arxiv.org/html/2610.11030v2", "source_sha256": "f19c13498976d00b504302b241fa9d050b638959cee5c111f5d084c256e8a9b5", "metadata_sha256": "97f7b61771843446d400b5d98bfc32aa2e0109f5ff83c812dcfc808ae86ced2f", "locations": [ "four-pass policy compiler", "§4 evaluation", "banking benign utility limitation" ], "review_done": true, "automatic_retry": false, "automatic_implementation_queue": false, "implementation_status": "existing_evidence_not_upgraded_by_review", "semantic_diff_claimed": false, "new_paper": false, "review_scope": "Current version bounded defining-method/resource/correction check; no baseline equivalence or whole-article reading claim." }, { "id": "2610.08077", "title": "Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight", "published": "2026-10-06T10:06:47Z", "revision": 3, "priority": "P1", "decision": "reviewed_deferred", "resolution": "not_reproduced_high_cost", "reason": "SRD v3用事后经验监督行动前不可见轨迹的foresight，推理无需生成foresight；训练时可用hindsight不等于评测agent可读未来。Appendix C.2为4B/9B thinking-on、8 H200、每步128轨迹及十种agent任务，成本高，不复现、审查 DONE。", "review_level": "known_paper_latest_revision_bounded_review", "abstract_review_completed": true, "fulltext_review_completed": true, "full_article_read": false, "source_url": "https://arxiv.org/html/2610.08077v3", "source_sha256": "351f37a28c2daaffe91f0af0a653381d5c777ca71a6832d478a1a1ea37bbbd8a", "metadata_sha256": "704ee396cb0c5985bd40158232143f84d199b08275f3c5aed431d8861ecb2ad4", "locations": [ "prospective learning method", "Appendix C.2 Training Details" ], "review_done": true, "automatic_retry": false, "automatic_implementation_queue": false, "implementation_status": "existing_evidence_not_upgraded_by_review", "semantic_diff_claimed": false, "new_paper": false, "review_scope": "Current version bounded defining-method/resource/correction check; no baseline equivalence or whole-article reading claim.", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.08077v3", "basis": "prospective learning method; Appendix C.2 Training Details", "scale": "4B/9B、8 H200、128 on-policy trajectories/step、多任务", "available_resource": "单卡 A100 或 A30，或本地 Mac", "cost_type": "reported" } }, { "id": "2610.02626", "title": "Imagine the Future, Internalize the Gist: Efficient VLA Reasoning via Internalized Spatiotemporal Imagination", "published": "2026-10-02T00:31:51Z", "revision": 3, "priority": "P1", "decision": "reviewed_deferred", "resolution": "not_reproduced_high_cost", "reason": "IG-VLA v3先学习未来SigLIP状态，再将时空reasoning内化为gist；不是仅重复当前帧。§4.1原文训练reasoning用8 H100、gist用4 A6000，LIBERO 100K步和VLABench61K步。成本高，不复现、审查 DONE，单卡推理延迟结果不替代训练验收。", "review_level": "known_paper_latest_revision_bounded_review", "abstract_review_completed": true, "fulltext_review_completed": true, "full_article_read": false, "source_url": "https://arxiv.org/html/2610.02626v3", "source_sha256": "03c522eef9244cc0b9a16b7e1ffbb9fcdf48cd90a7565c34e0696eefb6aab77b", "metadata_sha256": "6fbe9d2ffed014bd58d30ea6258d81f183b1828c469f8de852130b5b2edf3d85", "locations": [ "latent spatiotemporal reasoning", "§4.1 implementation details" ], "review_done": true, "automatic_retry": false, "automatic_implementation_queue": false, "implementation_status": "existing_evidence_not_upgraded_by_review", "semantic_diff_claimed": false, "new_paper": false, "review_scope": "Current version bounded defining-method/resource/correction check; no baseline equivalence or whole-article reading claim.", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.02626v3", "basis": "latent spatiotemporal reasoning; §4.1 implementation details", "scale": "8 H100 + 4 A6000 training stages；LIBERO100K/VLABench61K updates", "available_resource": "单卡 A100 或 A30，或本地 Mac", "cost_type": "reported" } }, { "id": "2609.36484", "title": "The Teacher Is a Direction, Not a Destination: Extrapolating RL-Induced Representation Residuals in On-Policy Distillation", "published": "2026-09-29T01:44:39Z", "revision": 2, "priority": "P1", "decision": "reviewed_deferred", "resolution": "not_reproduced_high_cost", "reason": "RIDE v2逐层使用RL教师相对base的hidden-state残差构造外推回归目标，需额外冻结base前向；不将output-space概率外推等同hidden-space方法。当前版本方法与配置已核，整套teacher/base/student长rollout和多任务总成本无法廉价核定，不重训、审查 DONE。旧L1实现/回执保留，不因版本元数据变动宣布新版能力完成。", "review_level": "known_paper_latest_revision_bounded_review", "abstract_review_completed": true, "fulltext_review_completed": true, "full_article_read": false, "source_url": "https://arxiv.org/html/2609.36484v2", "source_sha256": "108d33567ecc517ff9c062abfb57660a804b75ab032b5f458395bc75468e7855", "metadata_sha256": "3bc897905d90239eb464c17279ea13be88708cee9f7fe231ff569c3c38e94bfa", "locations": [ "§4 Method", "Training and evaluation configuration", "Appendix G" ], "review_done": true, "automatic_retry": false, "automatic_implementation_queue": false, "implementation_status": "existing_evidence_not_upgraded_by_review", "semantic_diff_claimed": false, "new_paper": false, "review_scope": "Current version bounded defining-method/resource/correction check; no baseline equivalence or whole-article reading claim.", "cost_assessment": { "source_url": "https://arxiv.org/html/2609.36484v2", "basis": "§4 Method; Training and evaluation configuration; Appendix G", "scale": "teacher/base/student forward、逐层hidden-state目标与长rollout；完整成本未核定", "available_resource": "单卡 A100 或 A30，或本地 Mac", "cost_type": "materially_uncertain_after_read_only_checks" } }, { "id": "2609.04047", "title": "The Dice Roll Method: A Standardized Protocol for Repeated-Query Auditing of Large Language Model Brand Recommendations", "published": "2026-09-03T16:22:06Z", "revision": 2, "priority": "P1", "decision": "reviewed_deferred", "resolution": "correction_reviewed_fixed_tiers_withdrawn", "reason": "Dice Roll v2明确撤回v1由notebook合成数据覆盖和“Research”计数字段产生的固定次数tiers及若干不可复算表；改为先pilot再依方差分量求次数，不能继续引用固定5/10/15次建议。仓库没有该算法实现，保留旧审查快照并新增本勘误，不把旧接受判断当新版验证。", "review_level": "known_paper_latest_revision_bounded_review", "abstract_review_completed": true, "fulltext_review_completed": true, "full_article_read": false, "source_url": "https://arxiv.org/html/2609.04047v2", "source_sha256": "b408a60b1e07f12672f63d664e5f5a57b5c9ed2764c08a14b4fa925ac1493e9b", "metadata_sha256": "e326a732a8b4a0d1770e74feefb5fbd03003ff253a2877f7831d23ce9a1ed56f", "locations": [ "Correction (v2) full notice", "§6.2 external validation", "§8 conclusion" ], "review_done": true, "automatic_retry": false, "automatic_implementation_queue": false, "implementation_status": "existing_evidence_not_upgraded_by_review", "semantic_diff_claimed": false, "new_paper": false, "review_scope": "Current version bounded defining-method/resource/correction check; no baseline equivalence or whole-article reading claim." }, { "id": "2608.11669", "title": "Rubric Dropout: A Simple Way to Mitigate Reward Hacking in Rubric-as-Reward RL", "published": "2026-08-12T05:29:21Z", "revision": 2, "priority": "P1", "decision": "reviewed_deferred", "resolution": "not_reproduced_high_cost", "reason": "Rubric Dropout v2仍要求同prompt同step共享mask；还讨论相似criteria按block联合drop，至少保留三criteria及full-rubric评测。独立gold judge不是绝对真值，多数配置单run。1.7B–8B、16rollouts、600steps及双API judge完整多臂成本高，不复现、审查 DONE；不宣称旧adapter已覆盖v2 block方案。", "review_level": "known_paper_latest_revision_bounded_review", "abstract_review_completed": true, "fulltext_review_completed": true, "full_article_read": false, "source_url": "https://arxiv.org/html/2608.11669v2", "source_sha256": "693c6139bfd84cf85651c471ec2ae082690e327fc6c22741af8b624aa6661b93", "metadata_sha256": "0d997f6b79c8419a8de82f01373025a1d2b66ab87f002917865d76858df101f2", "locations": [ "§3.2–3.4 group-shared and block dropout", "§4.1 setup", "Appendix C limitations" ], "review_done": true, "automatic_retry": false, "automatic_implementation_queue": false, "implementation_status": "existing_evidence_not_upgraded_by_review", "semantic_diff_claimed": false, "new_paper": false, "review_scope": "Current version bounded defining-method/resource/correction check; no baseline equivalence or whole-article reading claim.", "cost_assessment": { "source_url": "https://arxiv.org/html/2608.11669v2", "basis": "§3.2–3.4 group-shared and block dropout; §4.1 setup; Appendix C limitations", "scale": "1.7B–8B、16rollouts、600updates、proxy+gold API judges、多臂", "available_resource": "单卡 A100 或 A30，或本地 Mac", "cost_type": "reported" } }, { "id": "2608.04962", "title": "SPARK: Learning from Verification in Speculative GRPO Rollouts", "published": "2026-08-05T15:32:27Z", "revision": 2, "priority": "P1", "decision": "reviewed_deferred", "resolution": "not_reproduced_high_cost", "reason": "v2由SpecRoll改为SPARK：同prompt group共享并每rollout清零的低秩logit correction，以top-16+tail的verifier反馈更新、跨rollout学习投影，原target verifier保持不变。与旧fast/slow SpecRoll不是仅名称变化；本地旧实现不算SPARK。21模型数据设置+GRPO训练全套成本高，不复现、审查 DONE；旧论文页已加版本警示。", "review_level": "known_paper_latest_revision_bounded_review", "abstract_review_completed": true, "fulltext_review_completed": true, "full_article_read": false, "source_url": "https://arxiv.org/html/2608.04962v2", "source_sha256": "0fa3996130da6ae00030b0c7dbbeeb51765799e0555c512eb4ceab84d77025af", "metadata_sha256": "0d5b10337734e4ebaedeb2e2d387923f3971428a8dbb08fb40bfa46422d36c0e", "locations": [ "§3.2–3.4", "Appendix A SPARK implementation", "§4 setup / limitations" ], "review_done": true, "automatic_retry": false, "automatic_implementation_queue": false, "implementation_status": "existing_evidence_not_upgraded_by_review", "semantic_diff_claimed": false, "new_paper": false, "review_scope": "Current version bounded defining-method/resource/correction check; no baseline equivalence or whole-article reading claim.", "cost_assessment": { "source_url": "https://arxiv.org/html/2608.04962v2", "basis": "§3.2–3.4; Appendix A SPARK implementation; §4 setup / limitations", "scale": "21 model–dataset settings、EAGLE3/DFlash草稿、target GRPO训练", "available_resource": "单卡 A100 或 A30，或本地 Mac", "cost_type": "reported" } }, { "id": "2604.12234", "title": "UniRec: Bridging the Expressive Gap between Generative and Discriminative Recommendation via Chain-of-Attribute", "published": "2026-04-14T03:13:50Z", "revision": 5, "priority": "P0", "decision": "reviewed_deferred", "resolution": "not_reproduced_high_cost", "reason": "UniRec v5已明确Shopee机构，含capacity-constrained SID、CoA、CDC以及RFT+DPO；§4.2生产A/B的PVCTR/订单/GMV证据通过，不能把离线token-hit和业务收益混写。训练依赖9天数十亿私有交互和三场景，忠实新版全套不复现、审查 DONE；旧CoA本地实验不自动升级为v5全方案。", "review_level": "known_paper_latest_revision_bounded_review", "abstract_review_completed": true, "fulltext_review_completed": true, "full_article_read": false, "source_url": "https://arxiv.org/html/2604.12234v5", "source_sha256": "4eca5d87f6761a3ce618e007a8b7cc08a67f50659b1f00ca3b8bbd95dfac2640", "metadata_sha256": "541c4fe0b0690b6aa29134d7302ba3c85cd88dbda696a13fb8169600194e55c4", "locations": [ "front affiliation", "§3.2–3.6", "§4.1 private dataset", "§4.2 production A/B" ], "review_done": true, "automatic_retry": false, "automatic_implementation_queue": false, "implementation_status": "existing_evidence_not_upgraded_by_review", "semantic_diff_claimed": false, "new_paper": false, "review_scope": "Current version bounded defining-method/resource/correction check; no baseline equivalence or whole-article reading claim.", "cost_assessment": { "source_url": "https://arxiv.org/html/2604.12234v5", "basis": "front affiliation; §3.2–3.6; §4.1 private dataset; §4.2 production A/B", "scale": "Shopee私有9天数十亿交互、3层SID×4000codebook、3场景、RFT+DPO", "available_resource": "单卡 A100 或 A30，或本地 Mac", "cost_type": "reported" } } ] }