{ "schema_version": 1, "as_of": "2026-10-21", "snapshot_commit": "d1b4851b0cfc5b2bfa0624e5ef4f987a511e1995", "scope": "Three frozen review registers plus the previously reported 54-source-lead cohort", "read_only_reconciliation": true, "ledger_modified": false, "review_watermark_modified": false, "all_history_implementation_complete": false, "sources": [ { "path": "docs/research-audits/oct07-review-closure/paper-register.json", "sha256": "cc252acfd0af1eba7883351b19a7e179d89c32093c76ae049d1de364354a8ae3" }, { "path": "docs/research-audits/oct12-review-closure/paper-register.json", "sha256": "dada63ae0d6a1792c156e77574026d74a28634085bb43da4a71a7f2dc2998073" }, { "path": "docs/research-audits/oct15-review-closure/paper-register.json", "sha256": "c2b0caf1fff3af87bc826e1bc4b0fd0e4ad994cb995c6893f327a32c8e3c95a4" }, { "path": "docs/paper-discovery-ledger.json", "sha256": "bd6a2ac6405ec865d5c3882d5c089f20f45402c2ecd365407d609866c1ebcf7a" }, { "path": "docs/research-manifest.json", "sha256": "799d644f68608b2b7fd42aa5101cfa836038fba68fad3c7428bf10e507e4690a" }, { "path": "docs/research-audits/harness2-priority-repair/backfill-inventory.json", "sha256": "9ce9b525d05c4acf360bc79fd8f4cc41bf38e77ab944574bb2f5f2f82cfdcb4e" }, { "path": "docs/research-audits/historical-backlog-reconciliation-20261021/data/date-checks.json", "sha256": "0424bba41827a9edd66fafa232a5abaae743d33be6e82b49b743aca6c303daba" }, { "path": "docs/research-audits/historical-backlog-reconciliation-20261021/data/global-review.json", "sha256": "ee798ac2d0fd915512d124ce3c5c996a0ee17758c75ba9eecba98364e1da7e4d" }, { "path": "docs/research-audits/historical-backlog-reconciliation-20261021/data/google-review.json", "sha256": "9fe69fc47bf55fd1a270e4b0353133669435c10a2ce36fb42098db3385654697" }, { "path": "docs/research-audits/historical-backlog-reconciliation-20261021/data/meta-review.json", "sha256": "3d7ee8e31c71a8b4dafebdea5fe92d2528047a0d2ac257fa7460c21b79d83be5" }, { "path": "docs/research-audits/historical-backlog-reconciliation-20261021/data/this-audit-dispositions.json", "sha256": "182af4579aa7d090dd548d3f9c74bf542b8a679edf8b6a1588aca64eed310771" } ], "counts": { "ledger_candidate_rows": 4943, "ledger_unique_identities": 4821, "ledger_repeated_identities": 113, "old_pending": 521, "accepted_union": 534, "accepted_catalogued": 26, "accepted_cost_or_source_terminal": 4, "accepted_not_catalogued_no_terminal": 504, "accepted_omitted_from_old_pending": 14, "deferred_misrepresented_as_accepted_delivery": 31, "duplicates_within_old_pending": 0, "catalog_matches_within_old_pending": 0, "duplicate_normalized_titles_among_unclosed": 0, "previous_google_new": 54, "pre_2026_source_leads": 9, "scope_or_document_type_diversions": 6, "current_2026_substantive_review_leads": 39, "source_lead_overlap_with_accepted_or_deferred": 0, "this_audit_full_experiment_terminals": 9, "this_audit_terminal_from_accepted": 7, "accepted_without_full_experiment_terminal_after_this_audit": 497, "this_audit_terminal_from_deferred": 2, "deferred_not_accepted_still_unclosed": 29 }, "source_unclosed_observations_by_track": { "agent": 206, "foundation-model": 212, "post-training": 85, "recommendation": 1 }, "accepted_not_delivered": [ { "id": "2510.15831", "title": "VISTA: A Test-Time Self-Improving Video Generation Agent", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/1021", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1470" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/973" ], "reason": "审查通过，待核心算法实现与公开验证。Google高优先历史补发现：结构化prompt规划、双向tournament保留champion、三维normal/adversarial/meta批评、DTPA修订后实际重生成/再选，属于可执行Agent自改进闭环而非视频主干创新。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": "not_reproduced_high_cost", "this_audit_full_experiment_review_done": true, "component_implementation_authorized_by_this_report": false }, { "id": "2602.16490", "title": "From Growing to Looping: A Unified View of Iterative Computation in LLMs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/1020", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1609" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/972" ], "reason": "审查通过，待核心算法实现与公开验证。Google合作历史漏收录：grow-first/loop-later组合配方与中间block recurrence有真实SmolLM训练、22公共任务及参数/训练/推理三类预算对照；可纳入循环模型研究候选，不称10月新文。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": "not_reproduced_high_cost", "this_audit_full_experiment_review_done": true, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00010", "title": "Heavy-Tailed Memory Traces in Long-Horizon Language Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/1019", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1468" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/971" ], "reason": "审查通过，待核心算法实现与公开验证。CTWM以访问排名幂律预算分配核心/尾部记忆，适合作为可执行记忆压缩机制候选；不能将近零任务成功率的token节约宣传成能力提升。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00012", "title": "When Do Causal World Models Help Modular LLM Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/1018", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1467" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/970" ], "reason": "审查通过，待核心算法实现与公开验证。FedCausalCompose的局部干预响应匹配、边发现与组合约束可作诊断性Agent世界模型候选；真实实现必须学习边，不能以oracle边替代。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00025", "title": "Measuring the Microtask Eligibility Gap: When Is an Off-the-Shelf SLM Enough for an Agent Harness?", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/1015", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1463" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/967" ], "reason": "审查通过，待核心算法实现与公开验证。四microtask用非LLM基线预定质量门槛和CI判定SLM eligibility；calibration上选logprob operating point，memory BM25 shortlist→SLM rerank及tool基线gating。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00040", "title": "DSSR-3D: Decoupled Reasoning for View-Dependent Referring in 3D Gaussians", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/1012", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1460" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/964" ], "reason": "审查通过，待核心算法实现与公开验证。DSSR-3D 用 FLAN-T5 分析目标/方位/锚点，语义 Gaussian softmax 定位锚点，在相机 XZ 平面做局部方向衰减并融合目标分数。Ref-LERF 公开场景支持验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00097", "title": "DramaAgent: Agentic Storytelling Video Generation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/1004", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1452" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/956" ], "reason": "审查通过，待核心算法实现与公开验证。DramaAgent 层级分镜、固定角色reference、逐scene候选视频与音频生成，以身份/语义/时序/音画加权评估，针对最弱维度改写并有限次数重生成。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00182", "title": "Localizing Post-Wire Semantic Changes in MCP Agent Frameworks", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/1002", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1449" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/954" ], "reason": "审查通过，待核心算法实现与公开验证。MCP post-wire差分测试用allowlisted projector保留语义atom与compartment，分别检查协议合法、类型语义留存、特定consumer可访问性，并沿各框架公开carrier边界定位。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00204", "title": "Query Independent Variable Rate Visual Token Coding", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/998", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1445" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/950" ], "reason": "审查通过，待核心算法实现与公开验证。RDTok 在 query-independent PCA 基底上建逐 token 量化率失真曲线，以整数动态规划分配总 bit 预算。ChartQA/COCO 公共图像及冻结 VLM 可验证压缩。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00251", "title": "The Null Is the Hard Part: Exact Tests for Memorization in Generative Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/996", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1440" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/948" ], "reason": "审查通过，待核心算法实现与公开验证。模型级随机标签置换与图像级匹配控制conformal/BH审计可补生成记忆评估，官方公开代码含真实测试入口与测量输出。仅接纳为研究评测候选，非已运行或可证明模型不记忆。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00313", "title": "Rules to Tools: Executable Checks for LLM Agents in Scientific Computing", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/992", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1425" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/944" ], "reason": "审查通过，待核心算法实现与公开验证。Rules to Tools 将公开物理方程/边界/输出约束编成可调用checker，修复时提供artifact实际残差，独立隐藏grader给最终成绩；匹配text与tool组保持书面规则、探针和budget一致。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00354", "title": "Proof-Gated Signing: Solver-Checked Transaction Guards that Hold Under State Drift for Onchain AI Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/987", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1418" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/942" ], "reason": "审查通过，待核心算法实现与公开验证。Proof-Gated Signing 先模拟交易抽取资产变化，再由 Z3 验证价格区间内策略，编译余额/收款/allowance/owner 后置条件，由钱包原子执行与回滚防止 check-to-inclusion 状态漂移。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00372", "title": "When Harnesses Lose the Signal: Causal Evaluation of Recovery in LLM Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/984", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1414" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/939" ], "reason": "审查通过，待核心算法实现与公开验证。CIR 从同一环境状态配对重放 refresh/不 refresh，区分 rescue 与 harm；三 logistic 模型分别预测异常及两种成功，OOF 阈值在 clean-harm 约束下选择，按时间最多刷新一次。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00385", "title": "FAER: Auditable Utility-Aligned Trajectory Replay for Language Model Post-Training", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/982", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1411" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/937" ], "reason": "审查通过，待核心算法实现与公开验证。FAER-UTILITY在非target calibration块拟合梯度对齐/metadata replay权重并无放回抽样，trace冻结后才join正确性；正文完成learner行用normalized alignment，optimizer-aware虚步仅扩展/诊断，不能混称已验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00399", "title": "Metacognitive Reasoning in Energy Based Models using Instance Based Learning Theory", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/979", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1408" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/935" ], "reason": "审查通过，待核心算法实现与公开验证。MERITED以191M EBT能量分布+问题embedding的经验相似性选择MCMC深度或逐步stop，20训练记忆例构建状态动作回报，可作动态计算机制诊断。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00400", "title": "Representation Transitions Reveal Emerging Safety Risks in Multi-Turn LLM Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/978", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1407" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/934" ], "reason": "审查通过，待核心算法实现与公开验证。DART 用危害减良性 transition 均值方向，去除良性均值和主协方差方向，在 segment 边界累积投影，阈值触发后定位最大贡献段并添加提醒再生成。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00431", "title": "ChainLoRA: Geometry-Preserving Task Vector Merging for Continual Learning in LLMs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/973", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1402" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/930" ], "reason": "审查通过，待核心算法实现与公开验证。ChainLoRA按QR carrier链式初始化、一侧几何正则，归档后SVD共同carrier与Procrustes对齐再共享/残差merge；真实持续学习三套数据。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00432", "title": "XOR-Trellis: Ultra-Low-Complexity Dequantization and Curvature-Aware Hadamard-Free LLM Quantization", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/972", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1400" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/929" ], "reason": "审查通过，待核心算法实现与公开验证。XOR-Trellis以16位状态和每步2位分支生成四值FP4 palette，LDLQ调整目标加D加权Viterbi保留曲率，无需Hadamard旋转。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00435", "title": "How AI Agents Discover Scientific Equations: From Hydrotope Rediscovery to New Water-Wave Amplitudes", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/971", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1399" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/928" ], "reason": "审查通过，待核心算法实现与公开验证。PI 将解析与数值探索分配给两位学生，共享记录并独立用 BG 递归找反例，重建跨频率 chamber 的全局振幅公式。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00487", "title": "ScaffoldM3C: A Multimodal Sequential Monte Carlo Framework for Generative Stable Construction Planning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/968", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1396" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/926" ], "reason": "审查通过，待核心算法实现与公开验证。ScaffoldM3C 以 Gemma 条件编码和位置/类别/候选似然三头生成砌块，SMC 对碰撞修复位移惩罚、采样重加权并选最长稳定构建序列。StableText2Brick 可重建监督。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00497", "title": "Gumbel Straight Flow: Distilling Autoregressive Models into One-step Flow Maps", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/967", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1395" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/925" ], "reason": "审查通过，待核心算法实现与公开验证。Google DeepMind Amsterdam参与GSF：AR Gumbel耦合直路径+单调Gaussianization+diagonal教师/semigroup自一致flow-map损失，官方正文含同架构同迭代FMLM-re控制及teacher-forced后验Gumbel训练算法。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": "not_reproduced_high_cost", "this_audit_full_experiment_review_done": true, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00499", "title": "Denoising Surface: Modeling and Predicting Inference Cost for Diffusion LLM Serving", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/966", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1394" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/924" ], "reason": "审查通过，待核心算法实现与公开验证。DWS将扩散解码负载分解为block存活概率和条件denoise-step存活概率，MiniLM两阶段多尺度/Brier监督，CPU INT8预测加硬件成本剖面驱动SJF。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00541", "title": "Random Recursive Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/963", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1391" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/921" ], "reason": "审查通过，待核心算法实现与公开验证。RRM逐样本逐步均匀有放回选择可复用层，真实随机递归训练；TRM+RRM保持完整TRM设置只换层顺序提供核心控制，纳入小规模递归架构研究候选，不称大模型能力已证实。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00559", "title": "PhysVista: Benchmarking Physical Intelligence in VLMs via a Perception-Reasoning-Assessment Loop", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/961", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1389" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/919" ], "reason": "审查通过，待核心算法实现与公开验证。PhysVista以公开视频帧和题目定义感知/因果/物理评估多任务，补充物理VLM评测覆盖，可接入为公共基础设施候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00568", "title": "Emergent Unfaithfulness: How Alignment Training Causes Language Models to Silently Override Task Faithfulness", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/959", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1387" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/917" ], "reason": "审查通过，待核心算法实现与公开验证。FaithConflict公开940对确认/冲突文档、FaithGap与8行为/7推理类型judge，22 checkpoint×prompt条件；官方代码数据已可读。任务忠实性和事实正确性必须分开，不能以复述有害虚构文档作为安全改善指标。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00573", "title": "FORTE: Adaptive Scoring and Exact Keyframe Selection for Long-Video Question Answering", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/958", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1386" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/916" ], "reason": "审查通过，待核心算法实现与公开验证。FORTE 在有限打分预算内用带状 RBF-GP 估计帧相关度和时间覆盖边际收益，再用精确 O(KN) 上包络动态规划选 64 帧。公开长视频 QA 可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00574", "title": "Make Sparse Rewards Count: Density-Aware Reward Aggregation for Multi-Reward RL", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/957", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1385" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/915" ], "reason": "审查通过，待核心算法实现与公开验证。DARA按各reward活跃group密度sqrt(max-density/density)有上限加权，Asym仅放大正优势再batch normalize；ToolRL/BFCL-v4与数学真实训练。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00604", "title": "MIKASA-Robo-VLA: Benchmarking Memory in VLA Models for Long-Horizon Manipulation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/954", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1382" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/912" ], "reason": "审查通过，待核心算法实现与公开验证。MIKASA-Robo-VLA提供ManiSkill3 90任务、两RGB+7D状态/动作统一接口、success_once与22500示范，可补真实执行的VLA记忆评估候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00613", "title": "Spatial Strategies, Not Actions: Vector-Quantized Geodesics as Tools for LLM-Driven Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/951", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1379" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/909" ], "reason": "审查通过，待核心算法实现与公开验证。离线250geodesic五步片段原点归一VQ/Kmedoids压5prototype，加5primitivewait工具；LLM一次描述、运行只选五步openloop，optionalcollisionfilter仅已观察障碍。单QwenVL格世界无高维泛化，signedmeasure动态增skill纯proposal未测；需保留真实geodesic采样与VQ，不用手写技能冒充。机制可作spatialtool建库候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00648", "title": "Incident-Arena: Getting agents to the last nine of reliability", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/945", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1373" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/904" ], "reason": "审查通过，待核心算法实现与公开验证。IncidentArena20K8s故障3substrates，hidden deterministic outcomeSLI+restart/reattack安全gate，healthy阈值先冻结，oracle/noop+adversarial QA；nativeharness3000trials，13Slack导致coverage不均、reasoning差异CI宽，rewardhack posthocjudge不等hidden scorer。官方Apache2任务/Harborfork公开，LinuxDocker8CPU16GB40–77GB；需隔离cluster/secret并不能用生产权限直跑。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00651", "title": "Agent Evaluation Reliability: More Tasks Won't (Always) Fix An Agent Leaderboard", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/943", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1371" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/903" ], "reason": "审查通过，待核心算法实现与公开验证。Bernoulli-logit Bayesian交叉variance分解区分model与model-scaffold部署对象，π²/3剩余项因单cell重试不足不可再分。29,923HAL/54model13scaffold，D-study后验传播显示加任务不能消掉model×scaffold/benchmark方差；均衡exchangeable且setupcost假设下的breadth结论非任意新任务更好。官方MIT R/Stan tests+公开outcomes、美元是价格假设非实测，可纳评测可靠性/预算分配分析工具。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00661", "title": "Exploring More, Reasoning Better: Stepwise Risk-Sensitive GRPO for Diffusion Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/941", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1369" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/901" ], "reason": "审查通过，待核心算法实现与公开验证。StepRS-GRPO按denoising maskedness调风险系数并变换组优势，保留原diffusion trainer；binary reward下仅状态/题目重加权，真实SDAR1.7/4B训练和mass/RMS controls。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00663", "title": "Backdoor Containment via Expert Quarantine and Shutdown in LLMs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/940", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1368" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/900" ], "reason": "审查通过，待核心算法实现与公开验证。QES用attention-derived软trigger定位、sample peakiness调制、四routing约束训练expert LoRA并关闭隔离expert；不依赖gold trigger，实际两任务三攻击多模型。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00666", "title": "VisionQ: VLM-as-a-Judge Taxonomy, Dataset and Benchmark for Qualitative Analysis in Computer Vision", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/938", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1366" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/898" ], "reason": "审查通过，待核心算法实现与公开验证。VisionQ具备新VLM judge基准执行与DPO训练代码，可作作者声明一致性的诊断评测，不能包装为客观视觉质量或自动审稿能力。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00675", "title": "LabBook: Harnessing Experimental History for Efficient LLM-Driven Discovery", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/936", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1364" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/896" ], "reason": "审查通过，待核心算法实现与公开验证。LabBook单自由文本semanticmemory+losslessexperimentlog，latest完整code/result+best/stagnation信息自动入context，agent最多5次History/Code/Result只读检索并共同输出完整新program与memory，不强制bestparent。49FrontierCS+9公开任务同模型prompt/evaluator，去memory/去retrieval、fullhistory/最高分检索消融；跨任务分数按各task最佳归一仅聚合，测试选择应validation隔离。代码willrelease但机制完整可候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00710", "title": "ReLiveGym: Evaluating Long-Lived Agents over Weeks of Replayed Reality", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/931", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1359" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/891" ], "reason": "审查通过，待核心算法实现与公开验证。ReLiveGym8任务真实流回放数周，clock只在wait推进不模拟LLM延迟，observations按timestamp隔离；deterministic timedcommit score+hardcostwallet，sleep/watcher/cron和dailyhindsightmemory/BoNprune。BoN仅过去已结算反馈选次日，无未来gold；某任务得益非全面。官方代码/8scorer/config公开但无数据随仓，DATA提供公开重建而bundle链接占位；需33GB，时间推断有误差，Linux无OSfence仅禁shell。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00717", "title": "Sequential Functional Structured Tucker Compression for Large Language Model Attentions", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/930", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1358" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/890" ], "reason": "审查通过，待核心算法实现与公开验证。FTC真实顺序student激活度量、共享QKV Tucker与稀疏core联合字节预算、WO独立teacher-map校准，含同度量SeqSVD/等存储core消融，作为模型压缩候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00767", "title": "Pre-training interventions, ex post facto: Grafting model beliefs across checkpoints", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/928", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1355" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/888" ], "reason": "审查通过，待核心算法实现与公开验证。在base checkpoint学synthetic-document增量再加到post-trained模型，减少reality drift；多个模型及284B规模实测，区别于旧GRAFT的轨迹交换。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00779", "title": "Effective Synthetic Data Curation Requires Group-Level Signals", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/927", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1354" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/887" ], "reason": "审查通过，待核心算法实现与公开验证。group-level influence评估下游训练效用，以组内梯度diversity优先分配昂贵group oracle调用，未审组用individual proxy；真实pretrain/RLVR训练，需隔离reference与test。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00864", "title": "Kinematic MeanFlow: One-Step Action Generation Policy for Robotic Foundation Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/918", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1344" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/879" ], "reason": "审查通过，待核心算法实现与公开验证。K-MF将MeanFlow时间导数拆成双区间JVP，提供真实的一步动作生成训练机制。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00878", "title": "UniTrackPLA: Unified Panorama-Language-Action Model for Instruction-Guided Navigation and Dynamic Person Tracking", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/915", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1341" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/877" ], "reason": "审查通过，待核心算法实现与公开验证。UniTrackPLA 将 DINOv2/SigLIP 全景多视图输入与时间/方位编码映射到机器人航点，WAC 预测动作后的潜变量并以观测误差触发重规划。OmniTrack/OmniVLN 可做路由隔离验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00885", "title": "FORALL-LEAN-AGENT for Auditable Reasoning in Formal Mathematics and Software Verification", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/912", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1338" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/874" ], "reason": "审查通过，待核心算法实现与公开验证。Forall-Lean-Agent隔离actor与fresh readonly reviewer、同artifact digest双批准；proofbodyless检索、OS全进程树sandbox/provider-only proxy、syntax/name-shadow/axiom审计，再serial clean验证。100VSB/672Putnam与2LeanEval；benchmarkrule/strict/independentkernel分母不同，legacy参考泄漏与filedrift有排除不能算能力。成本/compile/review预算非一致，不把harness收益归单模块。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00899", "title": "TOAST: Stochastic Robot Action Tokenization for Autoregressive Vision-Language-Action Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/908", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1336" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/870" ], "reason": "审查通过，待核心算法实现与公开验证。TOAST动作分词随机监督是清晰的基础VLA训练算子，可优先在公开LIBERO做同预算验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00952", "title": "A Matched-Budget Audit Framework for Recaptioned Image-Text Supervision Distributions", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/902", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1331" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/865" ], "reason": "审查通过，待核心算法实现与公开验证。Matched-Budget Recaption Audit在同源图片同文本预算下比较caption覆盖、prefix集中度、CBU密度与图像支持风险，适合作预训练数据筛选诊断。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00977", "title": "ABSENTIA: Detecting Broken Access Control Vulnerabilities in Web Applications", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/895", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1325" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/861" ], "reason": "审查通过，待核心算法实现与公开验证。ABSENTIA应用profile→双LLM抽pattern并集构routegraph→逐route声明invariant、查enforcer并source-grounded反例→共享file跨routecomposition，byte-identical依赖闭包缓存跨commit复用。BAC30已披露漏洞25项目，advisory/fix仅oracle；fixednegative仅对应bug非全repo安全，precision独立human73sample未确认LLMjudge增益。静态白盒defensive审计算子合格，代码正文willavailable不能声明已发布。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00979", "title": "RISED: RubrIcs for agentic multi-environment Selection and sElf-Distillation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/894", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1324" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/860" ], "reason": "审查通过，待核心算法实现与公开验证。RISED固定rubric词表(4环境82/主3环境59)，Qwen32B异步tag每rollout；环境roundrobin定quota下MMR pooled相关−已选冗余选64/128groups，正负rubric选样但teacher仅top3positive union，k3OPD重要性比+GRPO，负rubric用于steer。3环境Qwen3/4B，强baseline仅+1.2/.6pp；半optimizer步不等半总成本，62.9%zero-variance组OPD仍有信号；删distill消融同时删negative steering要注明。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00981", "title": "NarrativeFlow: Flow-Based Vision-Language-Action Model Using Robot Velocity Fields", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/893", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1323" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/859" ], "reason": "审查通过，待核心算法实现与公开验证。NarrativeFlow双任务token与叙述差分训练监督增强语言条件机器人flow，可作算法候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.00983", "title": "The Devil Is in the Reconstruction Loss Scale: Rethinking Optimization in LLM Quantization", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/892", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1322" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/858" ], "reason": "审查通过，待核心算法实现与公开验证。将PTQ重构MSE改为分组RMSE，使非零误差的输出梯度范数不依赖损失尺度；sample/channel/token/element是同一家族，可接现有PTQ训练对照。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01017", "title": "Pay for the Fault, Not the Flow: Label-Free In-Flow Multi-Agent Workflow Optimization", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/888", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1319" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/854" ], "reason": "审查通过，待核心算法实现与公开验证。InFlowOp cost=-log rubricsmatch造workflow，topdown拆/bottomup合并，运行中declaredinput/output contract判decomposition/assignment错，先换agent再局部重拆，invalidates descendants重跑之外缓存。无gold但semanticmatch非oracle；rollback不劣仅在检查可信/副作用可还原条件下。Braid8域，3闭源仅5域，singleLLM没工具无法读文件基线弱；同turnbudget单agent比较更有意义，profile成功筛选避免test跨样本泄漏。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01023", "title": "Groundability, Not Scale Alone: When Weak Reviewers Can Audit Strong Coding Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/887", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1318" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/853" ], "reason": "审查通过，待核心算法实现与公开验证。reviewability框架区分agent testimony/结构化/独立grounded证据；冻结cascade先空patch/新增staticerror再generatedtest（须base行为失败）再8Breview+abstain。official hidden test仅upperbound不能部署；32Django设计后121/122 GPT+59Gemini一次test，risk .33/.26、overreject .66/.67，generatedpass非证明。可纳独立核验/弃权策略候选，不能宣称可靠弱审强或总成本优势。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01037", "title": "SLIM: Simplex-Lattice Interpolation Merging", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/884", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1316" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/851" ], "reason": "审查通过，待核心算法实现与公开验证。SLIM用专家顶点与两两中点最小simplex二次格点设计拟合合并性能，再约束优化预测AVG；与同36次预算随机设计比较且另有heldout OOD。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01054", "title": "Capturing In-Context Learning Dynamics with Task Operators", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/882", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1313" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/849" ], "reason": "审查通过，待核心算法实现与公开验证。Task Operator从ICL前向提取逐head/位置的乘性softmax质量与加性context value，用于无demo replay；解析定义、公开实现、明确answer-excluded验证和测试分离满足候选条件。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01076", "title": "GLoC-EHR: Evidence-Cited Clinical Reasoning over Global Context and Local EHR Events", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/879", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1599" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/847" ], "reason": "审查通过，待核心算法实现与公开验证。GLoC-EHR新增global/local EHR memory接口与分阶段alignment、evidence-only sufficiency辅助loss；128350患者7:1:2隔离及外院验证。RL本身标准GRPO，接纳架构/grounded训练机制，gold只teacher/scorer可见。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01127", "title": "Counting and Min-Cost Encoding for Tokenization in Large Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/876", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1308" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/844" ], "reason": "审查通过，待核心算法实现与公开验证。CNF–MCE以直接substring统计/实际使用过滤建词表，再用带边界成本的全局DP分词，属清晰基础模型tokenizer算法候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01128", "title": "Grounding Large Language Models in DSGE Simulators for Policy Generation and Forecasting", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/875", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1598" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/843" ], "reason": "审查通过，待核心算法实现与公开验证。公开DSGE语言Agent环境：Snowdrop结构模拟器的状态深克隆、action校验、rolling forecast只提交首步、不可重写历史；matched PPO/GRPO共享起点/冲击。作者repo已可读，可作领域仿真任务；不是现实政策建议/实盘因果证据。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01172", "title": "Learning Rate Transfer for Hybrid Transformer-SSM Architectures", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/871", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1304" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/840" ], "reason": "审查通过，待核心算法实现与公开验证。混合Transformer-SSM的naive μP+AdamW跨width学习率迁移给出可执行参数化/选参配方，可作研究配置operator而非新理论保证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01192", "title": "FlashBack: Knowing When to Remember in Streaming Vision-Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/868", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1301" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/837" ], "reason": "审查通过，待核心算法实现与公开验证。FlashBack 分隔原生流式状态和临时 Recall 分支，以时间证据路由选择历史检索/SideKV，Recall 推理不写回原生历史。公开视频基准可测因果长期记忆。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01238", "title": "Mixture-Trained Merging for Unified Multi-Objective Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/864", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1595" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/834" ], "reason": "审查通过，待核心算法实现与公开验证。MTM训练mixture-biased soft experts，以merge-simplex多目标BO选分支混合并迭代复用merged起点；真实Qwen4B/OLMo7B多域与think模式，必须validation选择并计入搜索训练。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01243", "title": "When the Judge Acts: Auditing VLM-Guided Image Selection on Culturally Situated Prompts", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/863", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1298" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/833" ], "reason": "审查通过，待核心算法实现与公开验证。JudgeActs固定CulturalFrames候选池，三次cyclic顺序检查与预设abstention gate，对照精确随机期望及事后人评oracle；可作VLM选择可靠性审计。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01278", "title": "SCOPE-AD: Sequential cost-aware ordinal-belief planning with energy-based models for diagnostic agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/858", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1593" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/830" ], "reason": "审查通过，待核心算法实现与公开验证。SCOPE-AD以观测ordinal belief、离线soft Bellman EBM规划teacher，再masked KL蒸馏到Qwen；ADNI患者级1552/333/333，真实未取得值隐藏。限医学离线规划非通用OPD或临床部署。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01286", "title": "Dyna3: VLM-Guided Training-Free 4D Reconstruction via Depth Foundation Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/857", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1294" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/829" ], "reason": "审查通过，待核心算法实现与公开验证。Dyna3 用 DA3 特征跨帧最佳匹配相似度判运动，Qwen2-VL/SAM3 语义 mask 协同分离静态聚合和逐帧动态点云。DAVIS/TUM/Sintel/DyCheck 可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01316", "title": "What Wins a Vote? Formatting, Length, and Lexical Diversity in the French Compar:IA LLM Arena", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/856", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1293" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/828" ], "reason": "审查通过，待核心算法实现与公开验证。公开偏好评估分析入口：投票时截断对话、风格/长度/词汇联合Bradley–Terry与model内对照；官方代码及derived data可读，原French ComparIA数据有gate。137293 decisive votes/116 models，complete-case127092；调整rank不是已证实更接近能力，不能因观察关联推断风格因果效应。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01492", "title": "Q-SPT: Learnable Query-Based Compression for Low-Frame-Rate Speech Tokenization", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/844", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1281" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/820" ], "reason": "审查通过，待核心算法实现与公开验证。Q-SPT 用滑窗 query 交叉注意力降采样 SenseVoice/DAC 特征，语义 FSQ+条件声学 RVQ，联合频谱/对抗/文本重建等训练。LibriSpeech 重建、ASR/TTS 支持公开验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01496", "title": "SALD: Self-Referenced Advantage Learning for Diffusion Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/843", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1280" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/819" ], "reason": "审查通过，待核心算法实现与公开验证。SALD 同模型 easy-noise 无梯度教师与 hard-noise 学生比较，保留 batch-normalized 可微优势权重、样本 EMA 噪声课程和频域差异门控。公开扩散数据/检查点可真实训练。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01499", "title": "VTR-Bench: A Systematic Benchmark for Evaluating Visual Text Rendering in Video Generation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/842", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1279" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/818" ], "reason": "审查通过，待核心算法实现与公开验证。VTR-Bench公开300prompt、CoQ及carrier转录WER评测入口，支持新视频模型产物；核心keyframe agentic策略亦可研究但不能用未匹配预算增益作效率结论。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01506", "title": "MCRI: A Four-Dimensional Framework for Analyzing and Evaluating Agent Skills", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/841", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1278" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/817" ], "reason": "审查通过，待核心算法实现与公开验证。MCRI可执行文本rubric四维Metadata/Constraints/Resources/Instructions先各1–5再task-conditioned aggregate，重复mean score×mean confidence选skill。3公共bench50subtasks/345Mind2Websteps，58275calls；相关/popularity不等于因果或通用质量。AppendixD完整评分器足以作为候选预筛机制，但F明确/Resources未包含且无公开repo，不能声称复现实验或公开全数据。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01508", "title": "OverAct: Measuring and Mitigating Proactive Over-Authorization in LLM Tool-Calling Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/840", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1277" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/816" ], "reason": "审查通过，待核心算法实现与公开验证。OverAct以保守literal-minimal工具集定义越界（非普适授权法则），8域48tools/240双标原题扩720paraphrase，set操作无LLMjudge。SelfAudit先request-grounded justification再preexecution filter不需oracle，filter承担多数收益，justificationalone反恶化；TCR .94→.89。可纳入权限过滤机制候选，本文称附材有logs但未定位公开下载，不把720题宣称已接入可执行benchmark。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01513", "title": "Decision Titan: Test-Time Training for Long-Term Memory in Offline Reinforcement Learning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/837", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1275" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/815" ], "reason": "审查通过，待核心算法实现与公开验证。Decision Titan 将 RTG/state/action DT 与 Titans fast-weight 神经记忆联合，按轨迹 chunk 更新记忆。X-Maze 重复记忆任务可验证长度外推。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01537", "title": "FedFit: Federated Fine-Tuning of LLMs via Vector-Bank Parameterization and Quantization", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/832", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1270" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/811" ], "reason": "审查通过，待核心算法实现与公开验证。FedFit的双disjoint向量库、交替联邦聚合与低比特通信具有可复现研究价值；保留公式歧义核对门槛。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01560", "title": "AURAL: Adaptive Latent Reasoning with Joint Chunk for Speech Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/830", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1590" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/810" ], "reason": "审查通过，待核心算法实现与公开验证。AURAL用对角+低秩协方差GMM联合latent chunks、CSA、可微两遍scheduled sampling，再按latent坐标归一ratio做GRPO；真实语音/文本基准和延迟，应保持整个latent模型。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01564", "title": "Chaining Skills to Hijack LLM Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/829", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1269" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/809" ], "reason": "审查通过，待核心算法实现与公开验证。APEX跨skill中间artifact将工作进度错误提升为授权，独立sandbox执行反馈优化chain；防御taint+原请求核验。690attempt四定向families512成功，family任务重叠，utility按checks权重不是整任务成功；workloop token增不自动构成越权。官方MIT执行/trace判据/参考controlled variants公开，SkillsBench另取；只纳防御沙箱评测，不执行外部攻击。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01569", "title": "Managing Context and Communication in Distributed Agentic UAV Swarms", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/828", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1268" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/808" ], "reason": "审查通过，待核心算法实现与公开验证。分布式UAV新增peer-tag残差embedding+recency novelty最小距离gossip，管理员离线双阈值决定discard/compact/full，接收queue只是邻居已知的估计；三内存tokenFIFO、ID/path/TTL去重。10机SITL搜救，80211n解析模型非实机网络；proposal全完成而flood70–85%，后续耗时/误差仅成功run条件统计，不可只看省6–7MB到<0.5MB。通信/记忆策略可泛化候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01595", "title": "Before It Fades: Reinforcing Temporal Representations at Inference Time in VideoLLMs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/826", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1266" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/806" ], "reason": "审查通过，待核心算法实现与公开验证。TAI 比较视频正序/逆序的逐层激活差异，取输入特定峰值向量按衰减注入后层；冻结三 VideoLLM、四公开时间推理基准可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01605", "title": "Hob-VL: A Benchmark for Visually Grounded Boolean Reasoning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/825", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1265" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/805" ], "reason": "审查通过，待核心算法实现与公开验证。Hob-VL提供符号/NL视觉布尔组合及DeMorgan一致性公开评测，可执行新provider推理并离线严格评分。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01625", "title": "Beyond Domain-Level Adaptation: Margin-Oriented Semantic-Appearance Interaction Correction for Personalized Federated Vision-Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/824", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1264" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/804" ], "reason": "审查通过，待核心算法实现与公开验证。MOSAIC 在冻结 CLIP 上联合共享/私有 prompt、域类交互残差与低秩类别修正及样本门控。Office31/Home/DomainNet 的公开联邦域分类可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01637", "title": "Fusing Visual and Textual Representations via Multi-layer Fusing Transformers for Vietnamese Visual Question Answering", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/823", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1263" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/803" ], "reason": "审查通过，待核心算法实现与公开验证。MFT 将 PhoBERT/ViT 多层双向跨模态 attention 融合用于越南语 VQA 分类。公开 ViVQA 可训练并测 EM/token F1，作为窄域多模态算法候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01640", "title": "Not All Error Yields to Scale: Where Scaling Stops in Vision-Language Inference", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/822", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1262" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/802" ], "reason": "审查通过，待核心算法实现与公开验证。Separable Inference Scaling Law按backbone参数和真实视觉token数分别拟合误差floor+幂律，结合成本面选择heldout配置；可作为多模态预算规划诊断。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01649", "title": "CrossGMN: Graph Metanetworks for Cross-Architecture Weight-Space Transformations", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/821", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1261" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/801" ], "reason": "审查通过，待核心算法实现与公开验证。CrossGMN可作为小模型架构压缩/evolve元算子研究候选，不能宣称已支持LLM压缩。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01670", "title": "Do MLLM Judges Judge the Edit? Auditing Bias in Image Editing Evaluation with Verified Quality Preservation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/820", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1260" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/800" ], "reason": "审查通过，待核心算法实现与公开验证。EditJudgeBias独立judge偏差审计：quality-preserving cues、sham/retest噪声基线、顺序互换、paired统计。官方仓库有cue/judge/metric CLI，可评新judge，但不含全部图像/人工注释/作者I2EBench输出，不能宣称精确复现原数值。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01687", "title": "Architectural Sampling: Test-Time Scaling via Computational Diversity in Frozen Vision-Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/819", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1259" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/799" ], "reason": "审查通过，待核心算法实现与公开验证。Architectural Sampling 重复冻结 VLM 的层窗口，组合递归次数得到结构 rollout，再以无标签自奖励 TTRV-GRPO 训练。公开多模型/十二基准可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01710", "title": "CoEvolve: Construct-to-Edit Visual Grounding with Bidirectional State Refinement", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/817", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1257" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/798" ], "reason": "审查通过，待核心算法实现与公开验证。CoEvolve 以 RER-GRPO 训练有序多步 bbox 进展，再用 BDR 坐标去噪保留推理文本。RefCOCOg 训练和 RefCOCO/+ 持出评测可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01741", "title": "ATI-VLA: Action-Centric Predictive Vision-Language-Action Models via Actionable Alignment Then Adaptive Injection", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/816", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1255" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/797" ], "reason": "审查通过，待核心算法实现与公开验证。ATI-VLA共享动作/观察量化空间与分层侧路注入符合VLA新算法候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01769", "title": "CONTRA: Discovering and Qualifying Behavior-Changing Questions for Selective Clarification in LLM Code Generation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/812", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1251" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/793" ], "reason": "审查通过，待核心算法实现与公开验证。CONTRA先多轮发现问题，再要求相关且原需求未解决；每种可能答案采3程序、>=7共同输入，组内2/3稳定且至少一输入跨组不同才保留，失败非等价证明；history自适应选问或stop。419欠指定+80完整任务最多5问，GPT4o三票gold匹配；四模型macro F1 41.20vs27.32，程序隔离执行。通用执行证据澄清算子，plugin部署非线上A/B。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01815", "title": "Debias Anything: Fairness with Diversity without Supervision in Diffusion Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/809", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1248" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/791" ], "reason": "审查通过，待核心算法实现与公开验证。Debias Anything 在冻结扩散模型上训练语义 adapter，以属性方向、目标比例引导和 disagreement 多样性项进行批级去偏。公开人脸/扩散检查点可检验质量公平性权衡。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01892", "title": "Selection-Based Structured Reasoning: Toward Efficient Multimodal Search Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/802", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1588" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/784" ], "reason": "审查通过，待核心算法实现与公开验证。SSR全文不是普通工具应用：六共享推理模板以batched teacher-forced mean-logprob softmax选择，再插入全文生成action；SFT selector NLL+action NLL，GRPO一个categorical selector ratio+action token ratios，插入模板不逐token训练。AppendixB实际detach rollout竞争分数仅首步概率精确、后续近似。Qwen3VL2B/4B七multimodal搜索任务，效率须计候选prefill；保留原post摘要漏判并由正文纠正。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01917", "title": "MoLE: Mixture of Latent Experts for Complementary Visual Reasoning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/798", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1241" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/781" ], "reason": "审查通过，待核心算法实现与公开验证。MoLE的路由视觉证据、专家value变换与隔离潜在专家/summary拓扑构成可辨识新视觉推理结构，公开数据+详细附录足够作为保真实现候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01921", "title": "Cross-Lingual Alignment for Decoder-Only Models using MoE Routers", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/797", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1240" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/780" ], "reason": "审查通过，待核心算法实现与公开验证。XLR以平行语料的sequence-mean sparse router概率KL进行中层跨语言对齐，核心CPT辅助loss+packing优化明确，官方训练/评测公开，符合基础模型训练候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01955", "title": "Do Your Own Research: Learning to Forecast by Learning to Search", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/794", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1585" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/777" ], "reason": "审查通过，待核心算法实现与公开验证。prime-forecast公开工具环境/历史问题/rollout记录，point-in-time过滤、封锁市场价格、Brier和弃答插补可独立评估研究预测；训练/测试按解决日期隔离。实时检索和API过滤非静态可复现，过滤的完美小样本结果不是无泄漏证明。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.01962", "title": "SIEVE: Selective attention-value Suppression for Vision-Language Models Unlearning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/793", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1237" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/776" ], "reason": "审查通过，待核心算法实现与公开验证。SIEVE 交替优化 forget 的负 CE+attention value 归零和 retain 的正 CE+冻结参考 value 匹配，定义目标明确。合成 Profile Visual-QA/Textual-QA 可做选择性遗忘。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02005", "title": "Counting Moves, Weighing Voices: Bayesian Dialectical Argumentation for Calibrated Multi-LLM Councils under Persistent Adversaries", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/788", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1232" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/773" ], "reason": "审查通过，待核心算法实现与公开验证。BDA把Propose/Challenge/Concede转签名endorsement counts，积分Beta可靠性的DawidSkene式汇聚，约25标签校准同预算比较；K>2伪似然需tempering，校准理论仅二类无challenge等条件。标签身份稳定学习在位置MCQ失效，sleeper攻击/可靠性突变不保证；是新聚合机制非无条件鲁棒获胜。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02019", "title": "Controllable Multi-label Video Safety Detection via Adaptive Tversky Policy Optimization", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/785", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1230" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/770" ], "reason": "审查通过，待核心算法实现与公开验证。ATPO以FN/FP EMA log-ratio误差反馈调Tversky alpha/beta，固定和为2，可全局/类别控制多标签reward；真实VideoVLM训练，新增自适应reward控制器而非重命名标准GRPO。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02091", "title": "GeoLatent: Geometry-Guided Latent Structuring with Routed Optimization for 3D Reasoning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/777", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1223" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/765" ], "reason": "审查通过，待核心算法实现与公开验证。GeoLatent 分离共享几何成分与残差，用三阶段路由优化阻断答案的视觉捷径；DA3/VGGT 训练目标只辅助潜变量。SPAR/SpatialLadder、SPBench/ViewSpatial 可真实验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02140", "title": "Finetuning with Sampling: SFT Learns Better Than You Think", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/775", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1582" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/764" ], "reason": "审查通过，待核心算法实现与公开验证。Projection Sampling用expert条件proposal与Metropolis-Hastings分块采样使信息保留轨迹靠近base，再SFT；化学/数学/开放域真实训练。不得将有限MCMC近似等同理想投影精确采样。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02150", "title": "From Knowledge Access to Source Learning: Developing Source-Specific Competence", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/771", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1218" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/761" ], "reason": "审查通过，待核心算法实现与公开验证。SourceLearn持久source-grounded实体表示，Inspect/Deepen/Connect自学+consolidate；taskguided只用训练guidance答案诊断缺证据并重读source，更新representation/prompt不是模型权重。五任务、3次30/70guidance/test分割与AppWorld官方split；测试答案不得进入memory，QA judge与原生执行指标分开。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02161", "title": "DuoMind: Enabling Distributed Multi-Robot Coordination with Semantic Communication", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/769", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1215" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/759" ], "reason": "审查通过，待核心算法实现与公开验证。DuoMind每机器人VLM高层orchestrator、VLA动作，四字段意图/子目标/信念/可选不确定性通信，无联合特殊训练；7RoboPoly+8RoboTwin双臂改分布、每任务50双机器人demo转100单体训练。共享global相机+本地wrist，模拟两机器人非实机大群体；通信与π0替换消融支持模块可复用，不能宣称更大规模部署。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02190", "title": "Trust the Direction, Search the Step: Zero-and-First-Order Methods for LLM Fine-Tuning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/764", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1579" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/755" ], "reason": "审查通过，待核心算法实现与公开验证。ZFO沿基础一阶optimizer方向，用同batch正负两次前向估二三阶导，Taylor/Padé区间择步并奇点fallback；真实LLM微调与开销实验，须匹配额外前向预算。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02191", "title": "The Missing Primitive: Diagnosing and Repairing Mathematical Reasoning in Large Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/763", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1578" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/754" ], "reason": "审查通过，待核心算法实现与公开验证。Absorb以数学primitive为训练teacher特权，用reverse-KL单侧clamp限制压低学生偏好的强度，保留正向teacher指导；Prim/HLE/HMMT/Omni实测。gold primitive不得进入被评价student。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02196", "title": "InterEvolve: Test-Time Evolution of Reward Programs for Humanoid Loco-Manipulation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/762", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1211" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/753" ], "reason": "审查通过，待核心算法实现与公开验证。InterEvolve 在冻结 forward-backward 运动先验上进化分阶段奖励程序，内层 CMA-ES 调参数、外层 LLM 调结构，固定场景验证并复用已核验技能。OMOMO/GRAB 与模拟八类任务可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02202", "title": "ScholarCatalyst: A Benchmark for Retrieving Papers That Inspire New Research", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/760", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1209" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/752" ], "reason": "审查通过，待核心算法实现与公开验证。ScholarCatalyst207真实项目894查询（207Core/687Sub）作者验证inspiration与hard negatives，删solution泄漏并按源论文完成时间截断corpus。Recall标签不完备；oracle候选重排只诊断不供agent。公开191k文档、evaluation及受限本地corpus agentic-search脚本，CCBYNC4，适合检索/科研评测入口而非新训练loss。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02204", "title": "Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/759", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1208" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/751" ], "reason": "审查通过，待核心算法实现与公开验证。RPG从ABC视频重建MuJoCo任务并人工检查冻结evaluator；Runtime只部署观测，privileged独立episode供failureanalysis，Implementor修改skill/prompt，paired5seeds跨22任务gate平均增且单任务不降超.2，merge后重测。15轮单run/候选compute不匹配；savedround heldoutseed回顾评测非未见task，3实机各10trial统一标定冻结，不把privileged轨迹当部署policy。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02206", "title": "KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux with Runtime-Free Verifiable Rewards", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/758", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1577" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/750" ], "reason": "审查通过，待核心算法实现与公开验证。KaliBench公开8504 query-command/1642工具数据与runtime-free shlex评分，三个信息条件、alias-aware参数匹配可作工具使用评测。仅离线命令字符串评分，不等同实际环境执行成功；不得自动对真实目标执行命令。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02298", "title": "EditHero: A Benchmark for Long-Horizon Part-Level 3D Editing and Vibe Modeling", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/752", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1217" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/744" ], "reason": "审查通过，待核心算法实现与公开验证。EditHero由part library与操作日志重建457链2755turn的精确3D目标，self-rollout分开编辑区/保持区与no-op/regeneration对照，可补多轮生成评价。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02304", "title": "SimuVerity: Benchmarking Agents for Engineering-Grade Simulink Model Generation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/751", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1213" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/743" ], "reason": "审查通过，待核心算法实现与公开验证。SimuVerity101任务10工程域，publicspec/profile与hidden场景阈值隔离；交付/执行/工程qualification三gate后六维评分，非交付/候选故障计失败、基础设施故障重跑。6agent每任务3次，10强任务消融为选择性子集非总体因果。官方Apache2仓库公开101任务/scenarios/scorers/reference或importrecipe与结果验证；运行依赖逐任务MATLAB/Simulink商业产品和第三方资产许可。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02323", "title": "World-Calibrated Proposal-to-Action Flow for Vision-Language-Action Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/749", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1206" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/741" ], "reason": "审查通过，待核心算法实现与公开验证。ProAct具有完整公开训练核心，proposal/world-calibration/anisotropic flow可进入候选，先限LIBERO可执行路径。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02324", "title": "Slow-Fast Multi-Teacher On-Policy Distillation for Capability Preservation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/748", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1576" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/740" ], "reason": "审查通过，待核心算法实现与公开验证。SF-MOPD维持EMA慢学生，以slow-fast logprob方向去除teacher冲突部分并JSD蒸馏+原MOPD正则；QwenVL2/4/8B多domain实测，需真实EMA参数而非固定reference替代。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02330", "title": "Choosing Before Acting: Comparative Value Estimation for Long-Horizon Tool-Use Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/747", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1205" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/739" ], "reason": "审查通过，待核心算法实现与公开验证。CITA/CIM由observed同context成功率、Bayesian图synthetic无真实API、LLM模拟比较三源训练valueBCE+gapMSE+transitivity/confidence，Vc排名nexttool，confidenceweightedvalue与taskreward混合GRPO。Toolathlon/TOUCAN/TRAJECT、Qwen7B/Llama8B；value不是因果反事实真值，sandbox必须标synthetic，真实闭环评估不可由自身CIM评分替代。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02344", "title": "Drive vs. Decay: On the Training Dynamics of Joint-Embedding Predictive Architectures", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/746", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1204" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/738" ], "reason": "审查通过，待核心算法实现与公开验证。ResidualPred以残差预测器和对角注意力初始化抑制JEPA塌缩；表5五种子、同ImageNet1K30epoch预算控制支持新机制。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02349", "title": "MIRROR: Multipath Quorum Integrity for LLM Multi-Agent Communication", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/745", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1203" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/737" ], "reason": "审查通过，待核心算法实现与公开验证。MIRROR发送端一次生成canonicalbytes，5逻辑路由3payload2digest majority恢复；unkeyedhash非认证，安全完全依诚实endpoint且受控routes<半数/二次原像。AutoGen/CAMEL/MetaGPT和Gemma/Gemini评测只logicalroutes同backend、无真实独立故障域/时延测量；只抗in-transit篡改，不抗已被注入sender，须保留此条件。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02351", "title": "DeReAct: Decomposed Reasoning and Acting for Reliable AI Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/744", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1202" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/736" ], "reason": "审查通过，待核心算法实现与公开验证。DeReAct Brain提intent/command、独立Critic按约束failclosed gate、CM只保环境支持state并判完成，Brain另见首5+recentwindow。GAIA165/SWE489 native非LLM成功评分，3attempt pass1与pass3，groundingjudge另算；CM是模型判定非形式验证，Claudecritic/CM固定且非真实外部副作用验证，不能把successflag替换真实benchmarktest。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02359", "title": "Lexicographic Multi-Objective On-Policy Distillation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/743", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1575" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/735" ], "reason": "审查通过，待核心算法实现与公开验证。LMOPD按最高优先未满足reward路由expert，再对centered log-policy correction一次顺序投影成virtual teacher；真实多expert实验，但专有训练语料、单次非正交投影不保证全部约束或reward级无退化。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02361", "title": "SEDIMA: Cross-Run Hierarchical Insight Memory for Evolutionary Search Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/742", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1201" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/734" ], "reason": "审查通过，待核心算法实现与公开验证。SEDIMA每evaluation写raw→LLMinsight→attentioncentroid cluster，mutation前cluster/insight检索与证据trace→3建议，不改evolve搜索器。AlgoTune/ALE私有评分与publicfitness分离、memory由disjoint问题建；cold也在线写记忆。100候选匹配不等token/calls，主表单run，ablation/transfer才3seeds；boundedcontext非总检索常数成本。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02388", "title": "Octrees as an Explicit 3D Language", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/737", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1197" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/729" ], "reason": "审查通过，待核心算法实现与公开验证。OctLLM 以稀疏六级 octree、独立 3D 分支和 3D RoPE 共用冻结 Qwen2.5-VL，另训占据补全并接 TRELLIS 解码。Objaverse-XL/HSSD/ABO/ShapeNet 与 Toys4K/PointLLM200 可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02396", "title": "Inherit-MAS: Test-Time Evolution of Multi-Agent Systems through Workflow and Execution Inheritance", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/736", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1196" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/728" ], "reason": "审查通过，待核心算法实现与公开验证。InheritMAS对latest-incumbent DAG先prune再一次edit，最终单独judgelexicographic排名；继承仅exact resolvedrequest含模型/参数/上游/环境指纹且recordhash有效的只读节点，statechanging sink每次live。WorkBench/HotpotQA，部分baseline retrievalbudget非完全同，inheritedtoken只是等价历史量不是实际消耗；judge最高不保证真实utility单调。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02404", "title": "Trained Agentic Context Management", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/735", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1195" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/727" ], "reason": "审查通过，待核心算法实现与公开验证。训练agenticcontextmanagement：最小read_chunk/spawn工具，SFT oracle合成tree-reduce/sequential-fold/边界扩读策略，非成功RL方法（RL退化）。RULER每type5、OOLONG每family10且nativebaselinecontext不同，8kagentbudget可扩展但未正式量化成本；temperature经验选需另设validation，不能把oracle训练脚本开放给评测agent。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02425", "title": "Finding the Move Is Not Winning the Game: XiangqiBench for Closed-Loop Evaluation of LLM Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/734", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1194" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/726" ], "reason": "审查通过，待核心算法实现与公开验证。XiangqiBench119搜索+人工验证残局（非形式证明），gold解法不入agent工具；40ply/depth18固定defender、三次连续非法失败、简化重复规则。8568轨迹12模型2设置3trial，pass@3与全部成功分报；Sighted/Restricted提示也不同。官方MIT run/score与数据卡公开，PikafishGPL另计，原NNUE被替换须冻结engine+network不可直接换新版。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02444", "title": "Counterexample Generation via Per-Theorem Symbolic Verifiers: When Imitation Hurts and Reinforcement Repairs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/731", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1574" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/723" ], "reason": "审查通过，待核心算法实现与公开验证。SymCE公开反例任务/逐定理Python verifier/结构schema和dense/sparse reward，可作RLVR/evolve评测；作者仓库含corpus/splits、oracle/sandbox、runner及逐条结果。schema由canonical witness构造，需审查是否泄漏答案形状，隔离任意Python执行。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02460", "title": "CUEing User Simulators: Calibrated User Embeddings for Multi-Turn Benchmarking", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/728", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1189" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/720" ], "reason": "审查通过，待核心算法实现与公开验证。CUE将userturn与前序assistant attention编码、slotcommanddecoder及latentdiffusion采样，content抑制不保证去除任务/结果信息。completed-human-trajectory条件replay是回顾校准而非未来预测；无目标轨迹的sampledpersona才新用户模拟，sessionmixture不等于人群分布。483tauUSI及heldout SimulatorArena/PRISM；官方Apache2 runtime/training和HF模型/annotation/examplepool入口可用，未执行代码。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02462", "title": "Capability Scaling-Down Laws for LLM Compression", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/727", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1188" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/719" ], "reason": "审查通过，待核心算法实现与公开验证。按能力分别预测剪枝、量化和蒸馏损失：共享剪枝密度形状、位宽/组大小网格插值，以及独立数据量和训练曝光的蒸馏关系，可用于evolve的预算分配。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02472", "title": "APDMem: Agent-Controlled Progressive Disclosure for Query-Adaptive Long-Term Memory", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/725", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1187" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/717" ], "reason": "审查通过，待核心算法实现与公开验证。APDMem L0摘要索引/L3原文先建，L1事实/L2turnnotes按需缓存；hybrid+RRF search与agent自主drill，reflection无新证据或20步停，draft→organizednotes→answer。LongMemEvalS500，baseline从文献取非全重跑；onlinecost只本方法有且排除共享answergenerator，不能宣称等预算优于全baseline。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02478", "title": "Tropical Reinforcement Learning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/724", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1186" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/716" ], "reason": "审查通过，待核心算法实现与公开验证。Tropic对deterministic resettable环境维护已观测状态/片段图，每轮冻结actor重评分，max-plus prefix/suffix求最佳verified可达路径，frontier补采样与topL拼接必须整路径重放验证；best+低覆盖archive路径per-tokenNLL更新，无rewardbaseline/负向项。四环境两backbone三trainingseeds、heldout实例；不能用gold路径或不具可重置性真实API冒充。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02480", "title": "MEA: A Reward-Driven Multi-Agent System for Faithful Model Explanations", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/723", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1185" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/715" ], "reason": "审查通过，待核心算法实现与公开验证。MEA共享backbone Proposer工具策略/Actor解释，以遮挡再query原classifier衡量faithfulness，两段统一terminalGRPO advantage；vision mask面积λ.3与toolcountλ.05抑制gaming，text/tabular top25%约束。三模态及未训练Q8/Q9；遮挡是模型行为诊断非现实因果证明，必须保留真实classifier/XAI调用，不用LLM自评代替。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02523", "title": "Hypothesis-guided discovery of cognitive algorithms via program refinement", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/717", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1180" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/709" ], "reason": "审查通过，待核心算法实现与公开验证。从人写FlipPy认知程序出发，Game提出behavioral revision、Coding实现、Audit最多5修复、Inference算行为likelihood；846候选733可执行后语义cluster选24代表，不是全自主理论发现。heldout trial未参与生成/选择，72/137改善；synthetic recovery仅8.4%回收75%预测差距，限制明显。特定领域应用，但受约束程序改进与heldout likelihood选择可形成通用evolve算子；匿名OSF未提供可定位链接，不能声称材料已核实。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02525", "title": "Learning What to Investigate Next: Meta-Reasoning for Long-Horizon Research Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/716", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1179" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/708" ], "reason": "审查通过，待核心算法实现与公开验证。MIRA固定curation/executor仅训练外层workorder/terminate研究决策；共享actorcritic以单token离散return分布KL、决策级GAE与跨版本bootstrap bridge，ratio范围外mask policygradient但保留logratio²penalty，actor/critic独立buffer及optimizer顺序更新。四研究环境、proxy/gold分开；shared初始化来自critic故消融不隔离参数共享本身，内层tool/实验token不训练。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02542", "title": "How To Train Your World Model: Fine-tuning vs RAG for LM-based World Modeling", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/715", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1178" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/707" ], "reason": "审查通过，待核心算法实现与公开验证。LM worldmodel研究FT与RAG并提出HWM：同episode排除的历史transition检索条件NTP，query用observation/action分层检索可选；frozenpolicy统一规划流程。五模型环境20组合17胜是作者实验；oracle实际环境与top20后验gold重排只diagnostic，不能给部署策略。非constructedWM在9/20胜说明外推有限，transitionfidelity与taskutility必须分报。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02569", "title": "Pincer: Resource Authorization for Agents using a Digital Twin", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/712", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1176" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/704" ], "reason": "审查通过，待核心算法实现与公开验证。Pincer以OS完整资源中介与隔离twin为TCB；可信用户state+PRF逐路径段去语义accesspattern、content与precommittedpolicy三票合取。PRF保护只针对资源名优势，非全系统注入免疫；历史access不等于用户授权。合成两persona各单模型、条件ASR/BGR是资源gate非终局任务成功，implementation/data正文称即将发布，算法可候选但数据集不可声称已公开可运行。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02588", "title": "Open-Endedness Bench: Measuring Epistemic Process from Agent Records", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/711", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1174" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/703" ], "reason": "审查通过，待核心算法实现与公开验证。OEB将轨迹转act/proposition证据卡，substring核验引用、三次closed-question投票建epistemic图、确定性聚合；无机会abstain且<5机会不计。四能力轴与persona分离，不能总分混加。90人工audit、119已有runs/12tasks，相关非因果；median652judgecalls不宜伪称低价在线reward。官方MIT pipeline/converters及HF scored runs已核。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02593", "title": "Fisher-Guided Submodular Data Selection for Continual Pre-Training of Large Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/710", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1173" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/702" ], "reason": "审查通过，待核心算法实现与公开验证。Fisher双几何与流式submodular数据选择是明确的CPT算法候选，不以工业上线结果接入。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02597", "title": "GRAFT: Growing Agglomerative Foundation Models via Continual Teacher Distillation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/709", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1172" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/701" ], "reason": "审查通过，待核心算法实现与公开验证。GRAFT以历史student单anchor保存旧教师功能，TSRT与image-text关系KL接纳新视觉teacher，是定义清晰的持续蒸馏operator。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02614", "title": "What Is Lost in Post-Training? Default Collapse and the Loss of In-Context Steerability Across Diverse Perspectives", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/705", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1570" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/697" ], "reason": "审查通过，待核心算法实现与公开验证。SDM以相关性judge筛prompt、stance judge分桶，在维度群体分布上惩罚目标偏离并PPO；Llama8B两设置初步训练，不能把群体分布约束说成逐prompt保证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02616", "title": "VERSE: Verified Self-Evolving Optimizer for Agent Harnesses", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/704", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1168" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/696" ], "reason": "审查通过，待核心算法实现与公开验证。VERSE在固定权重harnessoptimizer上增加trace minimization、draft验证双通过、replay/perturb、失败模式audit与optimizer自身prompt/tool/hook更新。110train50val108Marchtest与107JulyOOD隔离；每行一次evolution、test重复3次，不是3个独立搜索seed，compute不完全相同。静态验证器/预算及val选择要保留；不能把单次flip当修复。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02617", "title": "WebUIProof: Benchmarking WebUI Code Generators with UI-Agent Execution Harness", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/703", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1167" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/695" ], "reason": "审查通过，待核心算法实现与公开验证。WebUIProof以DOM+视觉UI agent执行测试，build failure计零、Full+0.5Partial聚合，判定依赖grounding/stepbudget，100例人工一致率约80%非确定oracle；VisRL组合截图/交互/analyzer reward PPO。官方公开149general任务和build/eval/compute_acc代码，但论文219含70个3D任务未见repo对应，复现先限定149，RL/3D不能默认已开源。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02638", "title": "Batched Speech Decisions Without Decoding: Single-Token Supervision Lets a Frozen LLM Hear Beyond the Transcript", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/701", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1165" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/693" ], "reason": "审查通过，待核心算法实现与公开验证。DuplexJev冻结ASR/LLM只训connector，内容transcriptKL与决策单letterCE混合；typed选项随机位置；共享prefix KV+blockdiagonal suffix mask+restartposition精确批量logits读出。A多层fusion不胜B末层，speaker-disjoint真人语音，情绪训练ZJUcontent仍降6–17点、混loss非全解决。流式160–200ms仅部署构想未测。可纳语音决策/批量readout基础算子非完整duplexagent。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02657", "title": "Context-Tower Conversion Preserves Generation While Freezing Retains Knowledge: Low-Budget AR-to-Diffusion Conversion of MoE LLMs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/699", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1163" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/691" ], "reason": "审查通过，待核心算法实现与公开验证。低预算AR转diffusion通过frozencontexttower和layerKVconcat有匹配token/参数协议与公开任务证据，作为研究型转换operator而非serving加速接入。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02666", "title": "CHASE-VLA: Post-Training Quantization Framework for Vision-Language-Action Models with Chunk-Aware Scale Estimation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/697", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1161" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/689" ], "reason": "审查通过，待核心算法实现与公开验证。CHASE-VLA以先前生成action chunk预测逐层/去噪step组的激活尺度分布，分位损失与方差联合训练尺度估计器，保持W4A4 action expert。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02670", "title": "LEAP: Learning Efficient Action Proposals For LLM Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/696", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1160" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/688" ], "reason": "审查通过，待核心算法实现与公开验证。LEAP target canonical-action LoRA NLL训练tinydraft，无targetlogits；克隆环境先proposal，再同GPU并行targetverify、只提交匹配前缀并移交克隆状态不重复工具。OpenAGI/TaskBench无工具执行，tau2/BFCL为localbackend；1.10–1.63x whole-task speedup有任务配对bootstrap。真实外部副作用不能安全clone则不适用，targetauthorship不保证并发下trajectoryidentity。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02683", "title": "RAOA: Alternating-Operator Neural Computation with Programmable Radio Propagation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/695", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1159" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/687" ], "reason": "审查通过，待核心算法实现与公开验证。RAOA交替归一化能量梯度problem步和Hadamard mixer步，重复共享两个控制参数；零初始化残差adapter可在冻结LLM中研究递归结构。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02687", "title": "Decoupling Memory from Context: Structured Memory for Token-Efficient Test-Time Continual Learning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/694", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1158" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/686" ], "reason": "审查通过，待核心算法实现与公开验证。GraphMemory对ACE新增compact index entry选择与有向证据图，2hop/12nodes/2500tokens BFS，Beta(1+help,1+harm)均值排序，生成归因仅允许已载入节点，Curator显式创建/删除边。不是端到端O(1)，index随库增长；FiNER/Formula val选memory、heldout test，省token但query约1.8倍慢且部分准确率低，不能宣称全面优越。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02695", "title": "Test-time Calibration Learning for Large Language Model Reasoning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/693", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1569" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/685" ], "reason": "审查通过，待核心算法实现与公开验证。TTCL按答案一致性构造pseudo正确性/置信目标，reasoning-answer和confidence token分离优势，逐epoch冻结EMA目标cache；同benchmark无标签适应再评测，属于transductive TTT非独立未见测试。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02697", "title": "GeoScaffold: Learning Compact Geometric Latents via Reconstruction for Efficient Vision-Language Navigation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/692", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1157" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/684" ], "reason": "审查通过，待核心算法实现与公开验证。GeoScaffold 先训练冻结深度 VQ-VAE，再用八个几何 query 重建深度码/深度图/连通性/可通行性，联合模仿学习；部署移除重建头与 tokenizer。R2R-CE/RxR-CE 可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02700", "title": "Learning from Evolving Errors: Adaptive Iterative Repair for On-Policy Distillation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/691", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1568" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/683" ], "reason": "审查通过，待核心算法实现与公开验证。AIR-OPD按首个错误anchor开始，保留1-epsilon纠正概率质量的块边界repair区域；多轮指导teacher特权蒸馏和gamma深度/rho即时成功归一权重；参考答案只训练分支，实际DAPO训练。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02710", "title": "Self-Supervised Scaling of Terminal Environments for Scientific Domains", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/689", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1155" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/681" ], "reason": "审查通过，待核心算法实现与公开验证。SWR将500真实workflow包装成3000任务，public IO/hidden judge隔离；typed scenario、replay反例与hidden完美一致筛SFT，重复采样只是重权重非新轨迹。SWR100是同workflow的新task不是unseen-workflow；35k开发校验与独立200case冻结后audit分开，2%误接/误拒；oracle public-output选择仅反捷径诊断，不能供agent。方法可复用，正文未给可定位作者完整释放入口。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02713", "title": "WakeKV: Reactive, Reversible KV Residency for Heads That Change Their Minds", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/688", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1154" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/680" ], "reason": "审查通过，待核心算法实现与公开验证。WakeKV保存完整CPU reservoir，GPU按头LRU驻留并在需求变化时可逆取回，适合缓存迁移实验候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02718", "title": "Revisiting Visual Representation Enhancement of VLMs via Kernel Canonical Correlation Analysis", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/687", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1153" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/679" ], "reason": "审查通过，待核心算法实现与公开验证。3vKCCA将CLIP视觉、冻结DINO与文本作三视图投影训练，可作为带理论限制的视觉对齐候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02740", "title": "Prospective Hindsight: Self-Calibrating Reinforcement Learning via Prediction-Reality Gaps", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/686", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1152" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/678" ], "reason": "审查通过，待核心算法实现与公开验证。PH同policy共享参数自评只见state/action不可见privilegedfuture，ternary多数票与实际结果不一致则stopgrad(1+alpha)乘base RL/OPD/Combined loss；未知类中性。OLMo3-7B单轮verifier/多轮PRM+用户模拟、随机同质量reweight控制。校准解释非保证，alpha2可坍缩成预测失败；额外M forward训练开销必须计，不能让测试gold流入自评。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02781", "title": "OPD Before RL: Warm-Starting Rubric-Based RL with On-Policy Distillation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/682", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1565" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/674" ], "reason": "审查通过，待核心算法实现与公开验证。RP-OPD先用rubric特权teacher top256 forward-KL暖启动，再切换rubric reward GRPO；三rubric领域多学生。HealthBench正文另列重叠/不重叠评价，正式复现须使用完全隔离数据。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02814", "title": "VeriPy Source-Preserving Verification and Compatibility Checking for Python Components", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/681", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1149" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/673" ], "reason": "审查通过，待核心算法实现与公开验证。VeriPy 保留Python executable AST，显式admission/environment将spec直接降Dafny或Lean，proof sidecar与axiom gate、hash-bound guards、原body/model differential tests及scoped compatibility。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02828", "title": "FSPO: Policy-Consistent Risk and Pareto-Feasible Control for Budgeted LLM RL Post-Training", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/679", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1563" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/671" ], "reason": "审查通过，待核心算法实现与公开验证。FSPO policy-consistent risk-to-go ensemble、decision-conditioned calibration与Pareto continuation证书控制训练动作；正文区分24case ledger机制与Qwen3B真实GRPO trainer研究。接纳机制，不能把固定ledger效率当LLM能力提升。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02832", "title": "FastOPD: On-Policy Distillation for Lightweight VLA Deployment", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/678", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1148" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/670" ], "reason": "审查通过，待核心算法实现与公开验证。FastOPD单次学生flow-map跳跃采样on-policy状态匹配teacher对角velocity，并用中点自一致传播到有限步map；LIBERO/RoboTwin及真机器人，单步action-expert训练必须真checkpoint。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02835", "title": "All Work And No Play Makes Jack a Dull Boy: Understanding and Preventing Catastrophic Strategy Collapse in RLVR", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/677", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1562" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/669" ], "reason": "审查通过，待核心算法实现与公开验证。MeshLearning固定离线coach策略前缀（无中间计算和答案）探索，对正确轨迹相对pre-RLVR reference的策略logprob增长做uniform-KL均衡；推理时无coach，真实数学科学代码训练。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02840", "title": "PointWAM: 3D World Action Modeling for Dexterous Robotic Manipulation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/675", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1146" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/667" ], "reason": "审查通过，待核心算法实现与公开验证。PointWAM统一场景/手3D轨迹预训练再动作retargeting，符合机器人基础模型候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02841", "title": "How Robust Is Multimodal Claim Verification to LLM Rewriting?", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/674", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1145" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/666" ], "reason": "审查通过，待核心算法实现与公开验证。SciClaimEval同claim配支持/反驳图像，比较自然改写和单词注入引起的verdict变化，可作为多模态判别与reward鲁棒性诊断。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02856", "title": "Adaptive Mutual Distillation for Balanced Multi-Task Post-Training of Large Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/672", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1561" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/664" ], "reason": "审查通过，待核心算法实现与公开验证。AMD对MFT/平滑SFT两分支互蒸馏，三类全局短probe从相同状态试增减/保持权重，丢弃probe参数后按各任务方向validation选择；6基准多骨干实测，probe成本必须计入。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02876", "title": "Seeing, Saying, but Not Using: From Reportable Spatial Facts to Usable States in Multimodal Large Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/671", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1143" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/663" ], "reason": "审查通过，待核心算法实现与公开验证。OSS 对同一空间事实规范化并跨上下文对齐，以答案、状态对齐、状态轨迹三池标准自回归训练。派生样本继承 observation world 的原始切分，可验证 SpaceConflict 三类与成对评分。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02880", "title": "Found but Not Read: When Extracted Text Closes the Retrieval-Reading Gap in Document Vision-Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/669", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1141" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/661" ], "reason": "审查通过，待核心算法实现与公开验证。FoveDoc-Bench固定ColQwen2 top16页，对比同页图像与图像+CPU OCR/PDF text，lexical top16块各300字符；适合区分检索与读取瓶颈的受控诊断。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02885", "title": "PsyEvo: A Personalized Counseling Agent That Self-Evolves at Test Time", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/668", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1140" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/660" ], "reason": "审查通过，待核心算法实现与公开验证。PsyEvo HBSP固定population skill prior加client Gaussian posterior采样，SOCA双顺序agreement DAG选回复与mean-preserving ordinal delayedcredit；session间group-normalized weighted DPO更新shared LoRA。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02887", "title": "Revealing Epistemic Uncertainty in MLLMs via Causal-Invariant Masking", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/666", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1139" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/658" ], "reason": "审查通过，待核心算法实现与公开验证。CIM 对保留问题相关 ROI 的遮罩输入比较语义分布，以 EED 期望嵌入漂移估计不确定性。VQAv2/OKVQA/AdVQA/POPE/MME 有公开数据与等采样预算对照。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02925", "title": "Positive-Unlabeled Learning for Agent Safety False Alarm Auditing", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/664", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1137" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/656" ], "reason": "审查通过，待核心算法实现与公开验证。Safe-alarm PU ranking以OOF role boundary transfer TN、mass-constrained soft-protection减弱误警负梯度；frozen references median rank/reliability-gated distillation，hierarchical TN与cross-alarm约束KL rank refinement。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02928", "title": "Discriminating Fixture Coverage in Agent-Infrastructure Verification Suites", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/663", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1136" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/655" ], "reason": "审查通过，待核心算法实现与公开验证。Discriminating Fixture Coverage 冻结projection suite/hash后对外部mutants测kill，instrument区分未激活与污染未可见，按预测七input维度加fixture、原oracle不改复测。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02951", "title": "Dynamic Expert Pruning for Multi-Agent Systems", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/662", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1135" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/654" ], "reason": "审查通过，待核心算法实现与公开验证。DEP已在既有队列：请求级文本encoder+MLP预测逐层expert topK，影响标签BCE与softmask自蒸馏KL联合训练，CPU专家池向GPU槽位差量加载；维持候选不重复新增。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02952", "title": "GTDD: Generative Test-Driven Development for AI Coding Agents with Adversarial Testing", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/661", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1134" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/653" ], "reason": "审查通过，待核心算法实现与公开验证。GTDD candidate/config先commit，source-informed或blind tester生成valid input不生成期望输出，可信reference差分、最短prefix/operation删减、最多k反馈；fresh审计与开发tests分离。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02970", "title": "A Guideline-Augmented Multi-Agent Framework for Schema-as-Code Biomedical Named Entity Recognition", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/659", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1132" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/651" ], "reason": "审查通过，待核心算法实现与公开验证。GAMA 从train gold归纳annotation rules并按train支持验证，TopN/type guideline memory→ranked span/type plan→schema-as-code→grounding/type/structure双环修复。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.02986", "title": "OLMo-Detect: A Multi-Stage, Confounder-Controlled Benchmark for Membership Inference on Large Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/657", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1559" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/649" ], "reason": "审查通过，待核心算法实现与公开验证。OLMo-Detect是跨pre/mid/post训练membership审计任务，member/nonmember按质量、时间、词汇匹配，15无监督/3监督MIA与dev/test分离；官方仓库有数据和评分pipeline。评估成员识别/污染风险，不等同提升LLM推理。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03002", "title": "Recursive Self-Improvement in Unified Multimodal Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/655", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1129" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/648" ], "reason": "审查通过，待核心算法实现与公开验证。统一多模态RSI以执行验证的绘图程序产生训练真值，诊断仅决定数据分布，分别更新生成/理解专家再合并，具备同训练步数与同verified数量控制，可作为多模态自改进研究候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03014", "title": "Beyond Predefined Sinks: Security-Aware Dependency Analysis for LLM Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/653", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1127" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/646" ], "reason": "审查通过，待核心算法实现与公开验证。AgentSecGraph permissive sensitive-op anchor→framework normalization→typed Security-ADG，Python AST backward trace识别覆盖constant切断，JS lexical trace，保留guard/trust-boundary provenance。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03020", "title": "DyadMem: A Long-Term Memory Benchmark of How Agents Work with Users", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/651", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1125" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/644" ], "reason": "审查通过，待核心算法实现与公开验证。DyadMem 分Capture/Update/Recall与full pipeline，six memory types含agent-user协作URAM，chronological delete/add/dedup，terminal valid-bank recall与grounded QA并测。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03034", "title": "Adaptive Second-Order Solvers for Fast Stochastic Diffusion Sampling", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/648", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1122" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/641" ], "reason": "审查通过，待核心算法实现与公开验证。PI 扩散求解器基于 stochastic Heun 一二阶误差和前次误差调步，将自适应轨迹聚合为静态模态专用时间表。FFHQ/ImageNet64、LM1B 与公开 EDM 检查点可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03036", "title": "WebFovea: When the Model Is Right but the Click Is Wrong -- Reliable Round Trips for Vision-Based Web Agents on Live Websites", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/647", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1121" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/640" ], "reason": "审查通过，待核心算法实现与公开验证。WebFovea hardens视觉action roundtrip：解析特殊token、坐标space映射、select/fill实际读回、反馈状态、五截图history与bounded DOM text通道，executor约束动作预算。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03039", "title": "HyperThink: Text-to-Parameter Hypernetworks for Efficient Reasoning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/646", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1120" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/639" ], "reason": "审查通过，待核心算法实现与公开验证。HyperThink以查询生成冻结LLM层bias并经VQ约束，保留独立算法候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03063", "title": "HARPO: Hallucination-Aware Reinforcement Learning for Faithful and Creative Language Generation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/642", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1558" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/635" ], "reason": "审查通过，待核心算法实现与公开验证。HARPO训练hallucination-span reward model，只有预测零幻觉才激活写作reward，并配创作到忠实cosine课程；属层级多目标reward机制，不是新的GRPO；gate不保证真实无幻觉。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03084", "title": "NegT2IBench: When Negation Changes the Picture. A Polarity Benchmark for Text-to-Image Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/641", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1116" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/634" ], "reason": "审查通过，待核心算法实现与公开验证。NegT2IBench以六类4800个正负属性/关系prompt和对象存在性先决条件，用detector/SAM/depth/SigLIP2逐条判分；可补图像生成的否定遵循评估。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03089", "title": "Securing Computer-Use Agents Against Branch Steering Attacks", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/640", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1115" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/633" ], "reason": "审查通过，待核心算法实现与公开验证。COBRA trusted-input预规划并承诺程序，BRH以先定branch/field约束解析观察，HTTP/MCP代理强制operation、参数、approved schema hash；fused5plans保留合法alternative。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03099", "title": "Beyond Single Videos: Benchmarking and Active Evidence Seeking for E-Commerce Cross-Video Reasoning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/638", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1113" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/631" ], "reason": "审查通过，待核心算法实现与公开验证。AdsCVR/AdSeek 多视频frame/audio选择逐轮证据，GRPO correctness+.1format，过滤no-tool/zero-var group并最多三次resample；teacher诊断missing evidence/reasoning→定向rectify→assistant-mask SFT。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03102", "title": "Ask, Relax, or Act? Evaluating Actionable Indeterminacy in LLM Preference Reasoning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/637", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1112" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/630" ], "reason": "审查通过，待核心算法实现与公开验证。Ask/Relax/Act 以admissible preferences/objectives下accepted action交集判定可行动/需问，infeasible则最小允许repair；四finite结构exact solver生成matched intervention pairs。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03123", "title": "Foresight: planning future perception in streaming VLMs without retraining", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/634", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1110" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/627" ], "reason": "审查通过，待核心算法实现与公开验证。Foresight 将共享冻结模型分为写入视频的 Ingest 与只读快照的 Think；规划采样率、保留类别、压缩和响应时间。StreamingBench/OVO 允许真实因果流式验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03190", "title": "Not Until the Evidence Says So: Teaching LLM Investigators When to Close a Case", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/627", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1555" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/620" ], "reason": "审查通过，待核心算法实现与公开验证。Nautil提供证据索引/request工具、close/open任务、来源内balanced accuracy及删关键证据对照；官方MIT代码/数据/模型公开。RLVR checkpoint选择用过64测试中22条，必须改验证集隔离，不能照搬其headline为无泄漏泛化。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03198", "title": "KV$^2$: A Self-Refining KV Cache", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/624", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1102" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/617" ], "reason": "审查通过，待核心算法实现与公开验证。KV²用KeyDiff跨层/头均值选chunk内稀疏重构query，append在完整cache后做真实因果前向；局部max-attention评分合成全局保留，也可迭代替换proxy。已在原队列，补全证据不重复计新文。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03199", "title": "Predicting and Repairing Merge Collapse in Large Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/623", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1101" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/616" ], "reason": "审查通过，待核心算法实现与公开验证。Prism对norm缩放后平均taskvector按跨专家方差softthreshold，并用cancellation/dispersion门避免删独有知识；确定性数据自由repair定义完整。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03213", "title": "Toward SLM-based agentic task-tool intent matching", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/622", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1100" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/615" ], "reason": "审查通过，待核心算法实现与公开验证。Gemma task-tool relevance 以Base→GEPA→LoRA SFT→DrGRPO curriculum，deterministic格式/标签/长度reward，候选逐工具Boolean判断。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03226", "title": "D2K-Bench: Can LLM Agents Turn Expert Designs into Efficient GPU Kernels?", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/618", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1097" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/612" ], "reason": "审查通过，待核心算法实现与公开验证。D2K-Bench 配对26task/85workload有无专家L1–3设计guidance，agent iterative Triton编辑/compile/profile/submit；数值与stateupdate、implement validity全通过才按geomean speedup选候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03315", "title": "Lightweight, Rubric-Guided Trajectory Evaluation for Production AI Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/615", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1094" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/609" ], "reason": "审查通过，待核心算法实现与公开验证。LiteTrajEval 离线domain rule profiles，在线规范化trace/弱failure hints，格式compaction+重复引用、tier waterfall预算与MMR证据选取，单rubric judge输出故障诊断。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03389", "title": "From Patching to Pruning Visual Computation in Vision Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/610", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1091" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/605" ], "reason": "审查通过，待核心算法实现与公开验证。P2P 用视觉投影缓存均值替换分析，前缀/后缀敏感度扫描选择安全层，真实跳过视觉 attention/MLP，可结合 patched-graph LoRA。公开七基准与独立 proxy/validation/test 可复现。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03448", "title": "Passing the Test You Trained On: Re-evaluating Prompt-Injection Detectors for LLM Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/607", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1088" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/602" ], "reason": "审查通过，待核心算法实现与公开验证。重评15 injection detectors：差分replay构造clean/injected tool输出，按default阈值/lowFPR测TPR，审训练重叠，并用task-aware yes/no logprob judge诊断缺失任务语境。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03516", "title": "XGenAct: Geometry-Enhanced World Action Models through Cross-Task Generation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/599", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1083" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/595" ], "reason": "审查通过，待核心算法实现与公开验证。XGenAct确定性几何codec与共享视频flow任务混合，可作world-action联合训练候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03524", "title": "From Benchmarks to Production: A Text-to-SQL System for Complex Financial Data", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/598", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1082" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/594" ], "reason": "审查通过，待核心算法实现与公开验证。FLINT lookup从reference表解析opaque ID，检索专家query bank结构示例，FK-chain schema linking→SQL→执行error reflection；多步骤并行。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03574", "title": "HyperBrowseComp: A Multilingual and Multimodal Stress Test for Web-Browsing Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/594", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1078" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/590" ], "reason": "审查通过，待核心算法实现与公开验证。HyperBrowseComp 423人工多语/多模态证据链短答，以no-internet过滤容易题；统一ReAct25步与provider-native/Exa/OWL三条件，保存实际web trace和短答judge。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03585", "title": "Threat-Preserving Representation Sensitivity in Agent-Security Benchmarks", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/593", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1077" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/589" ], "reason": "审查通过，待核心算法实现与公开验证。TPRS 固定任务、harm objective、policy、scorer与环境，仅替换tool representation，对ASB/MCPTox/AgentDojo分别量化ASR与benign utility。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03591", "title": "HazardWeaver: Scientific Route Selection for Hazard Analysis Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/592", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1076" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/588" ], "reason": "审查通过，待核心算法实现与公开验证。HazardWeaver 从文献编译Satisfied/Violated/Unresolved科学条件，以typed capability graph找可执行route，LLM在两者交集选择/修复，独立hidden route/时点/provenance/预算评分。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03631", "title": "NeutronGym: Physics-Graded Neutron Instrument Design for LLM Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/587", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1070" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/583" ], "reason": "审查通过，待核心算法实现与公开验证。NeutronGym 22 MCP tools管理McStas真实仪器、缓存编译后参数仿真；syntax/runtime/structure/science四级reward，程序化族经过无模型规则及固定答案shortcut admission，再GRPO训练。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03632", "title": "World Embedding Benchmark", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/586", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1069" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/582" ], "reason": "审查通过，待核心算法实现与公开验证。World Embedding Benchmark公开物理视频/文本检索、回归、pair判别及新模型评测入口，适合作为embedding物理理解公共评测基础设施。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03695", "title": "Language Models that Play Chess and Explain Their Moves", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/582", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1066" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/581" ], "reason": "审查通过，待核心算法实现与公开验证。Queen 连接冻结 Leela 棋局编码器与 SmolLM3，先四阶段空间/时间课程，再依据 Stockfish 核验的递归子局解释进行七轮自进化 SFT。Lichess/公开棋引擎与持出棋题可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03741", "title": "Logit-Aware MIMO AirComp for Distributed Mixture-of-Experts LLM Inference over Wireless Edge Networks", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/579", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1443" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/579" ], "reason": "审查通过，待核心算法实现与公开验证。以有限差分估计MoE聚合块对next-token logits敏感度，驱动有功率约束的MIMO AirComp收发器加权优化，可做无线噪声条件下的模型机制诊断。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03780", "title": "OceanMind: A multi-agent AI system for ocean diagnosis", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/577", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1415" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/577" ], "reason": "审查通过，待核心算法实现与公开验证。OceanMind 六角色按 query routing→skill planning→确定性工具编排→证据综合分析海洋状态，缺工具由受限 CodingAgent 补充，数据 artifact 通过引用与 role-specific memory 传递。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03796", "title": "ProsaBuddy: Assisting Mechanized Real-Time Schedulability Analysis with LLM-based Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/575", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1320" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/575" ], "reason": "审查通过，待核心算法实现与公开验证。ProsaBuddy primary两阶段lemma decomposition+领域guidelines/skills→隔离subagents，Rocq结果在独立OS进程AST审计：proofbody仅ProofStep/Query，外部只允许trustedimports，库readonly，禁止改theorem/axiom。17lemma成功、去skills12/去decomp14/去paperhints15；完整headline含人工proofhint必须分为hint-assisted不当无gold能力。Audit TCB超kernel未形式验证，难题弱模型省略非全矩阵。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03820", "title": "From Requirements to Attack Trees: Grounded LLM Agents for Design-Time Security Review", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/574", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1150" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/574" ], "reason": "审查通过，待核心算法实现与公开验证。从需求与架构图抽取带信任边界的实体图，按目标分派攻击树 worker；联合验证后只重建受影响分支，保留稳定 ID，最多三次。公开 Microsoft IoT Edge 威胁模型可作为参考任务，核心流程可进入安全 Agent 实现队列。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03894", "title": "Proxy Confidence: Auditing Black-Box LLM Agents with a Surrogate's Log-Probabilities", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/572", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1064" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/572" ], "reason": "审查通过，待核心算法实现与公开验证。ProxyConfidence 用独立开放 surrogate 的 teacher-forced/PMI与YES-NO/tool竞争logprob为blackbox actor行动打分，按regime校准方向后融合，并用于audit/修正选择。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03938", "title": "MLLMs Fail to Refuse when Using Tools Agentically", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/568", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1060" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/568" ], "reason": "审查通过，待核心算法实现与公开验证。同一MLLM在no-tool与ReAct四视觉工具条件下配对，测拒绝失败；按tool次数诊断context dilution，并逐轮重注入原请求比较缓解。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03984", "title": "Teaching Agents to Code Reliably", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/565", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1545" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/565" ], "reason": "审查通过，待核心算法实现与公开验证。在mutation/sibling补丁上给测试断言按加权coverage Shapley credit，门控正轨迹优势再注入assertion token，属新credit机制；gold tests仅scorer可见。全500例先诊断后230训练270holdout，不应宣称完全未接触测试。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.03994", "title": "Solving VeriContest with a Lean-Backed Rust Verifier", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/564", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1057" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/564" ], "reason": "审查通过，待核心算法实现与公开验证。Rust-Prover 将 Verus spec 转 Rust specification、由 Charon/Aeneas 转 Lean，proof agent给定不变代码/theorem迭代证明，干净clone复核byte identity与Lean axioms。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04002", "title": "Do Language Models Need a Trainable Input Embedding Table? Fixed Minimal Token Codes at 1.7B-Class Scale", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/563", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1056" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/563" ], "reason": "审查通过，待核心算法实现与公开验证。固定最小二进制token身份+确定lift可作为可执行架构/机制诊断变体，公开checkpoint；不作为质量或效率优胜算法。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04008", "title": "SkillScriptBench: Benchmarking Self-Evolution of Executable Agent Skill Packages Beyond Markdown", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/562", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1055" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/562" ], "reason": "审查通过，待核心算法实现与公开验证。SkillScriptBench评估SKILL.md和脚本一致性，AST-guided revision绑定需求到唯一语法节点，局部修改并显式保留正确行为，适合可执行skill/RSI迭代候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04009", "title": "SUAVE: Unified Video-Action Models via Masked Diffusion", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/561", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1054" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/561" ], "reason": "审查通过，待核心算法实现与公开验证。SUAVE统一离散视频/动作masked diffusion为真实新模型训练候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04047", "title": "Periscope: Extending Frozen Language Models Beyond Their Context Window", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/559", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1052" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/559" ], "reason": "审查通过，待核心算法实现与公开验证。Periscope明确给出冻结模型的局部/跨距探测与证据图，可作为长上下文推理和retrieval算子候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04118", "title": "LongSocialBench: Do Long-Context LLMs Understand Online Discussion Threads?", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/555", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1046" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/555" ], "reason": "审查通过，待核心算法实现与公开验证。LongSocialBench公开1462题/94完整讨论及精确MCQ协议，长上下文结构化社会推理覆盖有价值；只纳入无gold推理评测，oracle单列diagnostic。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04132", "title": "PB-GRPO: Learning Socially Adaptive LLM Agents from Persona-Driven Simulation with Preference-Batched GRPO", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/553", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1544" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/553" ], "reason": "审查通过，待核心算法实现与公开验证。PB-GRPO保持prompt-local均值，但在同潜在偏好bucket内合并方差作为分母；128训练persona/100未见persona真实Qwen7B，初步单设置证据需保守。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04134", "title": "Agentic Resource Allocation for Batch Multi-Objective Bayesian Optimization in Autonomous Materials Discovery", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/551", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1043" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/551" ], "reason": "审查通过，待核心算法实现与公开验证。MultiTaskGP 在每轮 qEHVI 与 diversified qUCB 六种配比之间分配实验batch；LLM根据近期HV/MI、历史和预算事件选择option，Adaptive还选β。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04139", "title": "From Sight to Foresight: Predictive Spatial Reasoning in Vision-Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/549", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1041" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/549" ], "reason": "审查通过，待核心算法实现与公开验证。SpatialMind 以有界对数深度修正、度量和运动编码对齐视觉 token，分 GROUND/CURRENT/DYNAMICS/FUTURE 训练状态链与答案。外部 VSI/OSI/VLM4D 支持迁移检验。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04140", "title": "ExpertMuon-Compass: Alignment-Guided Step Sizes for Mixture-of-Experts Training", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/548", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1040" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/548" ], "reason": "审查通过，待核心算法实现与公开验证。ExpertMuon-Compass保持Muon矩阵方向，以familyalignment和rowalignment标量半径调整expert步长，独立tuningseed与配对评测充分定义训练operator。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04152", "title": "Kepler4D: Controllable Future Video Generation via 4D Scene State Evolution", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/547", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1039" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/547" ], "reason": "审查通过，待核心算法实现与公开验证。Kepler4D 通过 DA3、SAM2 和 VLM 获取观测对象代理，Chain-of-Motion 产生结构化运动并渲染 3D 控制给 VerseCrafter。DL3DV/DAVIS 可做条件化延续实验。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04178", "title": "SHarP: Saliency-based Pruning of Agent Harnesses", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/543", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1036" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/543" ], "reason": "审查通过，待核心算法实现与公开验证。SHarP 对手工定义 harness 模块逐一禁用，以 dev 任务分数/成本的相对中位数及 paired t-test 判断 saliency；保护显著性能贡献模块，按效率信号剪枝后独立评测。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04196", "title": "Asynchronous Is Nearly Free for Evolution Strategies on Long-Horizon Agentic Tasks", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/542", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1035" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/542" ], "reason": "审查通过，待核心算法实现与公开验证。AsyncES 以参数版本记录 rollout lag，接受 bounded-staleness 回报、按 accepted cohort 标准化奖励并更新中央 ES 参数；worker 恢复无扰动权重并重放更新后再扰动。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04204", "title": "Clean: Second-order LLM Training at Linear Memory Cost via Nyström Sketching", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/541", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1034" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/541" ], "reason": "审查通过，待核心算法实现与公开验证。Clean/Q-Clean以Nyström双侧薄因子及四互补块moment降低优化器状态，定义足够明确且有LLM预训练实测，可作为新optimizer候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04225", "title": "FlashGaze: Training-Free Multi-Scale Patch Pruning For Efficient Video Understanding", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/540", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1032" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/540" ], "reason": "审查通过，待核心算法实现与公开验证。FlashGaze 在视觉编码前按运动补偿误差选择 Drop/Merge/Keep，以三层四叉树动态规划满足像素预算。Qwen3-VL/NVILA 与公开视频基准可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04261", "title": "Playing social deduction games with reinforcement fine-tuned large language models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/537", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1542" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/537" ], "reason": "审查通过，待核心算法实现与公开验证。社交推理 RFT 在狼人/Avalon 对同阵营共享policy训练，私有可见角色+公共历史形成prompt；用隐藏角色预测和后续投票影响奖励，终局win作外部观察。明确游戏引擎可构建验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04292", "title": "LMBuild: Evaluating LLM Agents for Generating Buildable and Functional Structures", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/532", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1027" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/532" ], "reason": "审查通过，待核心算法实现与公开验证。LMBuild用可检索/生成零件的交互工具生成3D结构，按结构、功能、设计和物理实现分层评测，适合多模态coding-agent能力候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04299", "title": "Questioning the Questions: Sustaining Self-Evolution in Reasoning Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/531", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1539" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/531" ], "reason": "审查通过，待核心算法实现与公开验证。R-Quest有效性初始化与采样成对新颖性门控改变questioner–solver自演化反馈；真实两家族实验。但正文按每域最高测试benchmark均值选checkpoint，复现必须另设validation，不能继承该选择为无偏正式成绩。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04303", "title": "What to Preserve in Recursive Computation: A Local Predictive Sufficiency Principle", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/529", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1025" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/529" ], "reason": "审查通过，待核心算法实现与公开验证。局部预测充分性用完整协方差平衡的ridge有限测试估计压缩缺损，并将真实optimizer proposal投影到分组预测约束半空间；覆盖语言记忆、TGN、VLA及递归策略，具备新可训练机制与对应控制。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04318", "title": "Rethinking Long-Video Efficiency: A Joint Allocation Perspective on Frames, Pixels, and Front-End Latency", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/528", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1024" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/528" ], "reason": "审查通过，待核心算法实现与公开验证。LoHi把一次解码的同一帧池分成低分辨率视频和少量高分辨率图像，统一token与前端解码预算；SemDiv用query相关性加DPP挑高分帧。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04328", "title": "Bidirectional Preference Synthesis: Learning Prompt-Conditioned Preferences from Boundary Failures", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/527", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1538" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/527" ], "reason": "审查通过，待核心算法实现与公开验证。BPS是独立离线偏好构造算法而非新loss：从部分失败生成student修正与安全/非泄漏achieved prompt，cycle-consistency验证后同一response pair在两prompt反向成crossed anchor，再标准DPO。IFBench及function-call有机制/下游实验；不把teacher同源的ranking诊断误当独立泛化。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04344", "title": "Hierarchical Credit Assignment for RLVR on Fused Gromov-Wasserstein Geometry", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/526", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1537" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/526" ], "reason": "审查通过，待核心算法实现与公开验证。HarA用正确/错误轨迹各自hidden-state位置图FGW重心，按span到重心贡献赋予正/负不同幂credit再乘GRPO优势；真实多模型多基准，不能以欧氏token分数替换FGW。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04371", "title": "Functionally Equivalent or Not? Graph-Grounded Differential Surrogate Execution for Code Equivalence", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/524", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1022" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/524" ], "reason": "审查通过，待核心算法实现与公开验证。FEAgent 基于 repository图义务与盲化双程序 surrogate 轨迹提出可执行反例，由确定性对账返回 Equivalent/Inequivalent/Unclear。EquiBench/SWE-bench Verified 可做二次审计。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04375", "title": "Do Tool Calls Execute as Intended? Measuring and Repairing Intent-Execution Correspondence in LLM Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/523", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1021" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/523" ], "reason": "审查通过，待核心算法实现与公开验证。IEC 以各跳接收器 parser 检测 intent-to-execution 第一个变异 hop，并单跳修复重放确认；IntAct 为对应 wrapper/process 边界生成不被改写的传递形式或拒绝。公开合成payload/CLI路径可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04379", "title": "AgentPersonaBench: Benchmarking Persona-Driven User Simulation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/521", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1019" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/521" ], "reason": "审查通过，待核心算法实现与公开验证。AgentPersonaBench用2460任务跨survey/chat/web/app检验persona约束是否转化成行为，自动泄漏检查加独立人审、确定性与LLM判据分报，适合用户模拟Agent评测候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04381", "title": "Beyond Plausibility: Verifiable Fine-Grained Image Editing on Structured Assets", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/520", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1018" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/520" ], "reason": "审查通过，待核心算法实现与公开验证。VeriEdit-Bench用可编辑SVG/chart/HTML源渲染确定目标与保护区，Lab距离分解fidelity/preservation/localization/magnitude及逐例调和分，可补图像编辑可执行评估。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04409", "title": "Understanding and Mitigating Hallucination Escape in Tool-Using LLM Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/518", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1016" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/518" ], "reason": "审查通过，待核心算法实现与公开验证。EscapeGuard 从冻结 LLM hidden state 学冲突 probe，按 selection/argument 阶段对当前 tool name/description/schema token 加 conflict强度 attention bias。BFCL/Seal/Glaive 有公开测评来源。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04429", "title": "Asking Earns Nothing: Scoring the Decision to Act in BFCL Multi-Turn", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/512", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1011" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/512" ], "reason": "审查通过，待核心算法实现与公开验证。将BFCL base与miss_func/miss_param配对，机械定位唯一empty-gold anchor；从结构化rollout辨别world-changing尝试而非lookup，计完整信息下act与缺失信息下hold的balanced accuracy。过滤坏key/多anchor/不匹配并人工认证成223pairs，strict与独立授权步骤credited分开。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04432", "title": "Video2World: Benchmarking Coding Agents for Interactive World Modeling from Embodied Videos", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/511", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1010" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/511" ], "reason": "审查通过，待核心算法实现与公开验证。Video2World将视频重建交互仿真作为coding-agent任务，分离可执行性、几何、动力学和功能oracle，公开222实例数据与评分接口可作为多模态Agent候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04488", "title": "Can LLM Agents Automate Reinforcement Learning for Text-to-Speech?", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/508", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1007" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/508" ], "reason": "审查通过，待核心算法实现与公开验证。TTS agent 在 CosyVoice2 上搜索 LM-carrier DPO 与 flow-carrier GRPO，再组合互补策略。AISHELL3/LibriTTS固定5000条件与 Seed-TTS-Eval 提供真实训练和外部观察指标。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04506", "title": "EgoExo-Next:Benchmarking Vision-Language Models on Visual-Option Next-State and Cross-View Reasoning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/503", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1001" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/503" ], "reason": "审查通过，待核心算法实现与公开验证。EgoExo-Next提供2503视觉4选1题、四个ego/exo当前匹配与未来状态任务，后续干扰帧防最远未来捷径；candidate-only/打乱序列/mismatch及CLIP/DINO检查支持诊断价值。官方HF有输入图、input_meta、分离answers与加载契约；仅输入元信息/图给模型，gold/time不给模型。注意论文§4.1两方向平均与HF样例both-directions-correct表述不一致，接入须固定评分版本并报告；源图继承六套许可、注释限非商业研究，不是统一开放商用。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04512", "title": "Correctness Is a Direction: Geometric Answer Selection in Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/501", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/999" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/501" ], "reason": "审查通过，待核心算法实现与公开验证。Geometric Answer Selection用50个标注校准问题的正确减错误平均归一hidden形成方向，对每候选一次前向投影选答案；适合冻结模型读出算子。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04517", "title": "EvoCast: Reliable Autonomous Research Agents for Iterative Forecasting Architecture Evolution", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/500", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/998" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/500" ], "reason": "审查通过，待核心算法实现与公开验证。EvoCast 将 LLM 的机制假设/源码修改与确定性 evaluator/晋升权威分开；先基线+可执行 ablation，再每轮有界源程序结构变异。TFB 公开预测任务支持真实实验。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04518", "title": "Length Generalization Needs Proper Regularization", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/499", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/997" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/499" ], "reason": "审查通过，待核心算法实现与公开验证。VPAD为每token非中心化激活计算保均值/方差的affine dropout，明确作用于RMSNorm后共享mask，具公开训练/评测/权重，可列长度泛化训练算子候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04555", "title": "$\\mathrm{TRIZ}^{a}$: Guiding Agent Evolution from Pattern Recognition to Solution Invention", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/495", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/993" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/495" ], "reason": "审查通过，待核心算法实现与公开验证。TRIZ-a 先冻结矛盾与比较协议，声明40类 typed mutation 的契约，再执行候选并根据固定参考群体的 useful/harmful 功能信息选择。公开三网络安全数据可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04575", "title": "From Probe Scores to Alarm Policies: Operational Validity of Activation Monitors for Language-Model Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/493", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/991" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/493" ], "reason": "审查通过，待核心算法实现与公开验证。按语义task/case聚合any-alarm maxima，fit/calibration/test严格分离；由独立负单位次序统计量Beta上界选择最大可认证rank并strict-exceedance报警，支持不足则abstain。observable/activation/joint同horizon比较，哈希绑定模型、阈值与一次性test，fail-closed claim compiler避免AUROC被升级成部署保证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04580", "title": "Strong Helps Weak: Directional Cross-Modal Alignment Transfer in Multi-modal LLMs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/492", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/990" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/492" ], "reason": "审查通过，待核心算法实现与公开验证。DCAT是有清晰闭式求解与公开项目代码入口的跨模态对齐迁移候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04596", "title": "DiffGate: Difficulty-Gated Teacher Guidance for On-Policy Distillation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/490", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1534" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/490" ], "reason": "审查通过，待核心算法实现与公开验证。DiffGate将teacher-student logratio以2tanh(delta/2)有界映射，并按组成功率只对失败轨迹调节teacher指导；固定8B teacher、0.6/1.7B学生数学代码训练。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04604", "title": "Anticipating the Consequences of Curriculum Decisions with Large Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/489", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1533" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/489" ], "reason": "审查通过，待核心算法实现与公开验证。LLM curriculum 根据完整 Craftax256目标文本构建跨任务关系/可行性图，与历史 learning progress 联合决定训练任务。公开环境和完整目标成功判定可训练验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04607", "title": "ForeAct3D: Policy-Grounded Future World Modeling for VLA Policies", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/488", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/988" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/488" ], "reason": "审查通过，待核心算法实现与公开验证。ForeAct3D从policy features的6queries解码当前/未来depth/3classmask/pose，未来仅接policy-generated action，训练用background staticity、按GTinstance的Kabsch rigidity、kinematics，推理不需要未来或GT。samepixel对应只是短时近似，强监督与刚体假设是实质限制；LIBERO98.3/CALVIN3.73与实机37.8%有消融。几何预测对照horizon不同且预测mask定位仍使用GTinstance，不可称无GT实例检测。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04611", "title": "Practical Runtime Enforcement of First-Order Temporal Requirements", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/487", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/987" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/487" ], "reason": "审查通过，待核心算法实现与公开验证。EnfFlash 将 MFOTL 可执行片段编译为表式 EF 因果/抑制运行时，用参考解释器及真实 trace 比较，是 agent 外部强制策略后端候选。40 formula/Agent 环境有可验证基线。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04616", "title": "PerturBot: Breaking Shortcut Priors in Vision-Language-Action Models with Perturbative Training", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/486", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/986" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/486" ], "reason": "审查通过，待核心算法实现与公开验证。PerturBot的V/C/R分别保持任务的视角干扰、状态/技能语言重标、已执行偏差运动重标，按原flow loss混合，不虚构失败目标恢复。配对GroundFscore以共享noise前向分别测null不变和causal敏感，需两条同状态expert记录标定Δaction，分母floor与过滤明确；这是诊断非任务能力证书。50种真实PnP布局与RoboTwin50任务含等预算/全因子消融，84%成功率非只离线分数。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04641", "title": "Back to the Future: Regressing Readability Features from LLMs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/483", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/983" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/483" ], "reason": "审查通过，待核心算法实现与公开验证。BTTF将AST-tagged片段embedding几何/聚类、冻结codeLM的构造级surprisal与32/256token上下文差及传统代码特征组合；跨模型稳定性共识选11特征，再以数据集等权的Ridge(alpha200)回归分位化可读性，输出具符号的特征贡献。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04646", "title": "SEIS: Self-Evolving Inference Systems", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/482", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/982" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/482" ], "reason": "审查通过，待核心算法实现与公开验证。SEIS 给 agent 私有代码 worktree、固定 correctness/throughput scorer 和预算，成功核验继承实际推理引擎/kernels；公开 Qwen 与真实 GPU 工作负载可验证代码自进化。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04651", "title": "GS-Codec: A Gaussian-Splatting Bottleneck for Neural Audio Coding", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/481", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/981" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/481" ], "reason": "审查通过，待核心算法实现与公开验证。GS-Codec以连续1D Gaussian内循环拟合潜变量+stopgradient/commitment训练编解码，再训练latent重建预测器替代300步Adam，训练后heldout校准的kmeans scalar量化参数；固定centers不传位置，bitrate仍离散。LibriTTS/LJSpeech/LibriSpeech评测需区分2339h与54kh配置，部分基线引文非重跑；L40S7.9ms仅encoder且3秒非因果片段，不是流式端到端延迟。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04658", "title": "RETRACE: From Entangled Repair Histories to Reusable Experience for CI Repair", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/480", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/979" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/480" ], "reason": "审查通过，待核心算法实现与公开验证。RETRACE 从 CI 完成修复历史做 endpoint 向后与开发历史向前的变更归因，生成 problem-level 三层可复用经验。CI-Repair-Bench565 PR/101仓库、最早30%历史与408后续目标可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04687", "title": "Understanding Errors in LLM-Based Question Answering over Imperfect Tables", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/478", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/975" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/478" ], "reason": "审查通过，待核心算法实现与公开验证。GBDI 对表格随机行排列做五个独立发现 context，映射回原行合并 union ledger，再按可验证证据 Edit/Drop/No-op。RADAR-T313例/53任务/27表可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04699", "title": "Not Self-Decidable: LLMs Cannot Draw the Boundary of What an Agent Verifier Can Check", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/477", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/974" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/477" ], "reason": "审查通过，待核心算法实现与公开验证。CoVer 将模型 corroboration 只用于提名，生成 fact-slot+规则并用 structure/groundedness/invariance/sensitivity 四门验证；使用基础设施 attested 字段、缺失或异常则 fail closed。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04721", "title": "Knossos and Ariadne: Benchmarking and Learning Complete Diagram Topology Extraction with Vision-Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/475", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/971" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/475" ], "reason": "审查通过，待核心算法实现与公开验证。Meta合著Ariadne先预测nodeinventory，再每组3active source对全predicted nodes输出typed edges，确定性并集后单独trace；推理不读GTinventory。Knossos六域18k/1200不同seed，另TopoBench162图外测；模板内seed隔离非未见grammar泛化。公开dataset/repo链接明确，方法与图结构评测皆有独立价值。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": "not_reproduced_high_cost", "this_audit_full_experiment_review_done": true, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04753", "title": "More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/473", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/968" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/473" ], "reason": "审查通过，待核心算法实现与公开验证。SAGA 解耦key/value头数，结合周期刷新阈值的近似top-N；预训练GQA转换通过数据加权key合并、最小权匹配和KL+CE蒸馏，而非简单删头。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04779", "title": "Asymmetric Repository Lineage Modeling and Verifier-Guided Coordination in Concurrent AI Coding Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/471", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/966" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/471" ], "reason": "审查通过，待核心算法实现与公开验证。MergeGym 区分 open-time authored-file overlap、真实 git merge-tree 冲突与调度重放；229 heldout repositories 的 T1、5936 可重放 T2 与97仓库调度任务是可执行基础设施候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04792", "title": "Do More Modalities Always Help? A Geometric Perspective on Missing-Modality Robustness", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/470", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/965" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/470" ], "reason": "审查通过，待核心算法实现与公开验证。GU针对缺失图像的Grassmann子空间编辑可作为多模态稳健算子。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04794", "title": "Knowing the Store: What a Memory Backend Must Write Down Before an Agent Can Read It", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/469", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/964" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/469" ], "reason": "审查通过，待核心算法实现与公开验证。可接入记忆治理诊断基础设施：区分 active/superseded/retracted/expired 与 current/past/change 意图，比较无值的状态摘要、返回记录 tags、简单名称 lookup，含444条同问题反事实存储版本。正文明确公开代码与数据，可忠实实现监测协议。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04805", "title": "ExStereo: Lifting 2D Vision-Language-Action Models to 3D with Explicit Stereo Representations", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/467", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/962" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/467" ], "reason": "审查通过，待核心算法实现与公开验证。ExStereo由冻结FastFoundationStereo、标定fxB/d和桌面RANSAC构造4方向正交depth/xyz/valid显式表示，以actionquery crossattention残差接入VLA；不替换prefix/VLM。7个3500episodes midtraining与10个评测任务/对象不交叠，post每任务100episodes；RoboTwin和三实机任务20次各验证。上游深度计算和标定必须计成本，不能把平均成功率百分点改写相对提升。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04824", "title": "Agent Behavior as Code: Efficient and Robust LLM Agents with Programmatic Specifications", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/466", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/961" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/466" ], "reason": "审查通过，待核心算法实现与公开验证。ABCAgent 以有未绑定输入变量的程序规定 agent 行为，代码 executor 管理作用域和类型状态，LM 负责设计/修改。GAIA、WorkArena、GSM-Symbolic、tau2 的公开基础任务可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04845", "title": "Which Preferences to Train On? End-to-End Multi-Objective Alignment with an Adversarial Preference Distribution", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/464", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1532" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/464" ], "reason": "审查通过，待核心算法实现与公开验证。MAESTRO用Dirichlet对抗偏好分布DirUEG更新，独立adversary/policy批次、向量reward GAE后偏好标量化；多目标HH/Beaver与摘要faithfulness实验。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04851", "title": "CURIO: Curiosity-Driven Test-Time Learning for Open-Ended Discovery", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/463", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1531" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/463" ], "reason": "审查通过，待核心算法实现与公开验证。CURIO训练hidden-state transition预测器并缓存更新前残差，epoch归一后只在非top-k采样token加bonus；它不是语义新颖性保证，encoder/predictor仍可能collapse；有科学发现任务实测。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04868", "title": "From Memory to Guide: Spatio-Temporal Composer for Procedural Coding Memory", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/460", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/956" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/460" ], "reason": "审查通过，待核心算法实现与公开验证。Composer 由 Selector 选择 Skill 并保留 whyNow，将规则绑定当前实体形成 Runtime Guide，按完成/失效条件终止并重新进入任务要求。固定 Skill Bank 的程序记忆适配机制可实施。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04882", "title": "ResOPD: Tail Residualization for Sparse On-Policy Distillation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/458", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1530" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/458" ], "reason": "审查通过，待核心算法实现与公开验证。ResOPD完整top-k质量加tail bucket基线，每步coarse梯度与稀有tail残差组成无偏估计（仅全支持祖先采样）；主数学实验top-p=.95/top-k20不满足该无偏条件，必须分别报告机制与成绩。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04889", "title": "ForkPilot: Self-Evolving Policy for Retrospective Search in Long-Horizon Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/456", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/954" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/456" ], "reason": "审查通过，待核心算法实现与公开验证。ForkPilot 使用恢复同状态的 search-vs-skip 成对完成轨迹训练 task-bootstrap ensemble，结合成本、支持度与校准边际决定 commit/probe/skip，任务内冻结、跨已完成源任务更新。六个公开 benchmark 可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04899", "title": "Rewrite What Matters: Adaptive Multilingual Query Rewriting for Reasoning via Agentic Reinforcement Learning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/454", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1528" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/454" ], "reason": "审查通过，待核心算法实现与公开验证。mRewriter-R1 顺序执行 language/structural/semantic 三类有限算子，由 GRPO 优化冻结下游模型结果。MGSM/BELEBELE/XCOPA 明确训练和标准语言测试可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04915", "title": "Are We Measuring Scientific Intelligence? Rethinking the Evaluation of AI Scientists", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/452", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/952" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/452" ], "reason": "审查通过，待核心算法实现与公开验证。以null/withdrawal/flip三类经统计量验证的反事实数据评估科学Agent是否依证据作答，并用EGA区别答对旧key与响应新证据，可接研究完整性评测。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04916", "title": "PreAct-Nav: Agentic Reasoning Before Action for Urban Navigation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/451", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/951" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/451" ], "reason": "审查通过，待核心算法实现与公开验证。PreAct-Nav 用 Navigation Memory 的中期子目标和按需 Predictive Sandbox 生成行动未来线索，冻结 VLM 先审后行动，证据未决或修订无效则回退原 policy。UrbanNav/CityWalker 可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04918", "title": "Residual Visual Credit Optimization: Conserved Evidence Routing for Multimodal Reinforcement Learning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/450", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/950" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/450" ], "reason": "审查通过，待核心算法实现与公开验证。RVCO用图像patch干预似然构建证据分数、IQR/entropy路由及residual校正精确保留响应credit质量；ViRL39K四VLM家族训练，需真视觉干预而非固定token启发式。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04921", "title": "Complex Agents, Shallow Tests: Demystifying and Enhancing Test Adequacy of Agent Harness in the Wild", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/449", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/949" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/449" ], "reason": "审查通过，待核心算法实现与公开验证。HarnessTester用coverage找未覆盖LLM依赖代码，AST精确检索构造器/fixture/mock绑定，指导保留调用契约并生成不同分支测试，适合harness工程测试候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04927", "title": "Assembling Insights for Agentic Machine Learning Engineering Systems", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/448", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/948" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/448" ], "reason": "审查通过，待核心算法实现与公开验证。MLE-InsightForge 从源竞赛 gold 解法提炼 competition/domain/improvement insight，按触发-动作条件注入 ReAct 解法树；160 Kaggle 任务/12域的源-目标隔离协议可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04940", "title": "Software World Models: From Consequence Prediction to Decision Value", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/447", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/947" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/447" ], "reason": "审查通过，待核心算法实现与公开验证。SWM 从恢复的训练世界执行反事实变化，SFT warmup+GRPO 学习 downstream blast set，再以采样概率决策迁移和付费检查；真实开源库与独立 synthetic world 可验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04950", "title": "How Should Teachers Be Prepared? RL on Student-Induced States for On-Policy Distillation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/446", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1526" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/446" ], "reason": "审查通过，待核心算法实现与公开验证。Prep-OPD先让teacher在固定初始student产生的prefix状态上RL，再冻结teacher做OPD，实测DeepMath及8数学基准；无gold解法灌入teacher，需保留两阶段状态分布。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04961", "title": "Building LLM Agent Systems the Deep Learning Way: From Modular Design to Architecture Search", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/444", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/945" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/444" ], "reason": "审查通过，待核心算法实现与公开验证。LLM building blocks将单调用/历史递归/ToT/GoT/多queryRAG组合为可执行harness，自反思更新metaprompt，Optuna在train30%搜索max tokens/head/retrieve参数后70%test冻结。publicGSM8K/StrategyQA/TSP可实现模块化架构搜索。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.04991", "title": "CIPO: Counterfactual Imagination Policy Optimization for Adaptive Tool Granularity Selection", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/440", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/941" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/440" ], "reason": "审查通过，待核心算法实现与公开验证。CIPO用预算BPE挖成功train工具链并实例化可binding/trace/interrupt技能，G4 rollout中两base各在first eligible state替换atomic/skill并同policy继续，bounded outcome difference补充reward再GRPO。适合Agent RL/toolgenome。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05023", "title": "Look Where You Say You're Looking: Self-Grounded Attention for Visual Reasoning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/438", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/939" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/438" ], "reason": "审查通过，待核心算法实现与公开验证。Self-Grounded Attention以GroundingDINO为每个CoT观察步骤定位，选定attentionheads的区域显著性得分加答案exactmatch/judge与format奖励后GRPO；head选择依VisualCoT相关性，不能将attention当因果忠实性。SaliencyR1-8k训练与25个VQA lmms-eval评测；需保留真实定位器/辅助judge和8H100训练预算，不能用答案窥视fixture代替。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05024", "title": "LightVLN: Efficient Aerial Vision-and-Language Navigation with Compact Memory and History-Guided Local Aggregation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/437", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/938" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/437" ], "reason": "审查通过，待核心算法实现与公开验证。LightVLN每historicalframe learnedquery单token FIFO，historyselectinstruction指导current16×16grid固定32局部区域attention汇聚，保留空间覆盖。OpenFly/AerialVLN分开训及未见campus，完整模型1.276B非仅.5B，4090profile/闭环路径应分别记；无额外memoryencoder不等于无visionbackbone。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05025", "title": "Triggering Generalist Reasoning via Predictive Uncertainty for Dual-System VLA", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/436", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/937" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/436" ], "reason": "审查通过，待核心算法实现与公开验证。TUD将π0拆heavyPaliGemma generalistcache与currentSigLIP/Qformer/actionexpert，同anchor跨观测重预测固定slot std衡量plan变化；runningmean阈值online百分位p70+maxskip触发刷新，执行仍committedchunk非每次probeaction。VLA-Arena/CALVIN/实机，variance是stalenessproxy非校准epistemic uncertainty，额外probe成本需计。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05026", "title": "GeoBridge-VLA: Geometry-Aware Residual Adaptation for Vision-Language-Action Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/435", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/935" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/435" ], "reason": "审查通过，待核心算法实现与公开验证。GeoBridge-VLA nativevisual中间feature→bridge/K4邻域geometrydecoder，stage1 simdepth或实机DA3pseudodepth监督；stage2冻结geometry+VLM，zero-init1×1residual gate≈.002对nativepatch相加，flow+RMSresidual罚。SmolVLA LIBERO/OMY，仅保持tokencount非零新增decoder计算；跨backbone未验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05030", "title": "EVISKILL: Grounding Skill Evolution in Replayable Evidence", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/433", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/933" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/433" ], "reason": "审查通过，待核心算法实现与公开验证。EviSkill由train trajectory span构建Replayable Evidence Cards链接局部edits，恢复prefix状态重执行验证/反思，globalvalidation更新ValidatedSkill；全局拒绝时仅replay支持edits留WorkingSkill ledger跨epoch再检验。适合Agent skill RSI。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05039", "title": "Causal Improvement Graph for Agentic Harness Optimization", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/430", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/930" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/430" ], "reason": "审查通过，待核心算法实现与公开验证。CIG以Evidence–Hypothesis–Intervention–Outcome typed graph作persistent improvement state，Observe/Plan/PlanSelect/Implement/Evaluate/Update局部proposer调用，由外部validation结果修订hypothesis assessments再portfolio选择。可接Agent harness RSI/evolve。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05047", "title": "Beyond Task Completion: Measuring Interaction Cost in Terminal User Interfaces", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/428", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/928" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/428" ], "reason": "审查通过，待核心算法实现与公开验证。TUINaut用Docker+Tmux共享真实TUI输入输出，独立状态oracle检验完成度；Agent-KLM分报推理token与原子按键数，适合交互Agent成本诊断。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05066", "title": "Salvation Lies Within: Eliciting Inherent Style Transfer in Step-Distilled Diffusion Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/427", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/927" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/427" ], "reason": "审查通过，待核心算法实现与公开验证。StyleForge无训练agent提render/color-light/composition/referencecontent四状态，统计色/亮度约束后按固定编译顺序保留style弱化composition排除referencecontent，再交few-stepFLUX2Klein；31style×16prompt及三Qwenjudge对照，非新扩散模型/蒸馏loss。可复用structuredprompt operator，judge自动分不当human偏好或生产AB。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05097", "title": "ReMAP: Restoring the Perceptual Cycle with Reasoning-Time Latent Visual Memory", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/426", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/926" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/426" ], "reason": "审查通过，待核心算法实现与公开验证。ReMAP使用全局/局部latent视觉memory、region定位器和分阶段形成/访问训练，再用RL学习何时取证；十基准，必须保留真实视觉memory而非文本oracle。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05099", "title": "Small Agents with Semantic Search: Efficient Multilingual Code Localization", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/425", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1525" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/425" ], "reason": "审查通过，待核心算法实现与公开验证。Compact code-localization agent用colgrep late-interaction CPU index，goldonlytrain turnweights SFT再DrGRPO/DAPO dynamicfilter异步rollout训练；publicSWE-Lite/MultiSWE-Flash按basecommit定位文件，可接Agent后训练候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05107", "title": "SearchJev: A Fast and Calibrated System-1 Model for Search Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/422", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/922" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/422" ], "reason": "审查通过，待核心算法实现与公开验证。SearchJev对schema legal-option token logits直接readout，CE+Brier softlabel按来源weight训练LoRA，再独立validation按outputtype温度校准；uncertain decisions交System2。可作Jev/Agent search跨轨算法。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05124", "title": "Memory Canonicalization: A Framework and Benchmark for Cross-Model Drift in Persistent LLM Memory", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/418", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/918" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/418" ], "reason": "审查通过，待核心算法实现与公开验证。写入记忆时做结构消歧、情绪字段提取与独立验证，CMSC-E以统一自然语言渲染隔离内容改写和JSON格式因素，适合作为记忆漂移诊断。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05126", "title": "Selecting Repetition Counts Across Model Scales in Data-Constrained Pretraining", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/416", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/917" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/416" ], "reason": "审查通过，待核心算法实现与公开验证。小尺度重复次数loss曲线经seedbootstrap+logN回归选top3、以及双项曲线argmin，是明确的预训练数据预算选择operator，target预测前冻结且测试独立。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05132", "title": "SpecAgent: Empowering Program Verification with Agentic Synthesis of Formal Program Specifications", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/413", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/914" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/413" ], "reason": "审查通过，待核心算法实现与公开验证。SpecAgent用函数/loop dependency计划与ACSL知识RAG生成annotation，Frama-C/WP反馈可回修upstream contract，critique用双次同input调用return equality completeness probe辨弱spec并继续修订。public C/SV-COMP任务适合formal-method Agent候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05139", "title": "Hidden in the Comments: A Context-Injection Attack Surface in Code LLMs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/412", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1523" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/412" ], "reason": "审查通过，待核心算法实现与公开验证。公开安全评测：双条件Flask prompt、generation harness、5组件detector及comment position对照。Zenodo API已证实open/published、Manuscript_Codes.zip及校验和。非同scaffold配对且task类型混淆；detector以实验条件作参考非漏洞gold，不能当纯因果或真实漏洞准确率。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05162", "title": "Memadapter: Counterfactual Adaptation Against Memory-induced Sycophancy", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/411", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/913" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/411" ], "reason": "审查通过，待核心算法实现与公开验证。MemAdapter对冻结retrieved memories先造counterfactual任务归纳条件使用边界，再当前context反思限定inferential角色，最终evidence-based回答；不改上游memory/retrieval。适合记忆可靠性机制。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05176", "title": "AECG: Asymmetric Experience Consolidation and Governance In Multi-Agent Systems", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/410", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/912" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/410" ], "reason": "审查通过，待核心算法实现与公开验证。AECG按team/event/agent保留scope，成功与归因可信失败不对称入账，slow累积/fast折扣Beta-style信念检出退化并按downstream exposure有限预算review；narrow/repair版本仅paired failure+disjoint success replay通过再启用。适合Agent memory/RSI。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05185", "title": "Recurrent Latent Visual Search for GUI Grounding", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/409", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/911" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/409" ], "reason": "审查通过，待核心算法实现与公开验证。ReLaViS LVS latent slots递归spatial distribution/evidence向量回灌，CWC对每candidate质量下界+final截断Gaussian KL；三阶段warmup/GTlatentteacherforce/predictedselfcondition，test只screenshot+query，3×3峰邻域坐标。100ktrain/5kval隔离五GUIbench，grounding非端到端agent任务。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05190", "title": "FORGE: Verification-Gated Behavioral Repair for Generative Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/408", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/910" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/408" ], "reason": "审查通过，待核心算法实现与公开验证。FORGE通过prefix/LOO定位毒性，候选token分类后用lm_head低秩QP或MLP零空间编辑，重新生成验证再commit/rollback；可作可审计权重修复机制候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05247", "title": "templar: agentic induction and evolution of standardized radiology reporting templates from large-scale clinical corpora", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/402", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/904" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/402" ], "reason": "审查通过，待核心算法实现与公开验证。TEMPLAR把模板与anatomical/diagnostic provenance graph作持久状态，Span-Triple双视图BM25/embedding图Leiden聚类归纳slot，结构reducer约束修订，外证据grounding及train结构化反馈校准粒度。可作多模态/医疗Agent机制候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05273", "title": "When and What to Prune? Stage-Aware Visual Token Pruning for Efficient VLA", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/399", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/902" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/399" ], "reason": "审查通过，待核心算法实现与公开验证。SAPrune无reward校准actionattention三depth稳定层，再QR解释度×attention集中度调top-attn/functionaldiversity双路径budget；只prunevisual保留text/action。三VLA LIBERO/SIMPLER+4实机，RTX5090 SDPA latency需保留与FA2区别，不借FLOPs替代实测。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05274", "title": "Answer with Evidence: Consistency-Aware Grounded Visual Question Answering for Roadside Traffic Scenes", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/398", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/901" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/398" ], "reason": "审查通过，待核心算法实现与公开验证。EtA先box再answer的规范序列；ECPO Grollout跨样本IoU聚类≥max(2,ceil(.25G))形成consensus pseudobox，Hungarian覆盖reward+format+referenceanswer后GRPO，不用GTboxes于RL但非无标签。RoadSceneVQA/gRefCOCO/CARPK跨任务，结构一致性存在性证明不保证真实完整grounding。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05282", "title": "Red-TTT: Test-Time Training for Automated Jailbreaking Large Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/397", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1520" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/397" ], "reason": "审查通过，待核心算法实现与公开验证。Red-TTT按单目标test-time LoRA适应，softmax(beta reward)-uniform风险偏好优势、随轮增强beta与elite replay；固定120查询预算研究。仅列算法研究，不实施攻击。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05284", "title": "Fusion is the New Mutation: Bandit-Guided Evolution on Workflow Graphs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/396", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/900" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/396" ], "reason": "审查通过，待核心算法实现与公开验证。DAGO将工作流搜索记录为generation lineage DAG，固定五parent组合arm，缓存concat embeddings，shared diagonal LinUCB从退火score-softmax提案中选择fusion；LLM从最强parent开始按总结做targeted融合，绝对子代validation成绩更新bandit。适合harness evolve算法。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05295", "title": "Readable Before Actionable: Causal Tracing of Indirect Prompt Injection", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/394", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/898" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/394" ], "reason": "审查通过，待核心算法实现与公开验证。Readable Before Actionable用content/format反事实role probe、component activation patch和独立hijacked-resisted suite差分方向，检验pre-action/tool-span固定activation edit对AgentDojo攻击行为影响。可收录安全steering实验机制。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05300", "title": "MESH-Harness: Self-Improving Agent Harnesses via Bandit-Guided Compositional Evolution", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/393", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/897" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/393" ], "reason": "审查通过，待核心算法实现与公开验证。MESH-Harness固定slot接口与容量池，缓存embedding、full-covariance LinUCB+混合起点coordinate search选择完整组合，每5评估trace驱动两slot变异与evict，保留surrogate统计。可接组合harness evolve。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05303", "title": "ASCENT: Online Test-Time Training of Long-Horizon Agents via Self-Distillation of Verified Experience", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/392", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/896" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/392" ], "reason": "审查通过，待核心算法实现与公开验证。ASCENT单次在线episode仅对已验证成功轨迹更新LoRA，以固定初始模型读过滤的自身hindsight作为teacher、forward-KL训练不见hindsight的student；无外部gold答案，须严格prequential协议。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05308", "title": "RubricArmor: Adversarial Evolution Improves LLM-Based Rubric Generation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/389", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1519" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/389" ], "reason": "审查通过，待核心算法实现与公开验证。RubricArmor独立质量判断确认满分但有缺陷的攻击，再经降分与独立review双验收迭代修补rubric；有2K WildChat、三policy真实GRPO验证。核心是可执行reward构造循环，须保留验证者非简单字符串rubric。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05318", "title": "Robust Parameter-Efficient LLM Adaptation on Analog Hardware", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/388", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/893" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/388" ], "reason": "审查通过，待核心算法实现与公开验证。冻结base用功能等价通道缩放，LoRA用校准异常通道拆分及数字残差累积保留亚阈值更新；可作为模拟硬件误差下的训练机制诊断。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05334", "title": "AgentDiscover: Autonomous Discovery with Minimal Search Scaffolding", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/385", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/891" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/385" ], "reason": "审查通过，待核心算法实现与公开验证。AgentDiscover用sandbox coding agent自行查询idea/candidate/session/lesson图数据库并提交隔离评分，跨session fresh context保留tools/skills；meta agent从trace提炼搜索指导，可接开放搜索evolve。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05342", "title": "IRSTD-Agent: Agentic Infrared Small Target Detection via Zoom-Guided Interaction Learning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/384", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/890" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/384" ], "reason": "审查通过，待核心算法实现与公开验证。IRSTD-Agent用proposal/zoom/detect/drop/refine视觉工具，基于train mask构造target-preserving/offset-recovery/negative trajectories，SFT预测工具及坐标，测试从native image获得证据。可作为多模态主动视觉Agent候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05367", "title": "AIProver: Agentic Auto-Formalization of Mathematical Research via Certificate-Driven Evolving Harness", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/383", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1517" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/383" ], "reason": "审查通过，待核心算法实现与公开验证。AIProver交替训练模型与进化harness，SAM双视图对比及CE、证书RLSF和独立fitness集accept/rollback；需真实形式验证链，不得用gold证明fixture替代。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05373", "title": "Towards Unbiased On-Policy Distillation for Block Diffusion Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/381", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1516" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/381" ], "reason": "审查通过，待核心算法实现与公开验证。Un-OPD筛选可见上下文对齐的block决策，对高teacher-support位置混合teacher与detach student logits校准，并复用rollout；真实BDLM实验。标题unbiased不可扩展为任意复用无偏保证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05383", "title": "Sibyl: An Efficient Small-large Model Collaboration Framework for Long-horizon Tasks", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/379", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/887" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/379" ], "reason": "审查通过，待核心算法实现与公开验证。Sibyl先hindsight skill OPD+RL增强本地policy，失败prefix replay找单次cloud action能逆转结果的decisive disagreement冷启动call_cloud，再以trajectory/matched-state/consultation counterfactual credit联合训练和skill internalization。适合Agent RL/OPD跨轨。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05398", "title": "MMPostTrainBench: Benchmarking Autonomous Research for Multimodal Post-Training", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/377", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/885" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/377" ], "reason": "审查通过，待核心算法实现与公开验证。MMPostTrainBench将多模态训练任务封装成Harbor独立verifier，MMResearch把媒体证据、假设、干预结果构成图并配global/core/archive记忆，适合多模态evolve控制器。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05402", "title": "Task Vector Descent: Learning from Non-IID Batches", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/374", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/882" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/374" ], "reason": "审查通过，待核心算法实现与公开验证。TVD将连续同分布小批更新的task vector部分整合，同时插值Adam一阶矩/步数、平方根空间二阶矩；固定样本预算与训练/评价分离，覆盖SFT及GRPO。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05418", "title": "EvoMem-VLA: State-Evolution Memory for Long-Horizon Robot Manipulation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/371", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/879" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/371" ], "reason": "审查通过，待核心算法实现与公开验证。EvoMem-VLA冻结paired DINO delta tokenizer，initial+recent+事件state/delta交织记忆，keyframehead预测未来record offset只在时间到达提交；task routing是配置而非learned classifier。训练GTsubtask/test生成subtask有暴露偏差，闭环只读已观测，RMBench/RoboMME+双embodiment80实机episodes。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05425", "title": "Unmentioned Checklist Findings Change How Reinforcement Learning Appears to Improve Chest Radiograph Report Checking", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/370", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/878" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/370" ], "reason": "审查通过，待核心算法实现与公开验证。公开checklist/reader格式稳健性诊断：图像writer不读待查句，checker不读图；matched-label pair和三个checker区分omission/order影响。官方代码/格式/analysis可访问，CheXpert数据须注册协议；医学专门诊断不作临床效益或通用视觉提升证据。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05473", "title": "Hierarchical Reinforcement Learning with Stable Temporal Abstraction for Language Model Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/366", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/874" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/366" ], "reason": "审查通过，待核心算法实现与公开验证。STAC在HiPER归一化boundary advantage上按已采样KEEP/SWITCH减去过早重规划与过时保留成本，PPO仅boundary token改变，reward/critic/subgoal/action保持host算法。适合hierarchical Agent RL。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05481", "title": "CodeForge-MA: Execution-Verified Multi-Agent Learning with Language-Conditioned LoRA for Multilingual Code Generation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/365", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1514" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/365" ], "reason": "审查通过，待核心算法实现与公开验证。PPO+SFT本身是标准配方，但语言条件稀疏Mixture-of-LoRA门控是可执行新增参数机制；Qwen1.8B与72B API实验。应复现路由和辅助约束，非仅重命名PPO。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05484", "title": "Universal Test-Time Training", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/364", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/873" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/364" ], "reason": "审查通过，待核心算法实现与公开验证。uTTT通过跨层共享快权重池与分离读写路由形成新长上下文算法；同总状态/活跃计算/训练token控制支持核心贡献。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05532", "title": "DelegationBench: Measuring When AI Agents Should Ask Before Acting", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/362", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/871" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/362" ], "reason": "审查通过，待核心算法实现与公开验证。接受为evolve权限/信息需求分离的评测基础设施；156场景48matched pairs、Act/Ask/Request/Refuse分别计分，Tool-Loop与post-approval测试行动而非只问判断。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05538", "title": "LiFT: Loop Flow Transformers", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/361", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/870" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/361" ], "reason": "审查通过，待核心算法实现与公开验证。LiFT以随机连续深度坐标监督共享DiT核心的逐步速度修正，stop-gradient初始速度锚与prelude辅助损失保留核心机制；同训练token、块设计、执行深度与近等FLOPs的dense对照支持接入研究候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05540", "title": "SALUS: Automated Auditing of NL-to-SQL Benchmarks through Weak Supervision of Multi-Agent Output", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/360", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/869" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/360" ], "reason": "审查通过，待核心算法实现与公开验证。SALUS对多agent候选SQL执行/结构/intent弱标记建generative label model，均衡高置信伪标签训练SQL-feature条件化agent reliability decision plane，融合verdict检测基准错误。适合评测质量基础设施候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05563", "title": "G-CARB: Graph-Localized Conformal Agent Risk Budget for Compositional Harm", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/358", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/867" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/358" ], "reason": "审查通过，待核心算法实现与公开验证。CARB对stopped-prefix全episode harm ledger做finite-grid corrected conformal risk calibration；G-CARB只向固定scorer提供prefix-visible source-to-sink依赖路径，保留proposal空结果与threshold nesting。接受安全门控机制候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05572", "title": "Disentangling Task Difficulty from Run-Level Failure in Agent Failure Prediction", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/356", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/865" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/356" ], "reason": "审查通过，待核心算法实现与公开验证。分解pooled AUROC为同任务/跨任务比较，以task-grouped交叉验证和绝对turn prefix区分任务难度与当前run失败信号，适合统一预算控制评测。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05615", "title": "DREAM: Dynamic Resolution Assignment For Multimodal Multi-agent Debate", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/352", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/862" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/352" ], "reason": "审查通过，待核心算法实现与公开验证。DREAM用多分辨率probe的ANLL均值减标准差形成confident set，五分支分配visual payload；丢弃probe历史后debate，按每agent ANLL轨迹回滚早期低不确定答案，再judge。适合多模态Agent推理候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05622", "title": "UndoBench: Separating Task Competence from Recovery Capability in Tool-Using AI Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/351", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/861" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/351" ], "reason": "审查通过，待核心算法实现与公开验证。UndoBench 对相同seed执行control/fault配对试验，以独立mutation log与state oracle区分能力、恢复、重复和漏执行，适合工具Agent容错评测。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05630", "title": "Atomic Visual Entailment: Enhancing Zero-Shot Vision-Language Reasoning through Atomic Fact Decomposition and Learned Selection", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/350", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/860" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/350" ], "reason": "审查通过，待核心算法实现与公开验证。AVE冻结Decomposer原句拆atoms，两VLM direct/atomic候选pool；XGBoost以labelscore/entropy/agreement等meta预测最终label，非必选pool已有label，再取代表justification GroundingDINO定位。47,203train中20%val选后固定SNLI-VE dev/test，94%oracle不能当部署结果；Flickr30kEntities是近似证据定位非完整faithfulness。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05637", "title": "Visual Grounding Safety in Vision-Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/349", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/859" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/349" ], "reason": "审查通过，待核心算法实现与公开验证。grounding safety以相同intent VQA/定位匹配评测，SFT混harmfulrefusal+有中性证据的benignselfdistill+1548capabilitygrounding；plain和placeholder格式分别训LLM。最终坐标/回答而非think判refusal，定位失败也可能假安全故单列PointArena能力/overrefusal；placeholder非真实可执行坐标。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05664", "title": "StageVLN: Spatial and Trajectory Auxiliary Guidance for Efficient Vision-Language Navigation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/348", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/858" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/348" ], "reason": "审查通过，待核心算法实现与公开验证。StageVLN多层current-image cosine对齐冻结geometryteacher，历史相对heading sin/cos与expert forward步比例progress辅助训练，pre-action hidden避免teacherforce读GT当前action。jointR2R/RxR train、valunseen，推理去teacher/projectors/heads；progress不是成功率也不是STOP规则。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05685", "title": "Spend Bytes on Breadth: Precision-Count Trade-offs for Decode-Time KV Compression in Long Chain-of-Thought Reasoning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/345", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/855" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/345" ], "reason": "审查通过，待核心算法实现与公开验证。BreadthKV在相同decode KV字节预算下联合KIVI量化和SnapKV-D淘汰，以独立AIME22–23校准位宽，研究更广低精度缓存优于少量高精度缓存。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05697", "title": "From Token-Max to Outcome-Max: How You Use AI Determines Its Productivity", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/343", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/853" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/343" ], "reason": "审查通过，待核心算法实现与公开验证。TokenEcom按verified output/cost而非token量设模拟监督策略，跨代只传递可重用playbook，执行反馈提炼双方skill。公共SWE/HumanEval/GSM8K/MATH适合cost-aware evolve评测候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05709", "title": "Nexus: An Execution Fabric for AI Agents Across Cloud, Edge, and Devices", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/341", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/852" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/341" ], "reason": "审查通过，待核心算法实现与公开验证。Nexus将调用持久化为task，使用invocation-scoped委托、operation journal、device receipt及cursor恢复，明确已提交结果重用与不确定外部副作用的资源端协调。接受执行基础设施候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05715", "title": "Rotated, but How Far? Diagnosing and Improving Object-Rotation Reasoning in VLMs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/340", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/851" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/340" ], "reason": "审查通过，待核心算法实现与公开验证。RotationCue 使用三个可学习关系查询，顺序预测对象同一性、旋转存在与角度范围，将条件化角度提示注入冻结 VLM。ALOI/ModelNet40 可构造训练，OR-Bench 为对象隔离的完整测试集。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05744", "title": "CIPHER-MoE: Balancing Efficiency and Routing Fidelity in Trillion-Scale MoE Training", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/339", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/850" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/339" ], "reason": "审查通过，待核心算法实现与公开验证。CIPHER-MoE保持tokenTopK，专家端cosine affinity与capacity admission加locality loss，区分严格drop与reroute，有真实大规模SFT效率/质量证据。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05774", "title": "AdaSpark: Adaptive DSpark with Online Learning for Tree Verification and N-gram Fill", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/337", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/848" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/337" ], "reason": "审查通过，待核心算法实现与公开验证。AdaSpark将DSpark候选与请求内ngram合并，按路径接受概率排序合法树前缀，用在线低分位延迟模型和接受率模型选择验证宽度；可作为推测解码控制器候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05778", "title": "MedicalHarness: A Controlled Evaluation of LLMs and Agent Harnesses on Medical Tasks", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/336", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/847" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/336" ], "reason": "审查通过，待核心算法实现与公开验证。MedicalHarness 的MH-Lab在相同执行循环中切换摘要、显式计划和工具发现接口，并将回答率、正确率、轨迹和成本分开；公开代码/MedTool可作为harness消融候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05833", "title": "Request Order Matters: Cache-History Sensitivity in Selective KV-Cache Reuse for Rolling Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/333", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/844" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/333" ], "reason": "审查通过，待核心算法实现与公开验证。接受Agent长时rolling context的KV重计算机制诊断：固定5% token预算，按doc均值deviation选择整span及连续余量；对照token-top-k和其他连续块，核心证据支持contiguity而非document边界/评分本身。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05842", "title": "HLA: Expressive Hybrid Linear Attention via Chunk-Wise Dynamic Mixing", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/332", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/843" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/332" ], "reason": "审查通过，待核心算法实现与公开验证。HLA以query gate对完整chunk仿射GDN transition作identity插值并按时序复合，核心不同于固定state加权，适合长上下文结构operator。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05861", "title": "Imagine to Act: High-Fidelity Data Synthesis via Image Editing World Model for Scalable GUI Agent Training", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/331", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/842" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/331" ], "reason": "审查通过，待核心算法实现与公开验证。Infinite-Dreamer训练时从before/after/action/nextaction提delta；合成时AppDesigner text→delta驱动FLUX2Klein截图编辑，VLM筛step后SFT GUIactor。worldmodel AndroidControl train/val，actor仅合成、AndroidWorld/MobileWorld在线评；合成可用future设计但评测agent不能读gold nextstate/plan，worldmodel并非真实应用状态机。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05872", "title": "Off-Policy Merging Beats On-Policy Self-Distillation for Continual Learning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/329", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/839" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/329" ], "reason": "审查通过，待核心算法实现与公开验证。以离策略donor微调增量、retention Fisher/KFAC敏感方向投影与缩放合并到当前模型，Qwen/OLMo持续学习验证；不同于2610.00767信念移植及旧GRAFT2609.37868轨迹交换。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05879", "title": "Learning to Learn a Language", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/327", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/837" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/327" ], "reason": "审查通过，待核心算法实现与公开验证。PFLM以每序列新采样的recurrent SCM先验进行byte-level预训练，再冻结权重从长前缀学习预测；核心数据生成定义与300M模型/权重公开，属于基础模型新训练范式候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05885", "title": "The Optimization Landscape of Learning Compacted Context Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/326", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/835" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/326" ], "reason": "审查通过，待核心算法实现与公开验证。Fast Diamond简化Perceiver KV writer有明确单次/递归压缩及预算对照，保留为学习型KV压缩算法；正文readout矛盾已由canonical附录解读为frozen并显式记录。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05896", "title": "TasteRoute: Personalized Routing for Video Generation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/325", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/834" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/325" ], "reason": "审查通过，待核心算法实现与公开验证。TasteRoute frozen request embedding+generator/user embedding与交叉特征，BTpairwise+ordinalquality，general五fold prompt隔离、personal每annotator独立train/val/test且userdropout.4；不生成video直接预算内排序。仅9annotator/3128video，冲突偏好准确55.2%增益有限，不泛化到大众用户，oracle verifier结果不可作可部署能力。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05897", "title": "Fitting Vision Adapters at Frontier Scales", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/324", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/833" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/324" ], "reason": "审查通过，待核心算法实现与公开验证。仅约50M vision projector连接冻结encoder/reader，sharedvocab affine embedding对齐初始化、SFT空think后仅adapter CISPO image-math RL恢复reasoning，正reward需非空reason/answer。Qwen4–32B同recipe可控，550/743B MoE训练/奖励不同不能规模因果比较；blind/图像破坏验证是关键。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05935", "title": "ThunderSyncRL: Lossless Acceleration of Agentic Reinforcement Learning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/323", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1511" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/323" ], "reason": "审查通过，待核心算法实现与公开验证。ThunderSyncRL用固定策略版本、GRPO双梯度累积和OPD按token延迟归一化在单次optimizer barrier前流式计算，数学更新等价；B200/H200数值核查和真实agent训练，不应把异步陈旧梯度代理当复现。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05940", "title": "ReMem: Streaming Video Understanding With Long Context Retention", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/322", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/832" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/322" ], "reason": "审查通过，待核心算法实现与公开验证。ReMem训练外SCM以固定generalquery注意力选择递归压缩、保留RoPE原位+时间offset；RVM clip pooledkey与query向量cos检索topclip，有限FIFO淘汰，非无限无损记忆。Streaming/OVO/长视频与NIAH；r=.7在StreamingBench SQA选择，正式复现必须重建validation并不能把调参集成绩作为独立test。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05978", "title": "Byte Language Models: Scaling, Emergent Abstractions, and Information Allocation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/319", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/829" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/319" ], "reason": "审查通过，待核心算法实现与公开验证。Byte模型256vocab+hash ngram/DeltaNetconv16+前30%四byte TST后70%NTP，与subword做IsoFLOPs而非相同text exposure；late-layer pooledspan干预为诊断，200M分段模型不可省略。speculative仅给定evaltext的oracle simulation非真实decoded latency，不能把3.4×接受token当3.4×提速。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05981", "title": "Mechanizing the User's Eye: Pre-Registered Deployment of a Sabotage-Validated Fail-Plausible Observer in a Production LLM Agent Runtime", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/318", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/828" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/318" ], "reason": "审查通过，待核心算法实现与公开验证。接受小规模Agent可靠性回归基准：五确定性signals→带verbatim substring证据闸门的LLM judge；14例公开offline corpus+逐detector sabotage+固定manifest/评分器，适合evolve防退化diagnostic而非通用能力分。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05993", "title": "From Transformation to Target State: Rethinking Query Representation for Zero-Shot Composed Image Retrieval", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/316", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/826" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/316" ], "reason": "审查通过，待核心算法实现与公开验证。ASAP-CIR冻结MLLM把reference+edit重构多holistic静态target描述与weighted atomicfacts，CLIP global及patchsoftmax匹配分别query-wise zscore融合。FashionIQ val/CIRR test1/CIRCO test，无任务训练但有dataset-specific prompts，不能称无任务设计/无额外MLLM成本。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.05994", "title": "How (and How Not) to Use Data Augmentation in VLA Post-Training", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/315", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/825" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/315" ], "reason": "审查通过，待核心算法实现与公开验证。VLA PPO critic-only图像增强在clean actor/rollout上通过advantage影响策略，是真实可复现训练干预；可作后训练/机器人研究候选，不是工业推荐结果。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06001", "title": "AgentSpy: Making AI Agent Behavior Observable", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/314", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/824" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/314" ], "reason": "审查通过，待核心算法实现与公开验证。AgentSpy 在隔离microVM外记录进程树系统调用/网络，再按命令、文件、主机的归一化计数构造star graph与weighted Jaccard，可作为独立可观测性机制候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06021", "title": "Scalable Minimal-Change Learning for Controllable Image Editing", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/311", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/820" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/311" ], "reason": "审查通过，待核心算法实现与公开验证。Arro VLM分别审计未实现/意外变化，group内LLM去重统一rubric再VLM Boolean核验，负count两channel独立标准化→加权→batch重归一后flowGDPO；不是简单单reward求和。MinEval10k/600disjoint另MagicBrush/AnyBench/EmuEdit，模型judge分数非生产AB，需计group auditor预算与外部judges。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06026", "title": "Differentiable Bit-Widths: Co-optimizing Pruning and Quantization via SVD for Ultra-Efficient LLM Compression", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/309", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/818" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/309" ], "reason": "审查通过，待核心算法实现与公开验证。DBW按SVD分量学习含0的位宽，STE加Hamilton剩余位分配、矩阵级预算和奇异值熵正则，联合剪枝与低比特量化。公开训练/评估实现适合压缩算子候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06064", "title": "TrustMI: Causally controlling how assistants trust their users", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/303", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/812" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/303" ], "reason": "审查通过，待核心算法实现与公开验证。TrustMI以冻结LM的全层BiPO steering矩阵+正向trust NLL抑制likelihood displacement；用户span和assistant span分别训练，Agent主实验把user-trained矩阵加到所有用户及工具输出。接受Meta高优先安全机制及公共评测候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": "not_reproduced_high_cost", "this_audit_full_experiment_review_done": true, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06100", "title": "From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/300", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/809" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/300" ], "reason": "审查通过，待核心算法实现与公开验证。Trace2Env从trace构造冻结worldbook，区分跨episode知识、当前state与逐步memory；适用性gate、schema/约束验证分阶段commit，配真实环境重放W2R评测，接受世界模型/Agent环境候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06147", "title": "Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/298", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/807" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/298" ], "reason": "审查通过，待核心算法实现与公开验证。TTC候选memory子空间稳定性校正，公开代码/特征包使其适合作为无梯度VLM适配候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06161", "title": "Introducing Code-Switched Contexts to Cognitively-Inspired Bilingual Model Training", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/296", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/805" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/296" ], "reason": "审查通过，待核心算法实现与公开验证。字典/POS受控code-switch与四阶段密度课程是可执行数据增强候选，适合小型双语预训练机制诊断；必须同时保留英中退化负结果。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06191", "title": "Judged Useless, Queried Anyway: Tool-Using Agents Rarely Turn Their Own Evidence Judgments into Stopping Decisions", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/293", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/802" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/293" ], "reason": "审查通过，待核心算法实现与公开验证。根据Agent自身USELESS判断累计失败检索，在连续五次无用时由harness触发整合/停止；固定步数对比和新300题确认，适合工具预算控制候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06192", "title": "Copies or Sources? Measuring How LLM Aggregators Count Restated Evidence in Multi-Agent Systems", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/292", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/801" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/292" ], "reason": "审查通过，待核心算法实现与公开验证。用精确Bayes可数来源任务测复制证据权重，比较去重声明、抽取唯一读数、只转发引用三种通信干预，可加入多Agent证据聚合诊断。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06193", "title": "Correct Code, Broken Contributions? SWE-CC: Benchmarking Repository Policy Compliance for Coding Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/291", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/800" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/291" ], "reason": "审查通过，待核心算法实现与公开验证。SWE-CC 将12个SWE-bench仓库的823条贡献政策编译成有前置条件的确定性检查器，同时评估轨迹与补丁，适合作为代码Agent合同评测候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06226", "title": "LeAVJEPA: A Minimalist Architecture for Audio-Visual Self-Supervised Learning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/287", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/795" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/287" ], "reason": "审查通过，待核心算法实现与公开验证。LeAVJEPA以共享早期融合ViT、模态专属局部视图、对称不变性与SIGReg构成可训练视听JEPA；附录C.3表A3同编码器/30epoch/批量512的控制实验证实局部视图机制。仅接受为研究候选，不宣称SOTA或已复现。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06269", "title": "Evolving in Thought Space: Training a Small Model at Test Time Unlocks Better Discoveries", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/285", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/792" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/285" ], "reason": "审查通过，待核心算法实现与公开验证。Guidance-TTT仅训练小guidance LoRA，用archive parent summary提策略、冻结executor读完整parent落实、独立verifier给reward；adaptive entropic LOO advantage+centered base correction的实际RL闭环，接受高优先evolve候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06286", "title": "DeferKV: Rethinking Eviction Timing for One-Shot KV Cache Compression", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/284", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/791" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/284" ], "reason": "审查通过，待核心算法实现与公开验证。DeferKV 等首个真实生成token进入decode获得query后再压缩prompt缓存；将其与尾部prompt query按时间折扣融合，沿用底层方法的平滑/预算/选择。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06324", "title": "Readout Blindness: VLM Scores Miss the Spatial Direction Their Frozen Encoders Retain", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/281", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/787" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/281" ], "reason": "审查通过，待核心算法实现与公开验证。ADR冻结encoder patch/独立词特征做Sinkhorn excess coupling或dense定位质心，再根据解析subject-object关系轴作有符号位移；无新训练且context-free/unary盲性定理有前提，不能泛称所有VLM encoder无法方向。九模型五direction-balanced集，derangement null校正grounded gain抵消语言prior；适合readout方法与diagnostic evaluator。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06404", "title": "Quantifying the Stability of Multi-Step Reasoning via Error Amplification", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/276", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/783" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/276" ], "reason": "审查通过，待核心算法实现与公开验证。随机有序子采样CoT步骤与STE量化感知SFT组合，降低step-to-step误差放大；CLRS-Text/LEGO上可复现实验及Jacobian诊断。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06406", "title": "What Did the Agent Actually Do? Evidence-Grounded Oversight for Long-Horizon Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/275", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/782" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/275" ], "reason": "审查通过，待核心算法实现与公开验证。EBG确定性证据→行为→scope图、需求驱动一跳扩展及原文回核；AgentMonBench分离遗漏规范、测试不变语义替换和事后反馈前缀，接受监督机制与条件性公共评测候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06413", "title": "SpatialChain: A Benchmark for Auditing Spatial Reasoning Faithfulness in VLMs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/272", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/779" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/272" ], "reason": "审查通过，待核心算法实现与公开验证。SpatialChain用GQA scenegraph生成/answer过滤28,350train+899test CoT，评测模型输入image/question，judge可读GT graph评AC/RF/RC；surface overlap不证明faithfulness，RF是graph一致性非模型因果忠实性。官方data+eval+judge CLI齐全，可作公共评测；需把gold/reference与agent隔离。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06439", "title": "Better Call Reward: Reward Hacking as Strategic Abstention in Legal Reasoning Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/271", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1509" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/271" ], "reason": "审查通过，待核心算法实现与公开验证。公开reward hacking诊断入口：LegalBench16任务分离surface proxy、真实可解析正确性、CTS与滚动GDM；Apache-2.0仓库有notebook与结果。刻意错误proxy训练用于检测format/abstention坍塌，不得宣传为能力改进。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06446", "title": "The Assistance Dilemma: Learning to Teach via Multi-Turn Reinforcement Learning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/270", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/778" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/270" ], "reason": "审查通过，待核心算法实现与公开验证。tutor无gold solution；masked近迁移post-test只保留学生tokens，配合correctness/no-handover gate；冻结学生与跨benchmark评估。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06454", "title": "AgentPrivArena: Evaluating and Auditing Real-world AI Agent Privacy", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/269", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/777" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/269" ], "reason": "审查通过，待核心算法实现与公开验证。AgentPrivArena/AgentPrivAudit 在六个容器服务中做389可执行隐私任务，读后提取来源/主体/接收方信息流、写前Pass/Abstract/Block，可作为真实工具状态评测与防护策略候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06479", "title": "Behavior-Preserving KV Cache Compression", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/268", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/775" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/268" ], "reason": "审查通过，待核心算法实现与公开验证。Behavior-Preserving KV 以全缓存下一token分布为目标，投影删除扰动经最终norm Jacobian/LM head近似到logits，用singleton候选池和mask-level边际KL贪心优化缓存。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06598", "title": "SimForcing: Distilling Simulation Motion Priors into Real-Domain Robot World Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/259", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/766" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/259" ], "reason": "审查通过，待核心算法实现与公开验证。SimForcing动作条件视频worldmodel，真实轨迹sim重放扰动训teacher，pairedsim/real同noise/time的相邻cleanlatent差L2蒸馏；逐block zero-initconv simcondition按样本共享dropout并噪声腐坏。Bridge/InternData视频质量及LIBERO下游，世界预训练不预测动作，sim侧刻意排除objectinteraction不能夸大为学通用物理。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06611", "title": "Lens3D: Target-Conditioned Visual Foveation for Fine-Grained 3D Understanding", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/257", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/764" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/257" ], "reason": "审查通过，待核心算法实现与公开验证。Lens3D先grounder目标box→depth-visible voxel增量coverage选多视图并wireframe标注→2DVLM；LensDistill privileged teacher caption监督native3DLLM+目标center，无test额外view/VLM。LensBench2068object/141valscene与561train分离但reference为3模型silver，100object人评；GTbox版本不得混作端到端grounding。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06616", "title": "Video Encoders Built on Image Representations", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/255", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/762" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/255" ], "reason": "审查通过，待核心算法实现与公开验证。image-first视频输入保留源位置，固定token预算下DPP/跨帧关联选择加可选18M时序refiner；Appendix C.1训练自由对照已隔离额外训练。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": "not_reproduced_high_cost", "this_audit_full_experiment_review_done": true, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06617", "title": "RealtimeWAM: One-Step Asynchronous World Action Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/254", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/761" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/254" ], "reason": "审查通过，待核心算法实现与公开验证。RealtimeWAM以teacher10步endpoint的interval velocity锚定CD，λ=.2等价1/t² endpoint权重，frozenvideo共享KV只LoRA action；CUDA双stream每层KV projection后event、action attention才wait的wavefront保持依赖。LIBERO/RoboTwin/Plus及同训练distill控制，延迟含VAE不含episode一次文本。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06636", "title": "Differentially Private Mixing of Public Datasets Improves Private Learning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/253", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/760" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/253" ], "reason": "审查通过，待核心算法实现与公开验证。DP-MixMin对公共预训练源权重进行有隐私会计的凸代理优化，是基础LM数据混合候选，不仅医疗应用；GPT2-Medium两个文本域有完整预训练/DP微调实验。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06637", "title": "Long-Horizon Textual World Modeling through Structured Reasoning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/252", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1508" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/252" ], "reason": "审查通过，待核心算法实现与公开验证。真实encode-transition-decode rollout tree按兄弟组credit，predictive gain与中间状态reward共同训练；无需gold textual state。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06647", "title": "LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/251", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/759" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/251" ], "reason": "审查通过，待核心算法实现与公开验证。直接累积随机投影gradient sketch、低秩update与predicted-KL步长；PPO实际多scale训练，H100结果不能外推A100。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06648", "title": "Representation-Space MMD for Diffusion Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/250", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1507" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/250" ], "reason": "审查通过，待核心算法实现与公开验证。冻结DLM token表示RBF-MMD，离散单denoiser pass采样后group REINFORCE，连续bootstrap后求导；OWT/GSM8K及16B DMax实验。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06650", "title": "Wikidata Search Traces: A Dataset for Training Knowledge Graph Search Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/249", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/758" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/249" ], "reason": "审查通过，待核心算法实现与公开验证。Wikidata约束图构造唯一可验证问题，持久Python REPL组合13类图查询函数，保留成功搜索轨迹，适合训练数据/Agent工具harness候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06666", "title": "What Matters for Latent Reasoning with Flow Matching", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/247", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/756" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/247" ], "reason": "审查通过，待核心算法实现与公开验证。FLaRe以symbolicCoT VAE+noise-shiftedFM和verified rollout全路径反传，包含独立可执行latent reasoning训练机制，不只是测量综述。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06671", "title": "Learning What to Imitate: Entropy-Aware Distribution Mixing", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/245", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1506" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/245" ], "reason": "审查通过，待核心算法实现与公开验证。ConMix/GeoMix由student top-K归一entropy控制teacher比重，用于采样SFT与forward KL；两种混合本身已有AMiD，新增是entropy调度。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06686", "title": "OVAL: Output-Aware Local Page Bases for KV Cache Retrieval", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/243", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/753" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/243" ], "reason": "审查通过，待核心算法实现与公开验证。OVAL 对键能量与键值输出敏感度矩阵做归一化混合，以主特征子空间构建页表示；decode 用低秩重建 logits 的 logsumexp 选页，无需每次读取值统计。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06729", "title": "Improving Diversity in LLM Short Story Generation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/238", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1505" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/238" ], "reason": "审查通过，待核心算法实现与公开验证。DivLM的CPT指令残差恢复与群体genre/tone/style/entity多样性奖励为可执行两阶段训练配方，需真实judge与语料。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06748", "title": "BazaarBench: Delegation Safety in Decentralized C2C Marketplaces Run by LLM Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/236", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/747" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/236" ], "reason": "审查通过，待核心算法实现与公开验证。BazaarBench 以可复制30天市场状态和固定20代理群做7天continuation，事件/库存真值与交易安全阶段分开，适合多Agent委托约束评测。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06750", "title": "Balancing Memory Pathways: Analyzing and Improving Memory Utilization in Hybrid LMs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/235", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/746" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/235" ], "reason": "审查通过，待核心算法实现与公开验证。保留混合LM两条记忆路径，额外attention跨segment屏蔽的递归辅助SFT损失有公开实现，符合可执行训练operator。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06751", "title": "MatrixFormer: A Foundation Model for Matrix Completion", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/234", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/745" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/234" ], "reason": "审查通过，待核心算法实现与公开验证。二维cell grid、行列轴向注意力与synthetic prior训练的78M通用矩阵补全模型，作为表格基础模型研究候选，而非工业推荐AB通过。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06782", "title": "T-Search: An Open Agentic Retriever and Playground for Hard Multi-Step Search", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/233", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1503" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/233" ], "reason": "审查通过，待核心算法实现与公开验证。公共检索Agent入口：冻结语料search/save_and_advance/finalize工具及多轮证据状态，TRuST/SynthComp公开基准；官方Apache-2.0 harness已可读。final-set Recall@10与生成Agent trajectory recall不可直接横比，融合3次须计预算。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06804", "title": "Sharpen Without Search: On-Policy Distillation of Sequence-Level Power Distribution", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/230", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1502" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/230" ], "reason": "审查通过，待核心算法实现与公开验证。Rényi梯度对应teacher幂分布×student分布的加权MLE，SMC沿student proposal重采样；GSM8K训练不读reference答案，保留整条SMC训练链。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06817", "title": "Paradee: Distilling Kokoro-82M into an 8M-Parameter Single-Voice Text-to-Speech Model", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/229", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/740" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/229" ], "reason": "审查通过，待核心算法实现与公开验证。Paradee 将 Kokoro 单音色蒸馏为 8.07M TTS，分阶段训练文本与解码器，保留多分辨率频谱损失、GAN 精调和相位锁定。WikiText103 教师生成语音与 200 句评测可重建。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06829", "title": "CLIFT: Conformal Self-Verification for Web Agent Training and Test-Time Scaling", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/227", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1501" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/227" ], "reason": "审查通过，待核心算法实现与公开验证。conformal自验证非负bonus和URL-stratified return优势改变credit；WAI训练、VWA bank transfer、OM2W无policy训练须分别报告。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06830", "title": "MemPilot: Orchestrating On-Demand Multimodal Memory Curation for LLM Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/226", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/738" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/226" ], "reason": "审查通过，待核心算法实现与公开验证。质量/成本/延迟分别组归一后合并，prefix probe边际答案质量修正阶段credit；训练latency用proxy estimator，不能冒称直接壁钟reward。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06843", "title": "Recursive Video In-Context Learning for Agentic Robot", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/224", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/737" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/224" ], "reason": "审查通过，待核心算法实现与公开验证。RV-ICL的多层示范索引与按子目标重入是明确Agent上下文访问机制；只接受示范条件化诊断候选，须将阶段目标/仿真状态来源与非gold能力评测隔离。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06844", "title": "Learning to Read the Contextual Tokens in Diffusion Transformers", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/223", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/736" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/223" ], "reason": "审查通过，待核心算法实现与公开验证。Contextual Reader用跨MM-DiT层contexttokens加layer/timestep embedding→2层Qformer→冻结LLM QA探测；核心CoAl以单早层1query aligner负cos匹配冻结SigLIP clean图teacher，jointflow训练并对null-condition加权抑制prompt捷径，推理丢aligner。COCO从零与SD3 Fine-T2I微调分开验证，不能把readoutprobe当新生成能力。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06892", "title": "Axiom Satisfiability of Linear Rewards in Alignment", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/219", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1430" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/219" ], "reason": "审查通过，待核心算法实现与公开验证。LP3在有限偏好图上共同最小化candidate slack和线性reward margin违反，以LP松弛替代ILP；Habermas真实偏好验证，不含LLM策略训练，也不保证未见response公理。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06932", "title": "RADC: Risk-Aware Dual Caching for Vision-Language Test-Time Adaptation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/213", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1033" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/213" ], "reason": "审查通过，待核心算法实现与公开验证。RADC前景/全局双cache与高斯风险准入是具体VLM TTA候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06947", "title": "FactorBench: A Portfolio-Aware Benchmark for Automated Factor Mining", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/212", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1497" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/212" ], "reason": "审查通过，待核心算法实现与公开验证。公共评测：统一symbolic/Python factor→date×asset signal契约，5市场、训练2017–21/验证2022–23/测试2024–25，组合和成本回测。官网仓库含data/evaluator/mining入口，可作为专门的自动研究任务，不宣称普适LLM能力或实盘A/B。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06950", "title": "Learning to Decide, Not to Reason: Parameter-Efficient Decision Operators via Low-Rank Activation Steering", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/210", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1004" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/210" ], "reason": "审查通过，待核心算法实现与公开验证。DecSteer训练共享低秩activation residual与token门控，初始identity但输出支路有梯度，集中学习决策而非全模型；真实SearchQA/LiveMath BC及多基准，需保留RMS+SiLU+gate结构。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06963", "title": "WavePrune: One period is often enough for RoPE", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/208", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/980" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/208" ], "reason": "审查通过，待核心算法实现与公开验证。WavePrune将每个RoPE二维通道的QK贡献限制在首旋转周期，FlashWavePrune按16维组取最大窗口实现TensorCore稀疏算子，可用于推理及重新预训练。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06964", "title": "Principles that Guide, Actions that Inform: Agent Evolution via Knowledge Abstraction", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/207", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/978" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/207" ], "reason": "审查通过，待核心算法实现与公开验证。SAGA 将执行轨迹抽象为分层经验/原则，检索后适配当前任务，并可 SFT/GRPO 优化 actor。ScienceWorld/ALFWorld 的证据 grounding 与 utility ledger 规则可执行。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06966", "title": "APEX: Active Protection at Execution Boundaries for LLM Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/206", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/976" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/206" ], "reason": "审查通过，待核心算法实现与公开验证。APEX 冻结用户授权合同，在执行边界以 WRAP 核验效应授权，以 PLANT deception receipts 暴露未获准信息使用，失败安全延续。AgentDojo/ASB/MCPTox/MSB/Skills 公开任务支持验证。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.06993", "title": "DART-ES: Difficulty-Aware Reweighting and Targeted Replay for Fine-Tuning LLMs with Evolution Strategies", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/202", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/936" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/202" ], "reason": "审查通过，待核心算法实现与公开验证。DART-ES用扰动种群正确率EMA构造连续难度权重，并以掌握阈值和零成功耐心控制稀有可解样本重放；实际LLM数学ES对比，不能把历史正确率oracle放入评价agent。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07043", "title": "TRIAGE: Direction-Aware Mismatch Stabilization of Native NVFP4 Reinforcement Learning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/197", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1494" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/197" ], "reason": "审查通过，待核心算法实现与公开验证。TRIAGE在负优势负gap token按segment门控并保持响应更新质量归一化，正优势加有界pseudo-Huber修复；真实NVFP4 learner/sampler实验，A100仿真不能宣称原生吞吐。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07118", "title": "AMBER: Training Long-Horizon Web Agents through Append-Only Memory", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/190", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/749" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/190" ], "reason": "审查通过，待核心算法实现与公开验证。append-only两回合memory/action共享GRPO终局优势，插入模板token mask；实验先SFT初始化，不得声称完全无SFT；需web环境。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07121", "title": "SoloQ: Calibration-Free Quantization for Diffusion Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/189", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/744" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/189" ], "reason": "审查通过，待核心算法实现与公开验证。SoloQ K-RPBH在permuted blockHadamard后Haar跨块混合，weight E8P/activation Lloyd或NVFP4分布quantizer，无校准数据；scale ||z||²/修正仅恢复投影非无误差。blockdiffusion仅commit KV量化、active保持高精度。五dLLM实验，H200 accuracy/低位硬件性能须区别A100可运行性。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07132", "title": "CroissantMiner: Automated Extraction and Validation of Croissant Metadata for ML Datasets", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/187", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/735" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/187" ], "reason": "审查通过，待核心算法实现与公开验证。CroissantMiner含102human-gold+500silver的30字段paper metadata，singlepass及4agentic抽取，rule/LLM分字段score与NULL惩罚；Sonnet4.5预填不参与排名，judge最终κ.708/hardset.663应替代初始.890。外部metadata enrichment与paper-only gold有边界冲突需分开；官方MIT仓库有data/evaluation/CLI，可接evolve信息抽取评测。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07177", "title": "CLM-as-a-Judge: Evaluating an Open Contrastive Decision Model on Public Judge Benchmarks", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/185", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1493" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/185" ], "reason": "审查通过，待核心算法实现与公开验证。公共judge评测与校准级联协议：calibration/eval隔离、双顺序、常数baseline及成本—风险读数。arXiv官方abstract页已枚举ancillary的pins、split、per-item预测、judges/run_pass/metrics/analysis源码。托管strong judge不可固定revision；本次未运行代码，单个附件浏览器读取报错。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07191", "title": "Agentic Design Space Exploration for Joint Hardware Configuration Selection and Mapping of AI Inference Workloads on Heterogeneous Edge SoCs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/184", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/733" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/184" ], "reason": "审查通过，待核心算法实现与公开验证。TraceDSE以真实硬件执行trace→工具化critic→proposer→Pareto/exploration state组成实测搜索闭环，适合evolve算子机制候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07212", "title": "Reward-Driven Learning under Prompt-Level Differential Privacy", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/183", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1492" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/183" ], "reason": "审查通过，待核心算法实现与公开验证。DP-GRPO以prompt group为DP-SGD裁剪记录、单更新GRPO为record loss；隐私会计与实际Qwen1.5B LoRA实验，保留DP机制而非仅加噪声示意。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07247", "title": "Learning What to Distill: Bilevel Top-K Token Selection for Self-Distillation in Large Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/180", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1490" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/180" ], "reason": "审查通过，待核心算法实现与公开验证。Soft-OR效用、softplus阈值内层与soft mask形成近似bilevel选择；阈值detach、全词表forward KL。200步最终checkpoint；TIP比较同时改变divergence，不可单独归因选择器。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07258", "title": "Lineage-Aware Memory Governance: A Derivation-Gated Framework for Privacy-Preserving Column-Level Access Control in Enterprise AI Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/179", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/731" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/179" ], "reason": "审查通过，待核心算法实现与公开验证。Lineage-Aware Memory Governance以敏感列派生标签、最新安全记忆检索及定义hash冲突提示实现明确可复现的共享Agent记忆治理；接受为机制诊断候选而非生产安全能力。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07261", "title": "Verifying Coordination in Parallel Coding Agents: NP-Bench and a Scheduling Planner", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/178", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/730" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/178" ], "reason": "审查通过，待核心算法实现与公开验证。Nerveplane被动观察git状态，按文件scope冲突着色分批并按契约依赖排序合并，NP-Bench含真实git集成与live coding agents，可作并行evolve协作评测候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07289", "title": "Catching Developers in the Flow: Low-Latency Agentic Program Repair at Google Scale", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/176", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/728" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/176" ], "reason": "审查通过，待核心算法实现与公开验证。Google FlowAgent以预/后执行过滤、ReAct generate-and-validate和最终测试完成低延迟presubmit修复；核心候选接受，未实现。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": "source_unavailable_archived", "this_audit_full_experiment_review_done": true, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07311", "title": "Understanding and Mitigating Inference-Time Overreliance Using Agentic Memory", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/175", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/727" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/175" ], "reason": "审查通过，待核心算法实现与公开验证。MemTrim用LLM规范化原子证据+task signature/answer support，独立key-value trie索引；删与当前重复或矛盾证据、support变化则抑制旧answer、task变化则typed reuse，接受memory治理机制候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07324", "title": "Scale-Invariant Training for Time Series Foundation Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/174", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/726" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/174" ], "reason": "审查通过，待核心算法实现与公开验证。ScaleIn将loss置于context-normalized target空间，作为时间序列基础模型training-config算子候选；本文是已存在策略的系统理论/实证审计，不宣称全新网络。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07332", "title": "Structuring MoE Expert Selection for Agentic Reinforcement Learning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/173", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/725" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/173" ], "reason": "审查通过，待核心算法实现与公开验证。operation标签和router互信息、相邻token专家集合一致性及entropy gate联合agentic RL；AppWorld/AutomationBench真实训练，需模型路由不可用proxy。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07335", "title": "Selective Critique for Cost-Aware LLM Agents in Long-Horizon Decision Making", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/172", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/724" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/172" ], "reason": "审查通过，待核心算法实现与公开验证。SAG对可执行动作长度归一loglik分布算entropy+top2margin门控，无gold critic纠偏；每10成功episode从同base重新LoRA-SFT累计action pairs，接受真实门控+在线学习候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07348", "title": "Stepped MoE: Segment-Level Routing with Configurable Inference Complexity", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/169", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/721" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/169" ], "reason": "审查通过，待核心算法实现与公开验证。SteppedMoE以前12dense层一次预测后44层专家集合，每段首token刷新、预算embedding与多稀疏度loss缩放联合训练，真实12B模型1Ttoken支持可变计算架构候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07349", "title": "RELACE: retrospective likelihood-based action credit estimation for long-horizon language agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/168", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/720" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/168" ], "reason": "审查通过，待核心算法实现与公开验证。执行动作原始/结果条件似然比修正局部return，终局信息仅用于训练credit；ALFWorld/WebShop真实训练，但主表混用文献baseline，应本地同预算复核。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07376", "title": "MemCo: Memory-Centric Collaboration for Generalizing LLM Agents to Unseen Environments", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/165", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/717" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/165" ], "reason": "审查通过，待核心算法实现与公开验证。MemCo本地trajectory graph提取条件化规则，episode-reduced正负证据经role抽象exact-key聚合，Wilson分数促进/撤回global records，state/phase/goal检索+grounding组合；接受Agent memory核心候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07385", "title": "Semantic Capability Acquisition and Specialization During Vision-Language Model Fine-Tuning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/164", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/716" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/164" ], "reason": "审查通过，待核心算法实现与公开验证。SSR训练时NAME/ATTRIBUTE routed causal隔离、融合summary及可选NAME dropout=.5，评测恢复普通EOT；P3无name及CAPP反事实区分描述可读性。6CLIP类模型跨数据验证，P3轨迹峰值为回溯诊断不是可部署选模，部署按P1选择。未保留fork revision及subset生成脚本降低精确复现强度但核心训练定义可实现。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07389", "title": "Inference and learning in sparse autoencoders as natural gradient flow", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/163", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/715" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/163" ], "reason": "审查通过，待核心算法实现与公开验证。BeFOND将Bernoulli稀疏编码的推断与字典学习同时构造为natural-gradient flow，具closed-form更新、可执行官方核心与公共SAEBench任务，适合作为表示诊断/干预算子候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07457", "title": "AlignQuant: Tile-Aligned Mixed-Precision Quantization for Efficient LLM Generation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/159", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/711" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/159" ], "reason": "审查通过，待核心算法实现与公开验证。AlignQuant 以64×64 tile统一量化分配、打包与GPU执行；prefill/decode Fisher敏感度分别归一化后取最大值，低敏感tile使用4bit，其余8bit；实际整数路径是核心组成。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07522", "title": "Activation Denoising: A Robustness View on Parallel vs Sequential LLM Quantization", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/151", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/704" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/151" ], "reason": "审查通过，待核心算法实现与公开验证。Activation denoising 用干净/噪声前向估计 H/C/N 二阶矩，先线性预变换权重，再在新度量下 GPTQ/QuIP# rounding，使各层可并行量化。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07532", "title": "Safeguarding LLMs via Model-Agnostic Latent Safety Signals from Dark Knowledge", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/150", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/703" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/150" ], "reason": "审查通过，待核心算法实现与公开验证。LADE按有害/良性首token概率差选安全token，跨tokenizer首有效子词和重复映射权重，L1归一特征到有害参考集的kNN距离门控；可实现防御机制。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07533", "title": "SkillFormer: Skill-Decomposed Adaptation for Audio Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/149", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/702" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/149" ], "reason": "审查通过，待核心算法实现与公开验证。SkillFormer question embedding kmeans六技能、r16并行adapter及question top2router，先每cluster500步再balanced2000步joint校准；50kQA、三audio模型/三bench。§2.4 PhaseB明确解冻LLM，与PhaseA“throughout冻结”用词有歧义，复现按分阶段并报告训练参数成本，不能宣传全程仅LoRA。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07557", "title": "CheckerBench: Can Long-Horizon Agents Synthesize Static-Analysis Checkers?", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/142", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/696" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/142" ], "reason": "审查通过，待核心算法实现与公开验证。CheckerBench/CheckerLab 按补丁与漏洞前后版本生成静态检查器，冻结产物后独立重建、差分、误报和语义门禁；300任务公共包及固定Linux runtime适合代码Agent评测。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07572", "title": "Two Vectors Replace In-Context Demos: Structured Task Adaptation via Embeddings", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/139", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/693" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/139" ], "reason": "审查通过，待核心算法实现与公开验证。STAVE在frozenLM/LMM输入结构组加两向量、paired有/无demo监督，是真实2d参数适配而非gold-inference或每层模拟。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07585", "title": "REViT-v2: Hierarchical Windowed Roto-reflection Equivariant ViT for Equivariant Feature Extraction", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/136", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/690" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/136" ], "reason": "审查通过，待核心算法实现与公开验证。REViT-v2以group-equivariant QKV卷积、完整representation-field多头局部窗口及群卷积downsample构成层次ViT，O(NM²C)非globalN²；ImageNet及RotMNIST实测，含同stem global控制。需保留有限p4m群/窗口划分几何条件，不能泛称任意角旋转严格等变。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07587", "title": "Large Language Model Orchestration under Heterogeneous Preferences via Explicit Persona Inference", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/135", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1487" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/135" ], "reason": "审查通过，待核心算法实现与公开验证。HARP以有限persona库、离线校准response scorer更新每Agent贝叶斯后验，posterior sampling集中规划；HARP+加reward-disagreement探索bonus。接受精确算法/机制诊断候选，不宣称开放对话Agent能力提升。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07639", "title": "HarnessSecurity-Bench: Do Security Mechanisms Really Protect Coding Agent Harnesses?", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/130", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/685" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/130" ], "reason": "审查通过，待核心算法实现与公开验证。HarnessSecurity-Bench 在固定合法任务下对原生安全机制ON/OFF配对，分离外部确定性utility与attack-effect oracle，适合离线隔离harness回归。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07643", "title": "Monte Carlo Estimation for KV Cache Eviction", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/128", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/683" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/128" ], "reason": "审查通过，待核心算法实现与公开验证。LORE-KV 从冻结模型采样多条短续写估计未来 query，以中心化 log-prob、熵和分歧加权，再按输出投影后的反事实删除代价选择 KV；采样轨迹随后丢弃。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07645", "title": "SkillPoison: Progressive Skill Poisoning via Successful Experiences", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/127", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/682" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/127" ], "reason": "审查通过，待核心算法实现与公开验证。SkillPoison揭示真实成功经历被提炼为脱离适用条件的技能时产生的治理漏洞，可作为防御性技能晋级回归评测/机制诊断候选，不接受为能力提升策略。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07654", "title": "Does On-Policy Distillation for Safety Pose Backdoor Risks?", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/126", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1486" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/126" ], "reason": "审查通过，待核心算法实现与公开验证。Lazy Defense对teacher-student sampled token log-ratio双侧clipping再stop-gradient，四teacher/student组合测试仅延迟backdoor，不是保证消除。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07657", "title": "Where Rules End and Judges Begin: Measuring the Judgment Boundary in Multi-Agent Systems Security", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/125", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/681" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/125" ], "reason": "审查通过，待核心算法实现与公开验证。DEFER提供27确定/相似度检查后残余LLM panel的实际host工具/记忆安全管线、四域testbed与outcome oracle，公开MIT代码/attacks/benchmarks/logs/scripts/tests，可作安全评测候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07689", "title": "Unlocking Fine-Grained Perception in CLIP via Structurally-Aware Latent Masked Modeling", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/123", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/678" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/123" ], "reason": "审查通过，待核心算法实现与公开验证。SALM通过patch关系/范数差双矩阵和teacher latent masked reconstruction改造CLIP，含无外部teacher的SALM-Self，核心loss和训练课程可保真复现。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07706", "title": "WASD: Wasserstein-based Knowledge Distillation for Large Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/121", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/676" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/121" ], "reason": "审查通过，待核心算法实现与公开验证。WASD用teacherembedding成本的去偏Sinkhorn divergence及stopgrad双potential差优化学生概率，有官方训练/评测脚本与等wallclock对照。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07729", "title": "Foveated Compression: Selective High-Resolution Preservation for Token-Efficient VLMs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/120", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/675" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/120" ], "reason": "审查通过，待核心算法实现与公开验证。Foveated Compression的混合分辨率表示与tie-aware region selector有明确算法，适合作为token压缩/选择诊断候选；不将test-oracle或弱于resize的真实结果包装为提升。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07731", "title": "Learning to Retrieve via Reinforcement Learning in Embedding Space", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/118", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1484" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/118" ], "reason": "审查通过，待核心算法实现与公开验证。vMF嵌入策略+G×G product rollout+RLOO，CMP投影降噪；ReasonRank训练、BRIGHT评测、三paired seeds。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07739", "title": "Cite What You Explore: Budget-Aware LLM Reasoning over Medical KGs with Verifiable Evidence", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/117", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1483" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/117" ], "reason": "审查通过，待核心算法实现与公开验证。证据有/无的同snapshot预测loss差构成paired gain，叠加预算与citation integrity，REINFORCE训练检索reasoner；需真实MIMIC/PrimeKG映射，非手写分数替代。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07758", "title": "Later Is Better: Token Reduction for ViTs Under Distribution Shift", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/112", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/670" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/112" ], "reason": "审查通过，待核心算法实现与公开验证。Late-concentrated幂律层间token预算是明确可执行的ViT/VLM压缩算子，官方代码公开；多shift、多backbone和五类reducer在保守iso-GFLOPs下有收益。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07767", "title": "TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/109", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1482" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/109" ], "reason": "审查通过，待核心算法实现与公开验证。将训练activation相邻FP4 codeword选为更接近rollout codeword并缓存mantissa/scale；四MoE训练验证。原native NVFP4路径需相应硬件，不得A100模拟冒称原速度。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07774", "title": "Reading, Not Manipulating: Leveraging Router Logits for Multimodal Safety in MoE Vision-Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/108", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/666" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/108" ], "reason": "审查通过，待核心算法实现与公开验证。将MoE各层最终prompt token router logits拼接，train统计标准化、class-balanced L2 logistic detector，prefill后τ=.4(val选)拒绝，否则原模型生成；两MoE模型HoliSafe及OOD、MMMU。安全信息可读不等于可通过expert steering控制，残差state也有效；不宣称router独有。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07809", "title": "MASKerade: Token-Routed Mask Experts for Dense-to-MoE Upcycling", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/104", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/663" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/104" ], "reason": "审查通过，待核心算法实现与公开验证。MASKerade在冻结FFN上训练多专家binary mask与token router，真实hard mask+STE、top1保留softmax概率确保router梯度；等标称激活预算和mask种类对照支持多模态MoE适配候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07810", "title": "SIFT: Search Intent-to-Filter Transformer for Multi-Task Personalized Filter Ranking at Airbnb", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/103", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/661" ], "track": "recommendation", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/103" ], "reason": "审查通过，待核心算法实现与公开验证。SIFT 的随机身份分流与酒店扩展线上证据满足门槛。过滤器点击人数 +20%，普通 booking +0.03% 不显著；酒店扩展 marketplace uncancelled booking +0.76%。原始旅程/过滤器/容量标签未公开，接受待实现不等于已实现。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07816", "title": "Do I Need the Cloud? Uncertainty-Aware Step-Level Handoff for Small Language Model Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/102", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/660" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/102" ], "reason": "审查通过，待核心算法实现与公开验证。StepGate以本地动作反事实失败标签训练校准门控，结合六类特征、阈值/预算、按需一致性采样逐步选择本地/验证/强模型；接受为真实模型路由机制候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07819", "title": "$α$Transfer: Coefficient Transfer for Efficient Model Merging", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/100", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/658" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/100" ], "reason": "审查通过，待核心算法实现与公开验证。alphaTransfer将小proxy搜索到的全局/任务/层合并系数转给同family大模型，层系数按相对深度复制/插值；目标taskvector编辑仍在大模型原生执行。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07832", "title": "Harness Engineering for Software Engineering via Modular Executable Dev-Primitives", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/98", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/656" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/98" ], "reason": "审查通过，待核心算法实现与公开验证。HERMES将持久文件包装为Dev-Primitive，按依赖动态激活、定向消息、执行诊断后至多3轮再激活修改；接受为文件级Agent编排核心机制。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07848", "title": "Dynamic Positional Attention Modulation for Parameter-Efficient Fine-Tuning of Large Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/95", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/654" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/95" ], "reason": "审查通过，待核心算法实现与公开验证。DyPAM根据hidden低秩特征生成RoPE配对维度Q/K缩放，叠加head/layerbias，core PEFT公式完整且有验证选checkpoint和多seed公共任务。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07851", "title": "RA-MoWE: Workflow-Affinity Embeddings for Query Clustering and Agentic Workflow Generation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/94", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/653" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/94" ], "reason": "审查通过，待核心算法实现与公开验证。RA-MoWE离线probe outcome affinity→cluster expert search→BGE+MLP预测路由是真实可执行工作流优化候选；只能接受text-only Guided-BB路径，test measured affinity列必须诊断隔离。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07860", "title": "WorkflowOps: Learning Agent Collaboration Priors for Multi-Agent Workflow Orchestration", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/93", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/652" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/93" ], "reason": "审查通过，待核心算法实现与公开验证。WorkflowOps以Minilm能力匹配、低充分性新Agent生成、合成workflow一阶转移先验添DAG软边/分层执行构成明确orchestration候选，但非经证明持续自演化。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07862", "title": "A self-learning scientific agent for X-ray diffraction", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/92", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/651" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/92" ], "reason": "审查通过，待核心算法实现与公开验证。Gan Jiang将XRD工具执行残差/失败转换为版本化skill修订并开发集成对选型、冻结后heldout比较；接受科学Agent机制及DeltaXRDbench公共评测候选，不能宣称完整闭源toolchain复现。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07868", "title": "CueRator: Agentic Search for Symbolic Rules to Adapt Frozen Multimodal Encoders", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/90", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/649" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/90" ], "reason": "审查通过，待核心算法实现与公开验证。CueRator在冻结ImageBind上agent发现AST约束closed-form门限式，训练数据Sobol per-video oracle仅诊断结构上限，另训练Gaussian contextual policy用K4 leave-one-out REINFORCE预测10参数；3×20候选val选后测test。可接evolve operator但不可给部署agent访问oracle GT；编码器不训练、seen-only训练且val包含unseen。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07887", "title": "Visual Abstention in Unified Multimodal Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/87", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/645" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/87" ], "reason": "审查通过，待核心算法实现与公开验证。VisTA在BAGEL上用feasible/infeasible配对：先reason再image或ABSTAIN，文本CE均训练、flowmatching仅feasible；38224对与1050对DoD图像/maze隔离。无reminder拒绝93.0%、误拒0.8%、edit74.3；paired minimal-change约束定义不可行，不能泛称现实世界不可能或安全对齐。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07898", "title": "FC-SWE: Failure-Conditioned RL for Long-Horizon Software Engineering Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/86", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1479" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/86" ], "reason": "审查通过，待核心算法实现与公开验证。真实R2E-Gym训练，失败时reset并提供patch/verifier上下文；每次attempt奖励独立、未执行padding不计优势。SWE-bench结果为verifier-assisted不是独立pass@K。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07903", "title": "Unsupervised Long-Tailed Adaptation of Vision-Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/85", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/644" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/85" ], "reason": "审查通过，待核心算法实现与公开验证。MARS冻结CLIP学prompt/文本adapter：zero-shot视觉类中心识别unsupported概率迁移并修正weak目标，strong-view CE；再以teacher软类别频率与margin EMA控制KL/伪标签CE权重。9数据γ10/20/50三seed，未知先验但标签集合已知，非开放类别发现。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07925", "title": "TF-PRVR: Training-Free Partially Relevant Video Retrieval", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/83", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/642" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/83" ], "reason": "审查通过，待核心算法实现与公开验证。TF-PRVR冻结VLM，帧特征速度方向变化的db4 DWT分层边界、层内语义/时间与相邻层IoU图，闭式(1−β)(I−βW)^−1传播后逐时跨尺度平均取最大检索分。三标准retrieval数据及三backbone，β=.7固定；离线图逆矩阵/特征成本应计，非新VLM训练。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07935", "title": "SIGMA: Self-Improving Alignment Generalization from a Model Spec", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/82", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/641" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/82" ], "reason": "审查通过，待核心算法实现与公开验证。Apple/JHU SIGMA的同模型taskdesigner+配对一事实变化+spec/rubric自验收+训练监督生成是Agent自改进候选；保留真实模型judges与GRPO，不能以规则fixture代替。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07940", "title": "Hybrid Latent Attention for Looped Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/80", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/639" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/80" ], "reason": "审查通过，待核心算法实现与公开验证。HLA为looped LM跨循环共享旋转潜KV并保留精确近期窗口；冻结骨干训练writer/reader，真实vLLM/A100同引擎速度与质量测量支持系统候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07972", "title": "Can Agents Work for Everyone? Cross-User Reliability for Mobile GUI Agents in Personalized User Interfaces", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/75", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/634" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/75" ], "reason": "审查通过，待核心算法实现与公开验证。PAIR可执行个性化GUI状态与RePAIR按跨用户UC子目标四rollout标准化优势构成可复现机制设计；接受为探索候选，接入必须重做独立validation选择。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07987", "title": "VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/73", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/632" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/73" ], "reason": "审查通过，待核心算法实现与公开验证。VisionWeave可学习gated 2×2 pooler+ViT中间granularity router，细/粗MRoPE中心对齐；三阶段teacher top512前向KL蒸馏依次pooler、soft-mix router、hard-route pooler+LLM，非OPD。训练子集校准50%节省，八bench和SGLang双A100长视频实测2.30×吞吐，FastV改为encoder后版本须标注。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.07990", "title": "A Broader Look at Model Merging: Rethinking Implicit Regularization Induced by Task Arithmetic", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/72", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/631" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/72" ], "reason": "审查通过，待核心算法实现与公开验证。S-RandOpt以辅助数据梯度扩展task-vector子空间一维后随机采样选模，另比较真正weight-GD与系数GD；支持模型合并搜索研究候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08048", "title": "DAEDALUS: Bootstrapping Agent Memory from Self-Generated Tasks", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/70", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/628" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/70" ], "reason": "审查通过，待核心算法实现与公开验证。DAEDALUS用自生且Explorer亲自完成的任务、失败提炼heuristic、reset后连续3次成功验收、难度调节与冻结整库注入；真实记忆生成/验证机制适合evolve候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08076", "title": "SpeedrunBench: Challenging LLM Agents with Video Game Speedrunning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/68", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/626" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/68" ], "reason": "审查通过，待核心算法实现与公开验证。SpeedrunBench 在固定轮次及计分调用预算下优化可回放动作轨迹；公开三款开源游戏、固定revision和独立计分容器，适合evolve公共评测。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08101", "title": "Beyond Corrected Memory: Execution Consistency in Multi-Agent Systems", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/63", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/623" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/63" ], "reason": "审查通过，待核心算法实现与公开验证。CAVERT将义务激活/观察覆盖/有效履行分离，证据不足返回U而非无违规，LLM提取+deterministic规则+precommit read-refresh gate是明确执行一致性机制。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08106", "title": "ChartBmkAgent: Harness-Governed Multi-Agent Construction of Chart QA Benchmarks from Sparse Error-Taxonomy Specifications", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/62", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/622" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/62" ], "reason": "审查通过，待核心算法实现与公开验证。ChartBmkAgent 将错误分类/难度作为不可变根约束，由 harness 审批数据、图表、机制锚点与 QA，失败局部修复或回退；适合可审计的评测数据生成候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08108", "title": "Enhancing Diffusion Language Models with Autoregressive Post-Training Weights", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/61", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1476" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/61" ], "reason": "审查通过，待核心算法实现与公开验证。A2D task-vector transfer与AR/diffusion增量插值有真实checkpoint benchmark；系数用anchor benchmark选取，本地复现必须改为validation隔离。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08133", "title": "VLA-ACL: Action-Consistent Visual Token Pruning for Efficient Vision-Language-Action Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/59", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/620" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/59" ], "reason": "审查通过，待核心算法实现与公开验证。VLA-ACL冻结VLA仅训练5层patch scorer，Laplace soft Top-K总质量K、黑图embedding平方根gate实现可微训练，L1 full-context teacher+GT动作一致性；推理hard topK保留RoPE位置。LIBERO及实机/A100实测，K64成功97.0对96.8、63.49→46.82ms；K32 Long退化须保留。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08155", "title": "Token-Efficient Multi-Agent Collaboration via System One-Guided Computational Division of Labor", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/57", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/618" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/57" ], "reason": "审查通过，待核心算法实现与公开验证。S1-MAS用冻结Laya离散决策encoder+Qwen0.6B证据定位，把Solve/Decompose/Reverse/Review/Repair/Contrast及停止交给轻量控制器，真实机制适合协同计算预算候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08159", "title": "Making COMET Comparable Across Scripts: Diagnosis and Correction of Tokeniser-Induced Script Bias in Indic MT Evaluation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/56", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/617" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/56" ], "reason": "审查通过，待核心算法实现与公开验证。公开COMET脚本偏差诊断与rank-safe quantile校准/LOLO纠偏统计脚本可作为评测审计基础设施候选；不是新LLM，也不是动态评估器已接入。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08162", "title": "The Failure Is in the Readout: Fine-Grained Emotion Recognition Benchmarks Measure Elicitation, Not Perception", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/55", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/616" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/55" ], "reason": "审查通过，待核心算法实现与公开验证。EmoNet读出审计以40类逐类yes/no首token概率对照自由JSON输出，包含answer-mass和单答案有效性门，可作多模态评价诊断。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08164", "title": "Align, Then Correct: Training-Free Two-Stage Low-Rank Compensation for Extremely Quantized Large Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/54", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/615" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/54" ], "reason": "审查通过，待核心算法实现与公开验证。两阶段低秩补偿：非对称激活目标经全输出 Fisher 加权 SVD，再对现有适配器与 K-FAC 自然梯度之和做秩约束 SVD；第二阶段替换而非叠加 rank-r 适配器。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08183", "title": "Compact Robot Policies Need Fine-Grained Visual Representations", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/53", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/614" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/53" ], "reason": "审查通过，待核心算法实现与公开验证。CoRP用可训练DINOv2+48-query resampler/FiLM文本及小DiT flow动作，λ=.01是跨层task-classification辅助CE（不是信息瓶颈惩罚），训练后丢分类head，10步Euler输出动作chunk。LIBERO/RoboTwin为任务内分布；实机另用私有1万小时/5887任务预训练，不能当仅公开sim数据的sim2real。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08341", "title": "DIPrune: Task-Aware Token Pruning with Dual Importance for Efficient Multimodal Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/41", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/602" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/41" ], "reason": "审查通过，待核心算法实现与公开验证。DIPrune结合文本到视觉注意力倒置rank归一与跨层rank上升的截断归一，保留早层低但后层新出现语义token并移除对应KV；多VLM及视频骨干评测、层对齐消融。LLaVA64tokens平均保留97.7%，POPE TTFT73.23→32.50ms、额外0.24ms，需以实际TTFT而非FLOPs主张速度。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08364", "title": "Transect: Retaining Observability for Long-Horizon LLM Agent Evaluations", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/40", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/601" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/40" ], "reason": "审查通过，待核心算法实现与公开验证。Transect为公共可执行评测观测工具：family Spec固定词表，structural/judged scanners→provenance tables→源链接报告；可用于evolve审计但不是任务能力评分器。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08401", "title": "GeoPID: Decomposing and Steering Visual Information in Vision-Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/38", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/599" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/38" ], "reason": "审查通过，待核心算法实现与公开验证。GeoPID从无标签校准表征估计模态共享/独占子空间，推理按I+αUUᵀ放大vision-unique；22×14组合每组100道独立校准题用标签选层/α后冻结，最终评测不读gold。任务Rényi信息分解为离线诊断，平均相对7.63%需保留308组异质/部分负收益。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08403", "title": "SSR: Sparse Segment Reduction for Ternary GEMM Acceleration", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/37", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/598" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/37" ], "reason": "审查通过，待核心算法实现与公开验证。SSR 将三值矩阵分成正负二值矩阵，按列块模式排序并剔除全零模式，以分段求和与树规约代替乘法；公开 C++/AVX2 实现可核验。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08414", "title": "Image Bitstream Fine-grained Understanding for Privacy-Friendly AIoT", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/35", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/596" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/35" ], "reason": "审查通过，待核心算法实现与公开验证。BFG将JPEG bytes经ByteFormer分层卷积/shifted-window编码为192维，projector到512维后全局上下文融合+局部cross-attention自回归描述，另120类head；CFU-D源ID/split隔离并测试替换/删除/重复腐坏。Dogs分类58.51%采用ImageNet初始化；不解码不等价正式隐私保证，不能包装为通用大规模foundation结论。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08452", "title": "Agentic AutoRAG: RAG Pipeline Optimization through Reasoning-Driven Agents", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/33", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/594" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/33" ], "reason": "审查通过，待核心算法实现与公开验证。Agentic AutoRAG的检索/生成归因→Diagnoser→KB/成本引导Proposer构成明确搜索机制，可接入evolve候选；准确率实验有100验证/300独立heldout隔离。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08463", "title": "UNREAL: Unifying Retrieval and Long-Context with a Single Model", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/32", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/593" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/32" ], "reason": "审查通过，待核心算法实现与公开验证。UNREAL用冻结decoder中层chunk表征、训练64检索token及跨层混合做MaxSim，再重编码所选文本生成；提供multi-positive InfoNCE、公开QA训练split与long-context独立验证协议。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08514", "title": "How Much Evidence Should a Coding Agent's Self-Correction Carry? Adaptive Dirichlet Evidence for Self-Distillation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/29", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/590" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/29" ], "reason": "审查通过，待核心算法实现与公开验证。EESD将公开执行before/after四种转移和词法相关性转为有效样本量1/sum(p²)，Dirichlet utility lower score加权CE并保留未加权forward KL；可复现Agent自修正训练机制。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08560", "title": "Have I Seen Enough? Frozen Video-Language Models Encode Evidence Readiness", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/25", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/586" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/25" ], "reason": "审查通过，待核心算法实现与公开验证。Readiness Gating以冻结Qwen残差的嵌套CV logistic probe预测证据就绪，按训练分位阈值joint readiness/confidence首次停止或deadline最高readiness选答；视频级留出、crossed-quad消除图像/问题独立捷径。固定窗口policy不读证据时间，growing-prefix诊断采样依赖标注不能当部署延迟；§4.4给matched-video-seconds增益。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08586", "title": "MINDSET: Energy-based Schema Evolution for Long Conversational Agent Memory", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/23", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/585" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/23" ], "reason": "审查通过，待核心算法实现与公开验证。MINDSET以immutable episodes和versioned schemas，按distortion/contradiction/history/fragmentation/inconsistency加权能量选择reinforce/supersede/split/new，再展开原始证据检索；定义完整可复现记忆算法。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08621", "title": "Recursive Game Creator: An Agentic Product-Level Experience-Oriented Game Harness", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/22", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/584" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/22" ], "reason": "审查通过，待核心算法实现与公开验证。Recursive Game Creator包含designer/build、可执行多策略player rollout、盲化版本review及实际artifact retention，保留偏好证据到下一轮，属于固定模型Agent迭代闭环。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08622", "title": "Agentic RCA for Internet-Scale Services Using Constrained Creativity", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/21", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/583" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/21" ], "reason": "审查通过，待核心算法实现与公开验证。E4以固定typed DAG语法约束RCA Agent，依次选playbook、组合已有operators、合成新operators；实际执行与admission形成可复现核心控制机制。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08639", "title": "Forensic Reserve: Eliciting Latent Knowledge for Image Forgery Detection", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/19", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/581" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/19" ], "reason": "审查通过，待核心算法实现与公开验证。RGE用冻结DINOv3的site-wise ICA及real/fake discovery定位reserve、固定输出basis U，只训练零初始化A的残差UA h；classifier/backbone冻结，benchmark不参与discovery。可复用视觉adapter方法，当前证据仍image forensic专门任务。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08647", "title": "SquidAgent: Parallelize Wisely, Coordinate Efficiently", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/18", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/580" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/18" ], "reason": "审查通过，待核心算法实现与公开验证。SquidAgent一次DAG规划同时估算输出token/层对齐成本，session fork继承上下文、预写约定，以sum(tokens)/(max(tokens)+alignment)>alpha决定并行；明确可执行Agent调度机制。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08649", "title": "Stable Scores, Unstable Answers: Frame Phase and Option Order in Video Multiple-Choice Evaluation", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/17", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/579" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/17" ], "reason": "审查通过，待核心算法实现与公开验证。PhaseFusion把3×8帧相位后验映射到内容空间作算术平均，独立于option排序控制；n442与32帧准确率差0.45pp不显著，origin flip率下降但order flip未解决；串行3.18s高于24帧2.86s，不能用attention单位冒充实测提速。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08659", "title": "Selective Transfer of RL Updates for Visual Reasoning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/15", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/577" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/15" ], "reason": "审查通过，待核心算法实现与公开验证。Selective-RL隔离R−B与B−V，逐矩阵rank1 SVD后恢复原Frobenius norm，分别重构donor插值和直接加RLdelta；3个8B VLM骨干，源验证选donor，H1验证选λ后冻结，需保留非视觉模块支持对齐。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08662", "title": "ParanoiaEval: Benchmarking Unnecessary Defensive Work in Agentic Coding", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/14", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/576" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/14" ], "reason": "审查通过，待核心算法实现与公开验证。ParanoiaEval作为evolve行为约束评测接受：200同任务E+/E−证据对，任务oracle和多余风险处置judge分离，可同时评测完成度与证据响应。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08670", "title": "Principled Under Pressure: Post-Training Decides Whether LLMs Act on Their Own Moral Judgment", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/12", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1474" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/12" ], "reason": "审查通过，待核心算法实现与公开验证。公共评测：判断—行动差距的筛选、四级测量校准与压力对照独立于训练算法；Apache-2.0 DeepSteer仓库可访问。仅74/208主情境通过screen，报告条件分布不能外推总体。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08674", "title": "EC-RAG: Event Chain Retrieval-Augmented Generation for Long Video Understanding", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/11", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/574" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/11" ], "reason": "审查通过，待核心算法实现与公开验证。EC-RAG训练外query模态拆解、CLIP variance取帧、ASR/OCR/DET事件描述与按时间相邻链，Contriever相关事件/邻接上下文辅助回答；三长视频benchmark，不是学习因果事件图，额外工具/token预算须计。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08678", "title": "Secure Speculative Decoding for Large Language Models", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/10", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/573" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/10" ], "reason": "审查通过，待核心算法实现与公开验证。SecureSD 按全局生成位置混合标准/宽松接受与拒绝恢复分布，早期严格验证，后续恢复快速验证；定义完整，可作为真实投机解码候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08691", "title": "ScienceClaw: Benchmarking Continual Self-Evolution of AI-for-Science Agents Across the Natural and Social Sciences", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/9", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/572" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/9" ], "reason": "审查通过，待核心算法实现与公开验证。ScienceClaw冻结模型，以typed workflow执行修复提取linked Skill–Operator bundles，boundary/source reset replay必须实际使用新版本，只有独立validation严格提升且hard constraints/预算通过才原子晋级；符合真实evolve闭环候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08713", "title": "SpaTime: Streaming Vision-Language Models for Spatio-temporal Reasoning", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/8", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/571" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/8" ], "reason": "审查通过，待核心算法实现与公开验证。SpaTime冻结因果几何/视觉编码器，2×2 merge投影相加，训练LoRA/state/head并建模首个Response分布作时间损失；strict只计最早回答，charitable末帧强制回答须分开。StreamVSTI训练、StreamVSI任务OOD。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08718", "title": "When Forgetting is not Catastrophic: On the Mechanics of Spurious Forgetting", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/7", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/570" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/7" ], "reason": "审查通过，待核心算法实现与公开验证。新事实微调造成共享表征偏移与逐事实侵蚀分离；对更新矩阵删首奇异分量的干预在OLMo2-1B有同范数随机rank1对照，可登记机制诊断候选，不能宣传为通用抗遗忘能力算法。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08773", "title": "AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/5", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1473" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/5" ], "reason": "审查通过，待核心算法实现与公开验证。冻结world model与judge，分阶段课程—执行者、对手—执行者训练；三seed rollout与150任务对照，不等于真实生产A/B。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08775", "title": "Agent in a Bottle: Can LLM Agents Turn Their Capabilities Into Cheap, Scalable Artifacts?", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/4", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/568" ], "track": "agent", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/4" ], "reason": "审查通过，待核心算法实现与公开验证。BOTTLED作为evolve可执行评测基础设施接受：Agent在预算内生成可复用artifact并完成大批无标签任务，明确隐藏gold与预算约束，兼容模型/脚本/路由器候选。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08778", "title": "Sherpa: Teaching LLMs to Teach Adaptively", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/3", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1472" ], "track": "post-training", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/3" ], "reason": "审查通过，待核心算法实现与公开验证。逐回合gate掩码return并按到达同回合的轨迹leave-one-out归一，训练teacher适应学生偏好；有MATH分离训练/评测及教学benchmark。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08779", "title": "ALIVE: Interaction-Aligned Object Insertion for First-Frame-Guided Video Editing", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/2", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/566" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/2" ], "reason": "审查通过，待核心算法实现与公开验证。ALIVE以source+noisy target通道拼接、干净首帧和按时间overlap/描述长度归一的cross-attention；Qwen VLM从同输入预测P4而不是读target。35,800对训练与两benchmark；部分基线额外P2提示不完全同条件。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.08791", "title": "World Models' Last Exam in Physics", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/0", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/565" ], "track": "foundation-model", "old_computed_status": "implementation-pending", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/38/candidates/0" ], "reason": "审查通过，待核心算法实现与公开验证。40项物理关系测量任务覆盖九类现象，官方公开任务图像/提示/评估器及外部视频批评估入口，可纳入evolve公共评测候选；不是新世界模型算法，也不是已接入。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.18157", "title": "Fan Out or Relay? Ownership-Aware Routing for LLM Agents Querying Siloed Tables", "accepted_evidence": [ "docs/research-audits/oct15-review-closure/paper-register.json#/papers/876" ], "track": "agent", "old_computed_status": "reviewed", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/44/candidates/48" ], "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。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.18684", "title": "Knowing the User Is Not Knowing What We Established: Grounding Awareness in Personal Agents", "accepted_evidence": [ "docs/research-audits/oct15-review-closure/paper-register.json#/papers/3066" ], "track": "agent", "old_computed_status": "reviewed", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/44/candidates/144" ], "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。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.18780", "title": "BLADE: Auditing Information Use for Failure Attribution in LLM Multi-Agent Systems", "accepted_evidence": [ "docs/research-audits/oct15-review-closure/paper-register.json#/papers/811" ], "track": "agent", "old_computed_status": "reviewed", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/44/candidates/44" ], "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。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.19131", "title": "Plant Memory: Technical Language Processing and a Verified Maintenance Ontology for Agent-Accessible Failure Knowledge from SAP PM Work Orders", "accepted_evidence": [ "docs/research-audits/oct15-review-closure/paper-register.json#/papers/770" ], "track": "agent", "old_computed_status": "reviewed", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/44/candidates/41" ], "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。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.19139", "title": "Online Tokenization with Optimal Delay in Linear Time", "accepted_evidence": [ "docs/research-audits/oct15-review-closure/paper-register.json#/papers/257" ], "track": "foundation-model", "old_computed_status": "reviewed", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/42/candidates/207" ], "reason": "在线tokenization以稳定输出前缀与delay形式化为基础，组合special-token、pretokenizer和BPE，并给出更低delay的联合算法与不可解边界；公开Morgan Stanley上游实现可做CPU级精确ID/流式等价性验收。接受为低成本算法候选，本轮不实现、不跑完整吞吐研究。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.19316", "title": "Truthful and Expected-Payment-Optimal Model Routing in Graph-Based Agentic Workflows", "accepted_evidence": [ "docs/research-audits/oct15-review-closure/paper-register.json#/papers/745" ], "track": "agent", "old_computed_status": "reviewed", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/44/candidates/38" ], "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。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.19345", "title": "When Is Smaller Greener? Right-Sizing Language Models for Agentic AI", "accepted_evidence": [ "docs/research-audits/oct15-review-closure/paper-register.json#/papers/3057" ], "track": "agent", "old_computed_status": "reviewed", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/44/candidates/135" ], "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。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.19474", "title": "Beyond Retrieval: Evidence Access and Event Aggregation in Agent Memory", "accepted_evidence": [ "docs/research-audits/oct15-review-closure/paper-register.json#/papers/3043" ], "track": "agent", "old_computed_status": "reviewed", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/44/candidates/121" ], "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。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.19659", "title": "Measuring and Reducing Unusable Replies of LLM Agents", "accepted_evidence": [ "docs/research-audits/oct15-review-closure/paper-register.json#/papers/689" ], "track": "agent", "old_computed_status": "reviewed", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/44/candidates/31" ], "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。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.19801", "title": "Delta Tokenizer: Reference-Identical Tokenization of Session Deltas and Segment-Parallelism for Long Prompts on Multicore CPUs", "accepted_evidence": [ "docs/research-audits/oct15-review-closure/paper-register.json#/papers/670" ], "track": "agent", "old_computed_status": "reviewed", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/44/candidates/28" ], "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。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.19818", "title": "One Monitor Does Not Fit All: Formalism-Agnostic Runtime Verification for Agentic AI", "accepted_evidence": [ "docs/research-audits/oct15-review-closure/paper-register.json#/papers/3036" ], "track": "agent", "old_computed_status": "reviewed", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/44/candidates/114" ], "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。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.20067", "title": "Which Knowledge Should LLMs Trust? Uncovering a Generalizable Steering Direction Across Knowledge Conflicts", "accepted_evidence": [ "docs/research-audits/oct15-review-closure/paper-register.json#/papers/649" ], "track": "agent", "old_computed_status": "reviewed", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/44/candidates/25" ], "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。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.20303", "title": "SpectrumQuant: Spectrum Alignment for LLM Quantization", "accepted_evidence": [ "docs/research-audits/oct15-review-closure/paper-register.json#/papers/57" ], "track": "post-training", "old_computed_status": "reviewed", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/43/candidates/9" ], "reason": "SpectrumQuant 用校准激活协方差构造闭式 alignment+mixing 正交旋转；原文7B校准约0.26 GPU小时、4.51 GiB，可保留128条真实校准样本的7B原算法路径；不承诺70B或全套多模型下游扫评；接受为低成本审查候选，本轮未实现、未实验，亦不自动进入执行/evolve。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false }, { "id": "2610.20413", "title": "When Does Momentum Help in Reinforcement Learning with Verifiable Rewards?", "accepted_evidence": [ "docs/research-audits/oct15-review-closure/paper-register.json#/papers/36" ], "track": "post-training", "old_computed_status": "reviewed", "ledger_evidence": [ "docs/paper-discovery-ledger.json#/batches/43/candidates/6" ], "reason": "只接收论文原始十臂 bandit 的 bias-corrected EMA/安全 η/(1−β) 机制和解析性质检验，可用 CPU 数值运算保留原目标；不接收4B/32B RLVR或50,000-replicate完整统计活动，不据此声称LLM提升；接受为低成本审查候选，本轮未实现、未实验，亦不自动进入执行/evolve。", "catalog_match": false, "source_ledger_cost_gate_closed": false, "user_implementation_authorization_inferred": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null, "this_audit_full_experiment_review_done": false, "component_implementation_authorized_by_this_report": false } ], "accepted_catalogued": [ { "id": "2610.11167", "title": "PIVOT: Perplexity-Informed KD-to-RL Transition Scheduling for Vertical-Domain Few-Shot Distillation", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/311" ], "entries": [ { "key": "pivot", "domain": "post-training", "detail_path": "post-training/2610.11167-pivot/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.11247", "title": "Why On-Policy Distillation Sometimes Fails: Vanishing Learning Signals", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/299" ], "entries": [ { "key": "opd-learning-signals", "domain": "post-training", "detail_path": "post-training/2610.11247-opd-learning-signals/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.11251", "title": "V-CoLA: Vision Token Compression with Linear Attention", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/298" ], "entries": [ { "key": "vcola", "domain": "multimodal-models", "detail_path": "multimodal-models/2610.11251-vcola/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.11291", "title": "When Do We Need On-Policy Distillation? Distilling on Offline Student Rollouts Is Often Better", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/294" ], "entries": [ { "key": "semi-opd", "domain": "post-training", "detail_path": "post-training/2610.11291-semi-opd/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.11332", "title": "ReCal: Calibrating Structured Pruning for On-Policy Distillation Recovery", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/285" ], "entries": [ { "key": "recal", "domain": "post-training", "detail_path": "post-training/2610.11332-recal/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.11519", "title": "Residual Advantage: Student-Relative Teacher Guidance for RL with Verifiable Rewards", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/256" ], "entries": [ { "key": "residual-advantage", "domain": "post-training", "detail_path": "post-training/2610.11519-residual-advantage/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.11659", "title": "DIAL-OPD: Learning More from Fewer Tokens in On-Policy Distillation", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/233" ], "entries": [ { "key": "dial-opd", "domain": "post-training", "detail_path": "post-training/2610.11659-dial-opd/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.11854", "title": "GRPODropout: Less is More for Online Reinforcement Learning Rollouts", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/216" ], "entries": [ { "key": "grpo-dropout", "domain": "post-training", "detail_path": "post-training/2610.11854-grpo-dropout/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.12061", "title": "When Should Agents Think? Adaptive Reasoning via Cross-Turn Estimation", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/196" ], "entries": [ { "key": "race", "domain": "agent-research", "detail_path": "agent-research/2610.12061-race/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.12367", "title": "Which Skill to Distill? SGUID: Selecting a Compact Skill Bank for Model-Skill Co-Evolution", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/169" ], "entries": [ { "key": "sguid", "domain": "post-training", "detail_path": "post-training/2610.12367-sguid/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.12417", "title": "WOVEN: Weaving Visual World Modeling into Multimodal LLMs", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/163" ], "entries": [ { "key": "woven", "domain": "multimodal-models", "detail_path": "reproductions/2610.12417-woven/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.12707", "title": "Unsupervised Multimodal Reward Modeling", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/140" ], "entries": [ { "key": "umrm", "domain": "multimodal-models", "detail_path": "reproductions/2610.12707-umrm/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.13025", "title": "Evo$^2$: Evolving the Evolver through Weight-Harness Co-Evolution", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/112" ], "entries": [ { "key": "evo2", "domain": "agent-research", "detail_path": "reproductions/2610.13025-evo2/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.13092", "title": "When Faster Decisions Do Not Make Better Agents: An Empirical Study of Jev in Agent Workflows", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/99" ], "entries": [ { "key": "jev-workflow", "domain": "agent-research", "detail_path": "reproductions/2610.13092-jev-workflow/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.13194", "title": "Janus: Enhancing Long-Horizon Autonomous Research with Co-Evolving World Models", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/87" ], "entries": [ { "key": "janus", "domain": "agent-research", "detail_path": "reproductions/2610.13194-janus/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.13257", "title": "CAST: Credit Alignment for Cross-Tokenizer On-Policy Distillation", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/80" ], "entries": [ { "key": "cast-opd", "domain": "post-training", "detail_path": "reproductions/2610.13257-cast-opd/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.13285", "title": "BoundaryEvo: Data-Model Co-Evolution at the Capability Boundary for Long-Chain Reasoning Distillation", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/71" ], "entries": [ { "key": "boundary-evo", "domain": "post-training", "detail_path": "reproductions/2610.13285-boundary-evo/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.13348", "title": "Test-Time Memory Construction", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/62" ], "entries": [ { "key": "t2mem", "domain": "agent-research", "detail_path": "reproductions/2610.13348-t2mem/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.13481", "title": "Branching On-Policy Distillation", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/50" ], "entries": [ { "key": "branch-opd", "domain": "post-training", "detail_path": "reproductions/2610.13481-branch-opd/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.13521", "title": "Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/45" ], "entries": [ { "key": "road-mopd", "domain": "post-training", "detail_path": "reproductions/2610.13521-road-mopd/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.13630", "title": "Multi-Stream Attention Residuals", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/36" ], "entries": [ { "key": "msar", "domain": "foundation-models", "detail_path": "reproductions/2610.13630-msar/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.13735", "title": "Orthros: Phase-Aware Heterogeneous Attention for Efficient Transformers", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/23" ], "entries": [ { "key": "orthros", "domain": "foundation-models", "detail_path": "reproductions/2610.13735-orthros/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.13767", "title": "DualQuant: Dual-scale Calibration-free Weight-and-Activation Quantization for LLMs", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/18" ], "entries": [ { "key": "dualquant", "domain": "foundation-models", "detail_path": "reproductions/2610.13767-dualquant/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.13806", "title": "FORTE: A Production Framework for Feature Evaluation and Deprecation in Industrial Recommender Systems", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/13" ], "entries": [ { "key": "forte", "domain": "recommendation", "detail_path": "reproductions/2610.13806-forte/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.13889", "title": "RATED: Reliability-Aware AI Agent Tool Retrieval via Execution Dynamics", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/6" ], "entries": [ { "key": "rated", "domain": "agent-research", "detail_path": "reproductions/2610.13889-rated/README.md", "document_exists": true, "capability_complete_inferred": false } ] }, { "id": "2610.13912", "title": "RSI-Anything: Unlocking Recursive Self-Improvement with World Models", "accepted_evidence": [ "docs/research-audits/oct12-review-closure/paper-register.json#/papers/3" ], "entries": [ { "key": "rsi-anything", "domain": "agent-research", "detail_path": "reproductions/2610.13912-rsi-anything/README.md", "document_exists": true, "capability_complete_inferred": false } ] } ], "accepted_terminal": [ { "id": "2610.02626", "title": "Imagine the Future, Internalize the Gist: Efficient VLA Reasoning via Internalized Spatiotemporal Imagination", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/702", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1166" ], "terminal_evidence": [ "docs/paper-discovery-ledger.json#/batches/45/candidates/5" ] }, { "id": "2610.05417", "title": "Render to Reason: Novel-View Semantic Prediction Improves Spatial Understanding in VLMs", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/372", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/880" ], "terminal_evidence": [ "docs/paper-discovery-ledger.json#/batches/54/candidates/11" ] }, { "id": "2610.07338", "title": "Logbook: Extremely Long-form Audio Event Understanding", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/171", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/723" ], "terminal_evidence": [ "docs/paper-discovery-ledger.json#/batches/54/candidates/12" ] }, { "id": "2610.08077", "title": "Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight", "accepted_evidence": [ "docs/research-audits/oct07-review-closure/paper-register.json#/papers/67", "docs/research-audits/oct12-review-closure/paper-register.json#/papers/1478" ], "terminal_evidence": [ "docs/paper-discovery-ledger.json#/batches/45/candidates/4" ] } ], "deferred_not_accepted": [ { "id": "2610.09027", "title": "Visual Memory Attacks Can Persist Through The KV Cache", "reason": "P-VMI whitebox图像攻击移除后仍经后续KV影响，cache-swap/summarycompaction控制；真KV交互/梯度攻击待实现，onlysoftprompt是diagnostic。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/24", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.09127", "title": "CADFather: Autonomous CAD Reconstruction through Coordinated Tool Use", "reason": "已读CADFather method/experiment：VLM看target/intermediate renders，协调pretrainedCADgeneration/numericalrefinement，真CadQueryDSLexecute feedback与bestvalidarchive，无新training。下一批须真实3D几何/多tool策略，注意testsubsetablation不能参与parameter选择。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/36", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.09159", "title": "SpecGuard: Proving a Task Is Broken Before the Agent Cheats", "reason": "Google DeepMind优先。已读SpecGuard方法、实验和限制：SpecAgent独立测试生成Lean specification，TestAgent分离，semantic connector与kernel证书只有在忠实autoformalization下保证。349 mutated SWE任务和手工60规范审计，不是线上A/B；下一批须防规范弱化/伪冲突。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/41", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": "not_reproduced_high_cost" }, { "id": "2610.09252", "title": "Adaptive Visual Token Reduction for Accelerated Image Understanding", "reason": "ReFIT multi-scale instructionwindowreshape+tokenrefine，只训reductionmodules；四VQA FLOPs非wallclock提升，真实feature模块待实现。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/56", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.09430", "title": "SkillCycle: Co-Evolving Agent Policies and Skill Banks", "reason": "SkillCycle bank/router冻结policy学习后冻结policy修rules，rule/bank环境配对验证；ALFWorld/WebShop真实controller/distill待实现，no-skill部署不可带bank。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/76", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.09484", "title": "MIMESIS: Learning User Simulators as Training Environments for Interactive Agents", "reason": "Meta优先。已读MIMESIS user simulator及agent posttraining：user-side midtrain、ThoughtTrace SFT、多域GRPO后冻结simulator，再agent GRPO+CSD token重加权；private thought仅retrospective teacher训练上下文。9 unseen user评测和8环境，下一批须真实双训练链而非fixed simulator。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/84", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": "not_reproduced_high_cost" }, { "id": "2610.09624", "title": "How Do Agentic LLMs Decide to Call Tools? A Tool-Call Vector Shaped by Suppression", "reason": "已读toolcall causalvector分析/跨model评测：500minimalverbpairs300discover200heldout，hiddenstateintervention/transcoder，7families与native tau2 generalization。下一批须真实activations causalnecessity/sufficiency，不用文字heuristic决定toolcall或按test筛vector。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/96", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.09761", "title": "PARC-Loc: Text-to-Point-Cloud Localization with Partial Assignment and Relational Consistency", "reason": "PARCpartialassignment+pairwise spatial关系和neighborboundary扩充，KITTI360/CityLoc；真实text/pointcloud特征与coarse/fine权重待实现。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/111", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.09856", "title": "Self-Evolve With a Reference:Anchored Training of Tool-Integrated Agents", "reason": "已读AnchorLoop方法/实验：历史Executor frozenanchor同时给cross-referenceadvantage和Curriculum一致性proxy，VeRL sandbox+交替真实训练；curriculum时anchor与executor同参数，不能冒称版本改善或correctness。下一批须保留dualagent训练和reference snapshot，不以selfconsistency当gold。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/124", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.09935", "title": "AgentTracer: Tracing Indirect Prompt Injection Attack through Fine-Grained Intention-Execution Alignment", "reason": "已读AgentTracer graphdesign/evaluation：userintent authorization与operationKB、taskintent implicitdependency、从anomalouscall backwardprune追injectionsource；successattacklogs加背景noise。下一批须完整trace与无attackfalsepositive控制，不能已知injection位置当attributor输入。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/134", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.09952", "title": "MSU Team at the Explainable Deepfake Detection Challenge 2026: Grounded Artifact Evidence for Deepfake Detection", "reason": "XPlainVerse DINO/Mesorchdetector、trainingartifactmask和patchcontrastive+QwenVL解释/GRPO简化；publictrainval/test可验，真多stage待实现。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/137", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.10062", "title": "Loud Failures, Quiet Failures: Fault Detection and Recovery in Tool-Using Language Model Agents", "reason": "已读loud/quietfailure method/results：自有BFCL statefultool layer四typedfault控制注入，24task1920trial、faultfree nondeterminism基线与检测/改plan/recovery分开。下一批须真实tool状态execute，gold finalstate仅scorer，reasoning无recovery提升负结果保留。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/145", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.10071", "title": "HGP:An on-device personalized agent memory via hybrid graph storage", "reason": "已读HGP方法/实验：confidence-driven LLMannotation在线训lightweightrouter、episode/semantic/proceduralgraphs+workingstate；LoCoMo/PAL选择与QA。下一批须真实router更新与图检索、on-device latency/资源实际测，不能只标签分类fixture。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/148", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.10120", "title": "RewardWeaver: Long-Horizon Interactive Learning for Language Agents via Self-Evolving Reward Adaptation", "reason": "已读RewardWeaver method/experiment：admitted rubric语义固定，阶段训练后lowoutcome backwardattribution动态reward allocation，独立controlled expansion；社会/议价/sales。下一批须实际policy训练、validatedrubric稳定，不能追分随意重定义reward或测试goal泄漏。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/153", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.10305", "title": "How Do Transformers Learn to Represent Symmetries?", "reason": "Transformerfiniteaugmentation学习pointcloud群不变性与equivariance，受控architecture/ablation；真实训练/OOD几何证据非fixture待实现。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/176", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.10786", "title": "Plan-and-Patch: Diffusion Language Models for Agentic Planning", "reason": "已读Plan-and-Patch方法/实验：真实diffusion planner parallelunmask、region infill保prefix/suffix，与独立executor失败region选择，NaturalPlan/ALFWorld/TextCraft训练。下一批须dLLM与联合generation/repair训练，AR字符串编辑非定义机制。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/228", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.11008", "title": "Why LLM Agents Favor Their Group: Stakes, Observed Norms, and Reputation", "reason": "已读群体favoritism实验/结果：任意group、costless/stake、scripted history与个体reputation控制，4400模拟societies。属于agent社会行为测量，保留下一批参考；不能把合成群体pointallocation解释为现实组织因果。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/254", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.11123", "title": "When Interfaces Speak: Data-Aware Generative UI Harness for Active Interaction", "reason": "已读GenUI task定义/实验：ToolAgent与GUICoder、DynamicUX实际sandbox交互reward、RewardAuditor meta-rubric，UI-TAU真实publicdatabase模拟任务与4B训练。下一批须真实UI/code执行和用户输入，judge reward hacking校验；rounds问卷不当生产A/B。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/264", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.11136", "title": "The \"10th Juror\": Open-Set Standpoint Screening for Bureaucratic Bias Detection", "reason": "已读MARS-Gov方法/实验：legal retrieval、open-set scout动态juror、conservative routing和rewrite语义verification，DGDB/LOCO评测。下一批须独立evidence与重写含义audit，法律背景检索不可视为法律建议/无偏性保证。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/265", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.11226", "title": "When Lower Reconstruction Loss Hurts: Distributionally Robust Refinement for Low-Bit LLM Quantization", "reason": "DRQ固定quant grid只refine integer codes，worst-case activation distributions纠正calibration reconstruction泛化，六PTQ methods dense/MoE公开ppl/tasks；下一批须真实整数码/量化forward/uncertainty inner problem。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/280", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.11384", "title": "Environmental Feedback Modeling Matters: Rethinking Feedback Treatment in Agentic Hindsight Self-Distillation", "reason": "已读SELF方法/实验：on-policy真实environment反馈建feedback-conditioned hindsight teacher，joint response modeling loss+selfdistillation；tau/AppWorld reward-free agent训练。下一批须真实双objective梯度，privileged未来feedback仅训练teacher不进部署policy。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/311", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.11573", "title": "Memory Type Varies: Empowering LLM Agents for Long-Term Memory with Diverse Strategies", "reason": "已读MemoType/TriMEM方法和评测：learned router识别memory/query类型选择retrieval，Contriever统一backbone、LongMemEval/LoCoMo/PerLTQA。下一批须实际router训练与检索标签隔离，人工分类型heuristic不能称核心复现。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/335", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.11578", "title": "Chronos Enables Code Agents to Reason over Software Evolution", "reason": "已读Chronos方法/设置：历史PR蒸馏experience cards+typed graph、semantic entry+multi-hop、双patch生成steward选择；6backboneSWE对比。下一批须截止目标issue前history防答案PR泄漏，真实patch执行而非检索静态gold。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/336", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.11862", "title": "LEVER: Adaptive Cost-Aware Proof Search Over AND/OR Graphs", "reason": "已读LEVER AND/OR graph与实验：partial proof realized objective加open subgoal预测，programmable cost/length/impurity，Lean kernel保correct；matched model/version/price。下一批须真实Lean搜索、预测前评分与cost含失败，不能gold proof排序代替。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/368", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.12248", "title": "EgoVoice: Proactive Spoken Assistance from Egocentric Multimodal Streams", "reason": "EgoVoice用HoloAssist source-separated/resynth audio和listen-or-speak SFT/DPO训练主动语音助手，170 sessions onset/content/end评测；多模态/语音后训练下一批须实际流式音视频与timing policy。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/399", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.12611", "title": "SoftwareGym: Scaling Computer-Use Agent Training with Diverse Synthesized Software Environments", "reason": "已读SoftwareGym方法/实验：合成真实可执行GUI app、固定v1 GUI测试迭代repair，182app12类120task，Qwen3.5-9B multimodal SFT、out-of-domain OSWorld。下一批须VM实际GUI交互与合成环境自测隔离，不能用静态截图/固定动作fixture。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/441", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.12655", "title": "CLEAT: Learning Value-Guided Clarification for Interactive Agents", "reason": "已读CLEAT摘要、intent alignment及实验：同history ask/proceed paired rollout估residual query value，controller与task policy co-train，新policy outcome更新labels；Qwen3-4B四bench十设置三seed。下一批须训练真实value predictor+policy，不能拿测试latent goal作agent输入。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/444", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.12794", "title": "Meta-Reflective Evolution Effectively Improves Production AI Systems", "reason": "已读CoCoEvolve方法与实验协议：archive parent、history检索、live target probing、自测mutation再external评分；train实例可见，validation仅score、test隔离。production AI目标及公开heldout基准不自动证明生产流量A/B，下一批须进一步核验live authority和真实evaluator，非registry标签。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/454", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.12957", "title": "JoyInside-VLM-Agent: Product-Conditioned Multimodal Policy Learning for Embodied Agents", "reason": "已读JoyInside方法/实验：产品条件multimodal trajectory、SFT、多teacher OPD及DPER token-level privileged evidence路由；部署单student不读teacher reference。私有120k混合数据、领域skill及两机器人case限制公开复现，下一批须保留真实多阶段训练。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/466", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.13144", "title": "Competence-Aware Orchestration for LLM-Based Multi-Agent Systems", "reason": "已读competence orchestration方法/实验：固定probe capability fingerprint加query-conditioned模型，learned role activation与communication graph；6 reasoning/codebench及unseen backbones。下一批须训练真实competence predictor，probe/test隔离而非gold推断能力。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/495", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null }, { "id": "2610.13166", "title": "Scaling Agentic Data for Long-Horizon Terminal Intelligence", "reason": "已读TerminalHorizon方法/实验：独立execution发现合成环境预先完成的关键工作，再reconstruct并扩展同goal任务；1.5k环境、Qwen3.5-27B SFT与12bench。production settings是任务评测不证明线上A/B；下一批须隔离environment gold verifier与agent可见包。", "evidence": "docs/paper-discovery-ledger.json#/batches/40/candidates/498", "structured_acceptance_found": false, "automatic_implementation_queue": false, "this_audit_full_experiment_resolution": null } ], "current_2026_google_review_leads": [ "2601.00372", "2601.19066", "2601.21149", "2602.02164", "2602.02690", "2602.06276", "2602.07266", "2602.07918", "2602.12481", "2602.16938", "2602.17554", "2603.01293", "2603.06981", "2603.10465", "2603.14609", "2603.14769", "2603.24591", "2603.28896", "2603.29470", "2603.29791", "2604.05364", "2604.09803", "2604.20329", "2604.23099", "2604.27723", "2605.26340", "2605.29729", "2605.30174", "2605.30322", "2606.00131", "2606.00369", "2607.02515", "2607.13614", "2607.25537", "2608.05578", "2608.11367", "2609.01588", "2609.01834", "2609.12464" ], "pre_2026_google_reference_leads": [ "2405.02699", "2507.23590", "2509.22963", "2510.03217", "2510.17670", "2511.02767", "2511.03782", "2511.10865", "2511.11552" ], "non_implementation_google_source_leads": [ "2605.03767", "2606.03237", "2606.12683", "2607.03906", "2607.20268", "2608.25924" ], "this_audit_full_experiment_decisions": [ { "id": "2610.07289", "resolution": "source_unavailable_archived", "full_experiment": "私有生产环境不可公开复现，审查 DONE", "basis": "Google private source, internal CI/model and production changes; Data Availability says no public reproduction package/data", "source": "https://arxiv.org/html/2610.07289v1", "location": "Data Availability; section 3", "component": "公开APR核心另行界定范围与成本；尚未实现，不声称生产复现" }, { "id": "2610.06616", "resolution": "not_reproduced_high_cost", "full_experiment": "成本过高，不复现（审查 DONE）", "basis": "Appendix C.3: learned refiner uses 65000 examples from the Video-ChatGPT training split, two epochs, 8 A100 for 12 hours (96 accelerator-hours), plus backbones and benchmarks; author-reported configuration, not a local measurement", "source": "https://arxiv.org/html/2610.06616v1", "location": "Appendix C.1 and C.3", "component": "training-free selection-only 是独立的优先低成本候选，未实现；不得将其冒充refiner训练" }, { "id": "2610.00497", "resolution": "not_reproduced_high_cost", "full_experiment": "成本过高，不复现（审查 DONE）", "basis": "Appendix C: batch 512, teacher and students up to one million steps, 167.8M main student, multiple datasets and conditions; full campaign exceeds a short single-device budget by scale-based estimate, no local timing claim", "source": "https://arxiv.org/html/2610.00497v1", "location": "Appendix C", "component": "小词表真实teacher/student核心损失是可另行成本核定的候选，未实现" }, { "id": "2602.16490", "resolution": "not_reproduced_high_cost", "full_experiment": "成本过高，不复现（审查 DONE）", "basis": "Section 4.1: 360M/1.7B SmolLM trained on roughly 200B/400B tokens with 22 benchmarks and recipe controls; scale-based assessment, no local timing claim", "source": "https://arxiv.org/html/2602.16490v1", "location": "Section 4.1", "component": "真实depth-growing和inference-looping小模型训练可另行核定；未实现" }, { "id": "2510.15831", "resolution": "not_reproduced_high_cost", "full_experiment": "成本过高，不复现（审查 DONE）", "basis": "Section 4: 100 public and 161 private prompts, Gemini/Veo; section 4.3 reports about 0.7M text tokens and 28 videos per iteration, video tokens excluded; full multi-round/control campaign expensive and partly private", "source": "https://arxiv.org/html/2510.15831v1", "location": "Sections 4 and 4.3", "component": "纯编排诊断不构成定义性视频重生成机制；不自动进入实现队列" }, { "id": "2610.09159", "resolution": "not_reproduced_high_cost", "full_experiment": "成本过高，不复现（审查 DONE）", "basis": "Full three-agent campaign: 349 SWE tasks plus 187 VERINA and 95 LiveCodeBench, repeated model/control runs, 120 bash turns and ten Lean-feedback rounds, one-hour task cap. Paper average $0.26/task implies about $90.74 for 349 tasks one arm, excluding repeats/additional suites; an estimate, not local spending", "source": "https://arxiv.org/html/2610.09159v1", "location": "Section 5", "component": "少量任务的真实D+C到Lean规范、独立测试与kernel链路为优先低成本候选，未实现" }, { "id": "2610.06064", "resolution": "not_reproduced_high_cost", "full_experiment": "成本过高，不复现（审查 DONE）", "basis": "Six 7B-70B models, ten-epoch steering training, 120B judge on H100, multi-task agent/capability evaluation and repetitions; full configuration exceeds the available single-device budget", "source": "https://arxiv.org/html/2610.06064v1", "location": "Section 4; Appendices C and D.2-D.4", "component": "小型真实LM的steering目标与应用范围可另行核定；未实现，不能用固定trust标签替代" }, { "id": "2610.04721", "resolution": "not_reproduced_high_cost", "full_experiment": "成本过高，不复现（审查 DONE）", "basis": "Appendix E.1/Table 10: four A100s, stages of 1125/5960/5801 steps for 8B/9B models; no original-scale run on the available single accelerator", "source": "https://arxiv.org/html/2610.04721v1", "location": "Appendix E.1/Table 10", "component": "纯图结果导入/合并接口不能冒充训练和真实视觉拓扑提取；未接受新的接口实施任务" }, { "id": "2610.09484", "resolution": "not_reproduced_high_cost", "full_experiment": "成本过高，不复现（审查 DONE）", "basis": "Appendix A.1/Table 4: mid-training on 32 A100s for about 4.8 days, SFT on eight A100s; simulator and agent RL use sixteen A100s with 512-rollout configuration", "source": "https://arxiv.org/html/2610.09484v1", "location": "Appendices A.1/A.3 and D.1/Table 13", "component": "固定simulator不能代替双训练链；没有确认可忠实接单的低成本实现" } ], "limits": [ "Accepted is a paper-review decision, not a human request to implement every candidate.", "No catalog match is an unfulfilled component/delivery observation, not proof of affordable faithful reproduction.", "Catalog membership and a diagnostic do not prove full capability reproduction.", "Priority-institution author blocks were not independently available for every old accepted candidate.", "New primary-source cost recommendations are separate from already committed ledger terminal decisions.", "Source unclosed observations by track refer to 504 catalog/delivery observations; full-experiment closure does not deliver the corresponding affordable component." ] }