{ "schema_version": 1, "batch": "20261021-google16-affordable-components", "accepted_ids": [ "2601.19066", "2601.21149", "2602.06276", "2602.07918", "2602.12481", "2602.16938", "2602.17554", "2603.01293", "2603.14769", "2603.29791", "2604.05364", "2604.23099", "2604.27723", "2605.26340", "2608.05578", "2609.12464" ], "component_count": 16, "gpu_validated_component_count": 5, "implementation_status_counts": { "component_only": 12, "evaluation_protocol_interface": 4 }, "full_experiment_resolution_counts": { "not_reproduced_high_cost": 14, "not_reproduced_resource_unavailable": 1, "not_applicable_theoretical_paper": 1 }, "all_components_are_diagnostic": true, "capability_reproductions_claimed": 0, "review_watermark_advanced": false, "all_sources_exhaustive": false, "records": [ { "id": "2601.19066", "title": "Dynamic Cogeneration of Bug Reproduction Test in Agentic Program Repair", "track": "agent", "key": "dynamic-brt", "component_key": "dynamic-brt", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "component_only", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l1_mechanism", "reason": "保留同轨迹fix+BRT的真实子进程fail-to-pass验证、回归过滤及按fix归组的ranked/dual test-aware选择。仅支持有界Python算术子集；没有生成Agent、私有Google基准或LLM语义审查实测，完整实验成本过高不复现。", "doc": "agent-research/2601.19066-dynamic-brt/README.md", "source_url": "https://arxiv.org/html/2601.19066v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "120个Google私有bug、六语言、5配置，每bug20条trajectory、每条25步，Gemini2.5 Pro；内部部署与人工代码审查被描述，但对照是离线重放，不是量化生产A/B。", "decision_zh": "完整5×120×20=12,000轨迹、最多300,000 agent步及私有monorepo测试，明确超出单卡/Mac短验预算；估算基于步数上界，不是实测。小规模真实patch/BRT选择核心可负担，未获本轮实施授权。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_reproduced_high_cost" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2601.19066v1", "read_sections": [ "§2.1 Our Agentic APR System", "§3 Dynamic Cogeneration", "§4.2配置", "§5结果摘要" ], "source_locations": [ "§2.1 Our Agentic APR System", "§3 Dynamic Cogeneration", "§4.2配置", "§5结果摘要" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "agent-research/2601.19066-dynamic-brt/README.md" }, "component_evidence": "docs/agent-research/2601.19066-dynamic-brt/metrics/component-contract.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": false }, { "id": "2601.21149", "title": "Mobility-Embedded POIs: Learning What A Place Is and How It Is Used from Human Movement", "track": "foundation-model", "key": "me-pois", "component_key": "me-pois", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "component_only", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_reproduced_resource_unavailable", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l1_mechanism", "reason": "实现ME-POIs Eq9–17的不重复原型InfoNCE、多尺度Gaussian时段分布迁移、锚点/稀疏KL及冻结文本cosine联合目标。仅损失/梯度组件，不以生成张量冒充访问Transformer或Maps能力；原Veraset/SafeGraph/Maps数据不可用，完整实验不复现（审查DONE）。", "doc": "foundation-models/2601.21149-me-pois/README.md", "source_url": "https://arxiv.org/html/2601.21149v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "Veraset LA一年/Houston20天私有移动数据；53.7M参数、4层8头d512；20 epoch预训练、100 epoch微调早停，V100 16GB；5项地图属性probe。", "decision_zh": "模型单卡可放下；原数据为付费/私有移动轨迹与Safegraph，原尺度训练与多个embedding API成本未量化，不能承诺完整复现。可负担有界机制候选，未实施；不把V100使用本身视为高成本证明。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_reproduced_resource_unavailable" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2601.21149v1", "read_sections": [ "§3.3-3.6", "§4.1", "Appendix A.1.4-A.3" ], "source_locations": [ "§3.3-3.6", "§4.1", "Appendix A.1.4-A.3" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "foundation-models/2601.21149-me-pois/README.md" }, "component_evidence": "docs/foundation-models/2601.21149-me-pois/metrics/component-a100-seed42.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/me-pois-a100-20261021.json" }, { "id": "2602.06276", "title": "Statistical Learning from Attribution Sets", "track": "recommendation", "key": "attribution-sets", "component_key": "attribution-sets", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "component_only", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l1_mechanism", "reason": "逐式实现归因集合风险估计、位置先验平方加权与固定分母ERM；有限样本常数函数校验发现v1印刷式不一致并保留反例，未宣称无偏性或推荐效果已验证。完整约7800次训练网格成本过高，不复现（审查DONE）。", "doc": "recommendation-components/2602.06276-attribution-sets/README.md", "source_url": "https://arxiv.org/html/2602.06276v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "MNIST、CIFAR10、Higgs；MLP/CNN；k=2^0…2^8；10学习率×10重复，MNIST200epoch、其余100epoch。正文是初步离线模拟归因集合实验。", "decision_zh": "完整网格约(9+9+8)×3×10×10=7,800次训练，100–200epoch，远超单卡短运行；核心loss小数据CPU机制可负担，未授权本轮实现。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_reproduced_high_cost" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2602.06276v1", "read_sections": [ "§3-4章节导航", "§5 Experiments" ], "source_locations": [ "§3-4章节导航", "§5 Experiments" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "recommendation-components/2602.06276-attribution-sets/README.md" }, "component_evidence": "docs/recommendation-components/2602.06276-attribution-sets/metrics/component-cpu-seed42.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": false }, { "id": "2602.07918", "title": "CausalArmor: Efficient Indirect Prompt Injection Guardrails via Causal Attribution", "track": "agent", "key": "causal-armor", "component_key": "causal-armor", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "component_only", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l1_mechanism", "reason": "真实计算用户/不可信tool段的批量LOO动作log-likelihood，按动作token长度归一化并定位支配跨度。保留选择性sanitizer和首个命中后assistant遮蔽、重新提案接口；缺模型回调时扣留动作，诊断实跑不声称安全能力。", "doc": "agent-research/2602.07918-causal-armor/README.md", "source_url": "https://arxiv.org/html/2602.07918v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "AgentDojo1.2.2四环境/三攻击template与DoomArena TauBench115场景×5seed；VertexAI多backbone/sanitizer，原环境8×A100。", "decision_zh": "原文8×A100与多模型/攻击/seed完整矩阵成本过高；不尝试下载或启动。LOO/masking有界组件可单卡/CPU模型接入，未在本轮批准实施。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_reproduced_high_cost" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2602.07918v1", "read_sections": [ "§3-4机制导航", "Algorithm 1", "Appendix C/D.1" ], "source_locations": [ "§3-4机制导航", "Algorithm 1", "Appendix C/D.1" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "agent-research/2602.07918-causal-armor/README.md" }, "component_evidence": "docs/agent-research/2602.07918-causal-armor/metrics/component-contract.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/causal-armor-a100-20261021.json" }, { "id": "2602.12481", "title": "Online Advertising with Spatial Interactions", "track": "industrial", "key": "spatial-ads", "component_key": "spatial-ads", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "component_only", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_applicable_theoretical_paper", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l1_mechanism", "reason": "实现分组伪圆盘真实LP松弛、x/3独立采样和原式二次稀释，核验组互斥及中心不被覆盖。不包含全广告匹配约化、支付或PTAS，不冒称18近似端到端复现；理论论文没有完整训练实验，不伪造高成本理由。", "doc": "recommendation-components/2602.12481-spatial-ads/README.md", "source_url": "https://arxiv.org/html/2602.12481v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "理论算法与证明；正文目录至hardness/附录，无实验章，全文检索experiment无结果；不是生产广告实验。", "decision_zh": "小规模LP/几何实例CPU可负担；无必须大模型/数据；未在本轮批准实现，不用缺A/B否认理论算法。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_applicable_theoretical_paper" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2602.12481v1", "read_sections": [ "§2.1-2.3", "Algorithm简单slot选择", "全文章节/experiment检索" ], "source_locations": [ "§2.1-2.3", "Algorithm简单slot选择", "全文章节/experiment检索" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "recommendation-components/2602.12481-spatial-ads/README.md" }, "component_evidence": "docs/recommendation-components/2602.12481-spatial-ads/metrics/component-cpu-seed42.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": false }, { "id": "2602.16938", "title": "ConvApparel: A Benchmark Dataset and Validation Framework for User Simulators in Conversational Recommenders", "track": "agent", "key": "convapparel", "component_key": "convapparel", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "evaluation_protocol_interface", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l0_interface", "reason": "实现用户模拟器PLSA的KS/MWU统计、HLS概率评估与跨策略来源/身份隔离。附CPU unigram/bigram logistic诊断判别器；不是论文Gemini微调模型，fixture高分不代表真人保真度或推荐能力。", "doc": "agent-research/2602.16938-convapparel/README.md", "source_url": "https://arxiv.org/html/2602.16938v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "4,146真人对话＋3,549模拟对话训练Gemini2.5FlashLite discriminator；各simulator生成300好＋300坏对话；90年代Turing类比不是生产证据；KS/MWU检验及HLS显示realism gap。", "decision_zh": "有界统计/validation框架可负担；完整真人数据采集、多个GeminiSFT与数千多轮rollout不在单卡短预算，原完整实验不复现。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_reproduced_high_cost" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2602.16938v1", "read_sections": [ "§4-6", "§6.1 discriminator配置", "§6.2结果" ], "source_locations": [ "§4-6", "§6.1 discriminator配置", "§6.2结果" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "agent-research/2602.16938-convapparel/README.md" }, "component_evidence": "docs/agent-research/2602.16938-convapparel/metrics/protocol-cpu-seed42.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": false }, { "id": "2602.17554", "title": "A Theoretical Framework for Modular Learning of Robust Generative Models", "track": "foundation", "key": "modular-robust-generative", "component_key": "modular-robust-generative", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "component_only", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l1_mechanism", "reason": "实现有限支持的三方primal-dual鲁棒gate、EG域对抗、全局质量约束、SIR与因果router缓存目标NLL；明确逐输入softmax不等于全局归一化，短诊断不宣称约束收敛或LLM鲁棒提升。原多专家/基线/多数据多seed完整campaign成本过高，不复现（审查DONE）。", "doc": "foundation-models/2602.17554-modular-robust-generative/README.md", "source_url": "https://arxiv.org/html/2602.17554v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "synthetic vocab100/seq10、1层2头ff32、embedding6/8/16/20；expert/baseline/gate800步，5重复；结构蒸馏5K序列。已读部分可明确CPU小机制可行。", "decision_zh": "正文小型Transformer实验可在Mac/单卡负担；不因为标题涉及LLM而判高成本；本轮未批准实施。真实数据§8.4-8.5尚未逐段审读，不据此声称完整实验复现可负担。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_reproduced_high_cost" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2602.17554v1", "read_sections": [ "§7.3 Algorithm5", "§8.1-8.3" ], "source_locations": [ "§7.3 Algorithm5", "§8.1-8.3" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "foundation-models/2602.17554-modular-robust-generative/README.md" }, "component_evidence": "docs/foundation-models/2602.17554-modular-robust-generative/metrics/component-a100-seed42.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/modular-robust-generative-a100-20261021.json" }, { "id": "2603.01293", "title": "Theoretical Perspectives on Data Quality and Synergistic Effects in Pre- and Post-Training Reasoning Models", "track": "foundation/post-training", "key": "data-quality-lsa", "component_key": "data-quality-lsa", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "component_only", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l1_mechanism", "reason": "实现Eq3.1–4.1的LSA分块prompt、population协方差初始化、自主多步rollout、教师SFT、最终OS回归目标及保留初始零空间的SFT闭式解。OS明确不是policy-gradient RL；仅数学/梯度诊断，GPT2多网格十重复长训练成本过高，不复现（审查DONE）。", "doc": "foundation-models/2603.01293-data-quality-lsa/README.md", "source_url": "https://arxiv.org/html/2603.01293v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "LSA d100/130epochs；GPT2架构9.5M、12层8头d256、batch64，SFT100步/OS12K步，10重复；synthetic covariance shift。", "decision_zh": "LSA核心CPU可负担；完整GPT2×多B/n/k×10run×12K OS步预计长训练，按网格规模推断超短预算，不启动；未在本轮批准LSA组件。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_reproduced_high_cost" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2603.01293v1", "read_sections": [ "§1任务/OS边界", "§8 Experiments" ], "source_locations": [ "§1任务/OS边界", "§8 Experiments" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "foundation-models/2603.01293-data-quality-lsa/README.md" }, "component_evidence": "docs/foundation-models/2603.01293-data-quality-lsa/metrics/component-a100-seed42.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/data-quality-lsa-a100-20261021.json" }, { "id": "2603.14769", "title": "POLCA: Stochastic Generative Optimization with LLM", "track": "agent", "key": "polca", "component_key": "polca", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "component_only", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l0_interface", "reason": "实现经验均值优先队列、完整评价历史和噪声方差、同一有放回minibatch比较及farthest-first epsilon-net过滤，显式接入Summarizer/Proposer。callback单测仅证明编排，不冒称已跑LLM程序优化或接入Evolve。", "doc": "agent-research/2603.14769-polca/README.md", "source_url": "https://arxiv.org/html/2603.14769v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "tau-bench/HotpotQA/VeriBench/KernelBench；Veri41题每题最多50 compiler calls×3run×多baseline；Kernel16题、Claude3.7与Gemini embedding。", "decision_zh": "完整多benchmark多baseline×3/6seed、每题50迭代与付费LLM/API的搜索矩阵成本过高；epsilon-Net/队列有界core可负担，未批准本轮实现。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_reproduced_high_cost" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2603.14769v1", "read_sections": [ "§1机制", "Appendix D VeriBench/KernelBench配置与预算" ], "source_locations": [ "§1机制", "Appendix D VeriBench/KernelBench配置与预算" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "agent-research/2603.14769-polca/README.md" }, "component_evidence": "docs/agent-research/2603.14769-polca/metrics/component-contract.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": false }, { "id": "2603.29791", "title": "Reasoning-Driven Synthetic Data Generation and Evaluation", "track": "foundation/agent", "key": "simula", "component_key": "simula", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "evaluation_protocol_interface", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l0_interface", "reason": "实现有界BFS分类树、best-of-N提案/critic、加权采样、meta-prompt复杂化、独立正反双critic拒绝与批内复杂度/Elo校准。语义步骤由显式模型回调提供，本轮仅fixture调度检验，不宣称已生成高质量真实数据。", "doc": "agent-research/2603.29791-simula/README.md", "source_url": "https://arxiv.org/html/2603.29791v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "Gemini2.5Flash；CTI/法律/GSM8K/GlobalMMLU多语言，每数据集512K唯一点、5系统版本；13gram/Jaccard decontamination；BS=N5评分。", "decision_zh": "每任务512K样本×多版本＋多次critic/复杂度与下游训练，付费多模态API数百万调用量级风险；完整实验成本过高。小有界taxonomy/critic核心可负担，未在本轮批准实现。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_reproduced_high_cost" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2603.29791v1", "read_sections": [ "§3 Experimental Setup", "§3.2数据规模", "Appendix E成本示例" ], "source_locations": [ "§3 Experimental Setup", "§3.2数据规模", "Appendix E成本示例" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "agent-research/2603.29791-simula/README.md" }, "component_evidence": "docs/agent-research/2603.29791-simula/metrics/protocol-cpu-seed42.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": false }, { "id": "2604.05364", "title": "TFRBench: A Reasoning Benchmark for Evaluating Forecasting Systems", "track": "multimodal", "key": "tfrbench", "component_key": "tfrbench", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "evaluation_protocol_interface", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l0_interface", "reason": "实现MASE、最多三轮五角色参考构建、S≥4且MASE<1准入与summary重新预测验收；独立one-shot留出评测拒绝晚于cutoff的新闻。标签反馈构建明确标作outcome-conditioned，不能冒充无泄露预测能力。", "doc": "agent-research/2604.05364-tfrbench/README.md", "source_url": "https://arxiv.org/html/2604.05364v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "10数据集/5领域，多Gemini/Claude预测和LLMjudge对比；候选生成验证访问未来ground truth，forecast本身不读；领域外部事件检索包含未来horizon。", "decision_zh": "完整10数据集×多backbone×多候选/3loop＋检索/judge的API矩阵昂贵；小规模benchmark协议与评分接口可负担，未实施。标签隔离必须明确。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_reproduced_high_cost" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2604.05364v1", "read_sections": [ "§3 Algorithm1/6steps", "§4结果表片段" ], "source_locations": [ "§3 Algorithm1/6steps", "§4结果表片段" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "agent-research/2604.05364-tfrbench/README.md" }, "component_evidence": "docs/agent-research/2604.05364-tfrbench/metrics/protocol-cpu-seed42.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": false }, { "id": "2604.23099", "title": "ProEval: Proactive Failure Discovery and Efficient Performance Estimation for Generative AI Evaluation", "track": "cross-domain", "key": "proeval", "component_key": "proeval", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "component_only", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l1_mechanism", "reason": "以排除目标模型的历史评分构造经验GP，实现有限题池BQ后验、解析方差减少和失败超水平集采样。只验证数学机制；不含encoder预训练、GMM与LLM失败合成，完整16模型/API矩阵成本过高不复现。", "doc": "foundation-models/2604.23099-proeval/README.md", "source_url": "https://arxiv.org/html/2604.23099v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "16LLM/VLM×reasoning/knowledge/safety/text/image多benchmark；failure discovery100迭代；GP使用历史外部模型score，目标test gold不能给evaluated agent。", "decision_zh": "现有历史score上的GP/BQ有界core可CPU负担；完整16模型多benchmark与LLM query synthesis昂贵，不本轮重跑；未批准实现。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_reproduced_high_cost" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2604.23099v1", "read_sections": [ "§2 Algorithm2", "§3.1数据/模型/feature", "Table5" ], "source_locations": [ "§2 Algorithm2", "§3.1数据/模型/feature", "Table5" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "foundation-models/2604.23099-proeval/README.md" }, "component_evidence": "docs/foundation-models/2604.23099-proeval/metrics/component-cpu-seed42.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": false }, { "id": "2604.27723", "title": "Optimized Deferral for Imbalanced Settings", "track": "foundation", "key": "mild", "component_key": "mild", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "component_only", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l1_mechanism", "reason": "实现论文不平衡margin的cost-sensitive logistic目标、解析梯度和有界线性router，推理仅接收特征、不读取真实标签或成本。当前仅数学/训练接口诊断；200epoch多数据多专家及LLM路由完整矩阵成本过高不复现。", "doc": "foundation-models/2604.27723-mild/README.md", "source_url": "https://arxiv.org/html/2604.27723v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "CIFAR人为synthetic expert覆盖比例7/2/1等；真实LLM routing有Qwen2.5 0.5B/1.5B等；实验区别synthetic oracle专家与真实模型。", "decision_zh": "loss/router有界CPU机制可负担；不必下载多个LLM就可审查，但不得把synthetic gold expert诊断当真实routing能力；本轮未批准实施。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_reproduced_high_cost" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2604.27723v1", "read_sections": [ "§5.3 Deferral loss", "Appendix C.1/C.2" ], "source_locations": [ "§5.3 Deferral loss", "Appendix C.1/C.2" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "foundation-models/2604.27723-mild/README.md" }, "component_evidence": "docs/foundation-models/2604.27723-mild/metrics/component-cpu-seed42.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": false }, { "id": "2605.26340", "title": "ScientistOne: Towards Human-Level Autonomous Research via Chain-of-Evidence", "track": "agent", "key": "scientist-one", "component_key": "scientist-one", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "evaluation_protocol_interface", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l0_interface", "reason": "实现CoE文件哈希/DAG、数值日志匹配CPR、五次canonical evaluator复算与自适应误差门槛、独立审计多数投票。缺I2/I3/I4语义证据时保留unverified，不用存在链接或哈希吻合冒充引文和方法正确。", "doc": "agent-research/2605.26340-scientist-one/README.md", "source_url": "https://arxiv.org/html/2605.26340v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "ADRS五系统研究任务×五research系统×3run，Gemini3.1Pro，最多20 solver迭代与2小时code窗口；evaluator5次复算、judges多数投票。", "decision_zh": "5×5×3=75完整research runs，单run上限2小时且可能重试3次；多轮500篇pool/100notes的外部研究；完整实验明显高成本。有界provenance/审计接口可负担但本轮未授权实现。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_reproduced_high_cost" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2605.26340v1", "read_sections": [ "§5 Integrity Audit", "§6 Experiments", "Appendix paper investigation" ], "source_locations": [ "§5 Integrity Audit", "§6 Experiments", "Appendix paper investigation" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "agent-research/2605.26340-scientist-one/README.md" }, "component_evidence": "docs/agent-research/2605.26340-scientist-one/metrics/protocol-cpu-seed42.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": false }, { "id": "2608.05578", "title": "Detecting Safety Training Modification in Language Models via Activation Analysis", "track": "foundation-model", "key": "activation-model-scanner", "component_key": "activation-model-scanner", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "component_only", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l1_mechanism", "reason": "实现官方48对概念输入的最后token激活提取、40–80%层扫描、均值差方向、pooled分离与两层结构检查；只运行已缓存小checkpoint，不生成回答，不把reference band当安全认证。14模型多权重族及行为生成完整campaign成本过高，不复现（审查DONE）。", "doc": "foundation-models/2608.05578-activation-model-scanner/README.md", "source_url": "https://arxiv.org/html/2608.05578v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "14model configs/4families；LOO threshold71%(10/14)、20JailbreakBench prompts/model；A10040GB扫描10–40秒但不含模型加载；bootstrap区间宽且部分跨PASS阈值。", "decision_zh": "单已缓存3B/7B模型扫描本身低成本，无需训练；14config权重总下载较大，完整矩阵不本轮下载。可负担core候选not_implemented；不能据71%判生产安全扫描已validated。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_reproduced_high_cost" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2608.05578v1", "read_sections": [ "§IV-F algorithm", "Resource requirements", "Data availability/摘要统计" ], "source_locations": [ "§IV-F algorithm", "Resource requirements", "Data availability/摘要统计" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "foundation-models/2608.05578-activation-model-scanner/README.md" }, "component_evidence": "docs/foundation-models/2608.05578-activation-model-scanner/metrics/component-a100-seed42.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": true, "gpu_validation_artifact": "docs/gpu-validations/activation-model-scanner-a100-20261021.json" }, { "id": "2609.12464", "title": "Beyond Vector Similarity: Hierarchical Context-Aware Graph RAG vs Standard RAG in Enterprise Code Migration", "track": "agent", "key": "hcrg", "component_key": "hcrg", "status": "implemented", "priority": "P1", "review_decision": "component_implemented", "implementation_status": "component_only", "resolution": "component_implemented", "review_done": true, "full_experiment_resolution": "not_reproduced_high_cost", "automatic_retry": false, "automatic_implementation_queue": false, "evolve_integrated": false, "promotion_eligible": false, "diagnostic_only": true, "evaluation_tier": "l0_interface", "reason": "提供明确限于Python AST的依赖图、架构骨架、三跳检索、SCC父依赖优先队列及部分静态指标。未实现论文Java tree-sitter/Spanner/Vertex和真实LLM迁移，不以Python图单测宣称跨语言效果。", "doc": "agent-research/2609.12464-hcrg/README.md", "source_url": "https://arxiv.org/html/2609.12464v1", "cost_assessment": { "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption", "paper_configuration_and_scale": "spring-petclinic Java→Python单repo；Gemini2.5Pro T0、Spanner/Vertex1h context cache；DRQ34.8→65.9%、PCC26.7→45.5%，但cyclomatic一致性下降。无多seed/多repo泛化证据。", "decision_zh": "小public repo的AST/拓扑检索与静态指标core可负担，付费云服务可替为同语义本地graph但须实际LLM迁移；未获本轮实施授权，未测总代价。", "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate", "local_cost_measurement_performed": false, "training_or_model_download_performed": false, "component_decision": "explicit user-authorized bounded component/protocol delivered", "full_experiment_decision": "not_reproduced_high_cost" }, "full_text_review": { "scope": "full-text", "source_url": "https://arxiv.org/html/2609.12464v1", "read_sections": [ "§4.2实验参数", "§5.2指标", "§6结果" ], "source_locations": [ "§4.2实验参数", "§5.2指标", "§6结果" ], "full_article_read": false, "proofs_read_in_full": false, "all_tables_read": false, "latest_revision_reviewed": false, "version": "v1", "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。", "supplementary_method_review": "agent-research/2609.12464-hcrg/README.md" }, "component_evidence": "docs/agent-research/2609.12464-hcrg/metrics/component-contract.json", "review_artifact": "research-audits/google16-components-20261021/README.md", "requires_gpu_validation": false } ] }