{ "schema_version": 1, "batch": "2026-11-04-priority-components", "date": "2026-11-04", "selected_scope_complete": true, "delivery_status": "bounded_scope_complete", "all_sources_exhaustive": false, "review_watermark_advanced": false, "review_watermark": { "before": "2026-10-15", "after": "2026-10-15" }, "fixed_component_keys": [ "agentheal", "gavagai", "proland", "rlid", "specopd", "strategist", "wesearch" ], "reference_only_keys": [ "xla-trenches" ], "counts": { "selected_papers": 8, "component_implementations": 7, "gpu_validated_components": 4, "cpu_only_components": 3, "reference_only_high_cost_review_done": 1, "new_first_publications_since_previous_capture": 7, "recent_overlap_backfills": 1, "capability_reproductions_claimed": 0 }, "papers": [ { "domain": "recommendation", "key": "wesearch", "title": "WeSearch: A Unified Model-Centric Retrieval Framework for WeChat Search", "paper_url": "https://arxiv.org/abs/2611.00530v1", "detail_path": "recommendation-components/2611.00530-wesearch/README.md", "priority": "P0", "first_author": "Kepu Zhang", "first_author_affiliation": "Tencent", "published": "2026-10-31", "published_at": "2026-10-31T01:35:20Z", "code": null, "code_status": "固定v1全文、arXiv资源与作者链接未找到WeSearch原作者代码；引用的OPQ/检索框架不等于本文实现", "cost_assessment": { "source_url": "https://arxiv.org/html/2611.00530v1", "source_location": "Sections3–6,especially6.1.1/6.5", "available_resource": "one NVIDIA A100 or A30, or local Mac", "model_scale": "private pretrained query encoder,feature-awareFFN,learnedOPQ and2-levelRQ-KMeans", "data_scale": "hundreds of millions to billions of private documents plus tens of millions fresh documents,private behavior/relevance teachers", "paper_configuration": "rolling training,atomic model/index refresh,distributed online serving and month-long paid-traffic A/B", "estimated_cost_basis": "exact GPU-hours not disclosed; corpus,private models and production deployment are unavailable in the single-device budget; not a local measurement", "decision": "完整实验成本过高，不复现（审查 DONE）；只执行CPU候选并集、固定OPQ量化/ADC恒等式及不可变新快照，不模拟完整检索器或线上收益。" }, "implementation_status": "component_only", "requires_gpu_validation": false, "gpu_validation_passed": false, "gpu_validation_artifact": null, "id": "2611.00530", "status": "implemented", "resolution": "bounded_l1_component_accepted", "doc": "recommendation-components/2611.00530-wesearch/README.md", "catalog_key": "wesearch", "catalog_detail": "recommendation-components/2611.00530-wesearch/README.md", "track": "recommendation", "intake_class": "new_first_publication_since_previous_capture", "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "component_implemented": true, "component_accepted": true, "diagnostic_only": true, "promotion_eligible": false, "evolve_integrated": false, "automatic_retry": false, "automatic_implementation_queue": false, "verified_priority_organizations": [], "full_text_review": { "scope": "defining method,evidence,resources,authorship and source-release review; not every reference or appendix claim", "full_article_read": false }, "source_provenance": [ { "source_kind": "fixed-primary-paper", "source_url": "https://arxiv.org/html/2611.00530v1" } ], "reason": "固定低成本L1组件已验收；完整高成本实验不复现，不宣称模型能力或Evolve改善", "component_evidence": "docs/recommendation-components/2611.00530-wesearch/metrics/component-cpu-seeds42-44.json" }, { "domain": "agent-research", "key": "agentheal", "title": "AgentHeal: Evaluating Goal-Preserving Resource Recovery of Web Agents under Access Barriers", "paper_url": "https://arxiv.org/abs/2610.39234", "detail_path": "agent-research/2610.39234-agentheal/README.md", "priority": "P1", "first_author": "Zhiyi Liu", "first_author_affiliation": "University of Science and Technology of China", "published": "2026-10-29", "published_at": "2026-10-29T02:51:58Z", "code": "https://github.com/Lzzzzzzzzzzy/AgentHeal", "code_status": "论文链接的作者官方代码与数据仓库公开；已核对agentheal、dataset、dataset_real、scripts与tests目录，未执行上游全部模型评测", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.39234v1", "source_location": "Section5.1 and evaluation protocol", "available_resource": "one NVIDIA A100 or A30, or local Mac", "model_scale": "13 models,including8B/9B/27B local checkpoints and paid API models,plus semantic LLM judge", "data_scale": "855 tasks,up to60 steps per episode,real web replay and independent audit", "estimated_cost_basis": "full actor/model/task/judge matrix is long-running and resource cost is not bounded by the lightweight evaluator;published configuration,not a measured local runtime", "decision": "完整实验成本过高，不复现（审查 DONE）；只执行独立评价侧状态合同，不调浏览器、模型或judge，不自动重试" }, "implementation_status": "component_only", "requires_gpu_validation": false, "gpu_validation_passed": false, "gpu_validation_artifact": null, "id": "2610.39234", "status": "implemented", "resolution": "bounded_l1_component_accepted", "doc": "agent-research/2610.39234-agentheal/README.md", "catalog_key": "agentheal", "catalog_detail": "agent-research/2610.39234-agentheal/README.md", "track": "agent", "intake_class": "recent_overlap_backfill", "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "component_implemented": true, "component_accepted": true, "diagnostic_only": true, "promotion_eligible": false, "evolve_integrated": false, "automatic_retry": false, "automatic_implementation_queue": false, "verified_priority_organizations": [], "full_text_review": { "scope": "defining method,evidence,resources,authorship and source-release review; not every reference or appendix claim", "full_article_read": false }, "source_provenance": [ { "source_kind": "fixed-primary-paper", "source_url": "https://arxiv.org/html/2610.39234v1" } ], "reason": "固定低成本L1组件已验收；完整高成本实验不复现，不宣称模型能力或Evolve改善", "component_evidence": "docs/agent-research/2610.39234-agentheal/metrics/component-cpu-seeds42-44.json" }, { "domain": "agent-research", "key": "strategist", "title": "Strategist: Evolving Reasoning with Agents", "paper_url": "https://arxiv.org/abs/2611.03618", "detail_path": "agent-research/2611.03618-strategist/README.md", "priority": "P0", "first_author": "Taiming Lu", "first_author_affiliation": "Princeton University", "published": "2026-11-02", "published_at": "2026-11-02T18:51:04Z", "code": "https://github.com/zlab-princeton/Strategist", "code_status": "作者官方仓库公开search、blueprint、retrieval、paired_pick和空初始库；已只读核对默认hidden_best与two_percent规则，未执行上游完整搜索", "cost_assessment": { "source_url": "https://arxiv.org/html/2611.03618v1", "source_location": "Experiment and implementation configuration", "available_resource": "one NVIDIA A100 or A30, or local Mac", "model_scale": "frozen8B executor plus external meta-agent inference;authors8xH200 setup", "data_scale": "30 benchmarks,10 iterations x5 candidates per task,about20K candidate-samples/task and108 baseline configurations", "estimated_cost_basis": "multi-benchmark repeated model calls and paid meta-agent searches exceed the single-device bounded scope;published configuration,not a local measurement", "decision": "完整实验成本过高，不复现（审查 DONE）；只执行CPU控制与测量协议，不进行模型搜索、下载或自动重试" }, "implementation_status": "component_only", "requires_gpu_validation": false, "gpu_validation_passed": false, "gpu_validation_artifact": null, "id": "2611.03618", "status": "implemented", "resolution": "bounded_l1_component_accepted", "doc": "agent-research/2611.03618-strategist/README.md", "catalog_key": "strategist", "catalog_detail": "agent-research/2611.03618-strategist/README.md", "track": "agent", "intake_class": "new_first_publication_since_previous_capture", "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "component_implemented": true, "component_accepted": true, "diagnostic_only": true, "promotion_eligible": false, "evolve_integrated": false, "automatic_retry": false, "automatic_implementation_queue": false, "verified_priority_organizations": [ "Meta" ], "full_text_review": { "scope": "defining method,evidence,resources,authorship and source-release review; not every reference or appendix claim", "full_article_read": false }, "source_provenance": [ { "source_kind": "fixed-primary-paper", "source_url": "https://arxiv.org/html/2611.03618v1" } ], "reason": "固定低成本L1组件已验收；完整高成本实验不复现，不宣称模型能力或Evolve改善", "component_evidence": "docs/agent-research/2611.03618-strategist/metrics/component-cpu-seeds42-44.json" }, { "domain": "foundation-models", "key": "gavagai", "title": "GAVAGAI: Referential Alignment for Infant-Scale Vision-Language Models via Balanced Optimal Transport", "paper_url": "https://arxiv.org/abs/2611.00319", "detail_path": "foundation-models/2611.00319-gavagai/README.md", "priority": "P1", "first_author": "Rajdeep Pal", "first_author_affiliation": "St Thomas’ College of Engineering and Technology", "published": "2026-10-30", "published_at": "2026-10-30T19:39:12Z", "code": null, "code_status": "固定v1全文、摘要及方法名作者名检索未找到可核验原作者实现；未找到/尚未发布", "cost_assessment": { "source": "固定v1 Section3.4：https://arxiv.org/html/2611.00319v1#S3.SS4", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "model_scale": "微调DINOv2视觉初始化与6层8头hidden512随机初始化文本decoder，12个区域proposal", "data_scale": "SAYCam train split，40epochs，batch256，另需手标2000-token内容词过滤训练", "paper_configuration": "作者报告4张NVIDIA A10040GB约11墙钟小时；3epoch caption-only warmup后加入对齐", "estimated_cost_basis": "4卡×11h为作者报告，不是本地测量，也不直接线性外推单卡墙钟；完整训练超出现有轻量预算", "decision": "完整实验成本过高，不复现（审查DONE）；仅实现可调用的容量上界OT、Nothing成本head与固定plan对齐损失，不声称SAYCam能力或完整防塌缩训练。" }, "implementation_status": "component_only", "requires_gpu_validation": true, "gpu_validation_passed": true, "gpu_validation_artifact": "docs/gpu-validations/gavagai-a100-20261104.json", "id": "2611.00319", "status": "implemented", "resolution": "bounded_l1_component_accepted", "doc": "foundation-models/2611.00319-gavagai/README.md", "catalog_key": "gavagai", "catalog_detail": "foundation-models/2611.00319-gavagai/README.md", "track": "foundation-model", "intake_class": "new_first_publication_since_previous_capture", "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "component_implemented": true, "component_accepted": true, "diagnostic_only": true, "promotion_eligible": false, "evolve_integrated": false, "automatic_retry": false, "automatic_implementation_queue": false, "verified_priority_organizations": [], "full_text_review": { "scope": "defining method,evidence,resources,authorship and source-release review; not every reference or appendix claim", "full_article_read": false }, "source_provenance": [ { "source_kind": "fixed-primary-paper", "source_url": "https://arxiv.org/html/2611.00319v1" } ], "reason": "固定低成本L1组件已验收；完整高成本实验不复现，不宣称模型能力或Evolve改善", "component_evidence": "docs/foundation-models/2611.00319-gavagai/metrics/component-cuda-seeds42-44.json" }, { "domain": "foundation-models", "key": "xla-trenches", "title": "XLA from the Trenches: Correctness on TPU Workloads at Scale", "paper_url": "https://arxiv.org/abs/2611.00624", "detail_path": "foundation-models/2611.00624-xla-trenches/README.md", "priority": "P0", "first_author": "Ignacio Cano", "first_author_affiliation": "Google", "published": "2026-10-31", "published_at": "2026-10-31T04:48:13Z", "code": "https://github.com/openxla/xla/blob/main/docs/hlo_isolation.md", "code_status": "原文明确macro-graph isolator已上游到OpenXLA，公开HLO isolation文档及evaluator；完整内部agent harness、TPU工具层、生产图集与微融合硬件调试路径未公开，不等于本文整套作者实现已开源", "cost_assessment": { "source": "固定v1 Sections2.3、3.1.3、3.2、4：https://arxiv.org/html/2611.00624v1", "available_resource": "单张NVIDIA A100或A30，或本地Mac，不具备TPU内部调试环境", "model_scale": "多代TPU上真实生产HLO/LLO融合图、内部ISA与硬件精确MXU模拟器", "data_scale": "约20万私有生产图样本、4万余静态测试定义和参数化后约20万测试，每轮跨后端逾百万次执行", "paper_configuration": "内部多agent harness与TPU工具层，硬件VREG逐步调试，最后以物理TPU确认", "estimated_cost_basis": "资源类型不匹配且私有生产/内部调试依赖不可获取，未虚构GPU小时或尝试CUDA替代", "decision": "完整实验成本过高且环境不可用，不复现（审查DONE）；只保留方法/证据/上游可见边界，不实现伪TPU检查器，不产生metrics或GPU收据。" }, "implementation_status": "deferred_high_cost", "requires_gpu_validation": false, "gpu_validation_passed": false, "gpu_validation_artifact": null, "id": "2611.00624", "status": "deferred", "resolution": "not_reproduced_high_cost", "doc": "foundation-models/2611.00624-xla-trenches/README.md", "catalog_key": "xla-trenches", "catalog_detail": "foundation-models/2611.00624-xla-trenches/README.md", "track": "foundation-model", "intake_class": "new_first_publication_since_previous_capture", "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "component_implemented": false, "component_accepted": false, "diagnostic_only": true, "promotion_eligible": false, "evolve_integrated": false, "automatic_retry": false, "automatic_implementation_queue": false, "verified_priority_organizations": [ "Google" ], "full_text_review": { "scope": "defining method,evidence,resources,authorship and source-release review; not every reference or appendix claim", "full_article_read": false }, "source_provenance": [ { "source_kind": "fixed-primary-paper", "source_url": "https://arxiv.org/html/2611.00624v1" } ], "reason": "Google TPU系统，reviewDONE/高成本不复现；不运行CUDA替身" }, { "domain": "foundation-models", "key": "proland", "title": "ProLand: A New Infeasible Manifold Optimization Method for LoRA Fine-Tuning", "paper_url": "https://arxiv.org/abs/2611.00236", "detail_path": "foundation-models/2611.00236-proland/README.md", "priority": "P1", "first_author": "Chenyu Zhou", "first_author_affiliation": "Shanghai Jiao Tong University", "published": "2026-10-30", "published_at": "2026-10-30T18:18:00Z", "code": null, "code_status": "固定v1全文、摘要及作者名方法名检索未找到可核验原作者ProLand实现；未找到/尚未发布，不把Muon或其他LoRA方法代码算作本文代码", "cost_assessment": { "source": "固定v1 Sections 3、5与Appendix L：https://arxiv.org/html/2611.00236v1", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "model_scale": "DeBERTa-v3-base、Qwen2.5-3B、Llama-3.2-3B、Moonlight-16B-A3B，多adapter ranks", "data_scale": "4项GLUE验证任务及7项语言benchmark，多模型与rank配置和消融", "paper_configuration": "附录L在H100 80GB上给出rank16、sequence512、effective batch640、290 updates的主实验和9次primary runs；另有消融配置", "estimated_cost_basis": "依据论文模型、训练与消融矩阵做只读判断；未测量完整单卡小时，也未为证明成本下载checkpoint", "decision": "完整实验成本过高，不复现（审查DONE）；本地仅执行小矩阵精确thin-factor优化器与独立稠密参考，不宣称11任务效果或生产吞吐。" }, "implementation_status": "component_only", "requires_gpu_validation": true, "gpu_validation_passed": true, "gpu_validation_artifact": "docs/gpu-validations/proland-a100-20261104.json", "id": "2611.00236", "status": "implemented", "resolution": "bounded_l1_component_accepted", "doc": "foundation-models/2611.00236-proland/README.md", "catalog_key": "proland", "catalog_detail": "foundation-models/2611.00236-proland/README.md", "track": "foundation-model", "intake_class": "new_first_publication_since_previous_capture", "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "component_implemented": true, "component_accepted": true, "diagnostic_only": true, "promotion_eligible": false, "evolve_integrated": false, "automatic_retry": false, "automatic_implementation_queue": false, "verified_priority_organizations": [], "full_text_review": { "scope": "defining method,evidence,resources,authorship and source-release review; not every reference or appendix claim", "full_article_read": false }, "source_provenance": [ { "source_kind": "fixed-primary-paper", "source_url": "https://arxiv.org/html/2611.00236v1" } ], "reason": "固定低成本L1组件已验收；完整高成本实验不复现，不宣称模型能力或Evolve改善", "component_evidence": "docs/foundation-models/2611.00236-proland/metrics/component-cuda-seeds42-44.json" }, { "domain": "post-training", "key": "specopd", "title": "SpecOPD: Branch-Parallel On-Policy Distillation for Speculative Decoding", "paper_url": "https://arxiv.org/abs/2611.02846v1", "detail_path": "post-training/2611.02846-specopd/README.md", "priority": "P1", "first_author": "Tianyu Liu", "first_author_affiliation": "Qwen Business Unit of Alibaba / University of Science and Technology of China", "published": "2026-11-02", "published_at": "2026-11-02T09:47:06Z", "code": null, "code_status": "截至2026-11-04，官方abs/HTML未给SpecOPD作者仓库；精确题名及GitHub检索未核实作者实现。原文引用的DFlash/Draft-OPD仓库不是SpecOPD开源证据。本地据公式独立重写", "cost_assessment": { "source": "Sec.3、Appendix A.2-A.8及B.2；https://arxiv.org/html/2611.02846v1", "model_scale": "8B全注意力、27B全-线性混合、125B稀疏-线性混合target及DFlash/DSpark/DFlash V2 drafter", "data_scale": "Open-PerfectBlend外层reference及多个dialogue/code/math/agentic评测；作者保留子集与准确checkpoint仍未完整公布", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "paper_configuration": "27B训练吞吐对照使用8张高性能GPU；8B drafter比较涉及10000步，27B主选择checkpoint为36000步；部分serving context40960、并发32", "estimated_cost_basis": "原文多模型、长上下文、多并发、万步训练矩阵明显超出单卡低成本预算；不将native吞吐倍数当作单卡训练耗时估算", "decision": "成本过高，不复现（审查 DONE）；仅L1小张量组件，不下载模型和数据、不完整训练、不自动重试" }, "implementation_status": "component_only", "requires_gpu_validation": true, "gpu_validation_passed": true, "gpu_validation_artifact": "docs/gpu-validations/specopd-a100-20261104.json", "id": "2611.02846", "status": "implemented", "resolution": "bounded_l1_component_accepted", "doc": "post-training/2611.02846-specopd/README.md", "catalog_key": "specopd", "catalog_detail": "post-training/2611.02846-specopd/README.md", "track": "post-training", "intake_class": "new_first_publication_since_previous_capture", "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "component_implemented": true, "component_accepted": true, "diagnostic_only": true, "promotion_eligible": false, "evolve_integrated": false, "automatic_retry": false, "automatic_implementation_queue": false, "verified_priority_organizations": [], "full_text_review": { "scope": "defining method,evidence,resources,authorship and source-release review; not every reference or appendix claim", "full_article_read": false }, "source_provenance": [ { "source_kind": "fixed-primary-paper", "source_url": "https://arxiv.org/abs/2611.02846v1" } ], "reason": "固定低成本L1组件已验收；完整高成本实验不复现，不宣称模型能力或Evolve改善", "component_evidence": "docs/post-training/2611.02846-specopd/metrics/component-cuda-seeds42-44.json" }, { "domain": "post-training", "key": "rlid", "title": "RLiD: Reinforcement Learning in the Decision Plane for Language Agents and LLMs", "paper_url": "https://arxiv.org/abs/2611.02455v1", "detail_path": "post-training/2611.02455-rlid/README.md", "priority": "P1", "first_author": "Ziyi Chen", "first_author_affiliation": "University of Illinois Urbana-Champaign", "published": "2026-11-02", "published_at": "2026-11-02T03:17:51Z", "code": null, "code_status": "截至2026-11-04，官方abs/HTML未给作者公开仓库URL；文中ldmrl/和llmdistill/raw/路径不是可访问开源链接，精确题名/编号检索未找到已核实作者仓库。本地据公式独立重写", "cost_assessment": { "source": "Sec.3 compute accounting、Sec.6及Appendix N；https://arxiv.org/html/2611.02455v1#S3", "model_scale": "Qwen2.5 0.5B/1.5B/3B/7B及Qwen3.5 2B/4B/9B LoRA rank16；另有跨模型学生", "data_scale": "ARC、Countdown、Sokoban、FrozenLake、ALFWorld多环境，多seed、教师轨迹收集、蒸馏及RL", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "paper_configuration": "ALFWorld每episode最多50步、每组8轨迹；完整训练-蒸馏链路作者报告12.7 H100 GPU-hours，对照17.3 GPU-hours", "estimated_cost_basis": "作者实测已为小时级，且其H100结果不能当成本地A100耗时承诺；多环境、多模型与完整轨迹蒸馏矩阵超出本轮低成本预算", "decision": "成本过高，不复现（审查 DONE）；只验证数值readout、决策PPO和同策略方差恒等式，不加载权重、不跑环境或完整训练、不自动重试" }, "implementation_status": "component_only", "requires_gpu_validation": true, "gpu_validation_passed": true, "gpu_validation_artifact": "docs/gpu-validations/rlid-a100-20261104.json", "id": "2611.02455", "status": "implemented", "resolution": "bounded_l1_component_accepted", "doc": "post-training/2611.02455-rlid/README.md", "catalog_key": "rlid", "catalog_detail": "post-training/2611.02455-rlid/README.md", "track": "post-training", "intake_class": "new_first_publication_since_previous_capture", "review_done": true, "full_experiment_status": "not_reproduced_high_cost", "full_experiment_review_done": true, "component_implemented": true, "component_accepted": true, "diagnostic_only": true, "promotion_eligible": false, "evolve_integrated": false, "automatic_retry": false, "automatic_implementation_queue": false, "verified_priority_organizations": [], "full_text_review": { "scope": "defining method,evidence,resources,authorship and source-release review; not every reference or appendix claim", "full_article_read": false }, "source_provenance": [ { "source_kind": "fixed-primary-paper", "source_url": "https://arxiv.org/abs/2611.02455v1" } ], "reason": "固定低成本L1组件已验收；完整高成本实验不复现，不宣称模型能力或Evolve改善", "component_evidence": "docs/post-training/2611.02455-rlid/metrics/component-cuda-seeds42-44.json" } ], "additional_dispositions": [ { "title": "FabScore: Fine-Grained Evaluation of Fabrications in Automated AI Research", "organization": "Google collaboration", "source_url": "https://research.google/pubs/fabscore-fine-grained-evaluation-of-fabrications-in-automated-ai-research-2/", "decision": "historical_duplicate_title_with_new_readable_source", "review_done": false, "blocking": false, "evidence": "Old official title already exists in the October archive; the new URL points to readable author PDF and MIT author repository https://github.com/chchenhui/fabscore. OpenReview/API challenge still prevents exact first-publication date confirmation. No November-new identity or implementation is invented.", "source_attempts": [ { "url": "https://chchenhui.github.io/papers/FabScore.pdf", "result": "readable35-pagePDF;method/figure/taxonomychecked" }, { "url": "https://api2.openreview.net/notes?id=6gYWcYQcWM", "result": "HTTP403" }, { "url": "https://openreview.net/forum?id=6gYWcYQcWM", "result": "browserverificationinterstitial" } ], "reopen_condition": "A primary exact original public date becomes available; no periodic retry without changed evidence", "automatic_retry": false }, { "title": "Entitlement Is Not Authorization", "organization": "Google", "source_url": "https://research.google/pubs/entitlement-is-not-authorization/", "decision": "outside_current_topic_scope", "review_done": true, "reason": "General entitlement consistency/billing model checking and seeded subscription simulation, not an LLM,Agent,JEPA or search/recommendation method. Google priority does not remove topic boundaries." }, { "title": "What Limits Us? Analyzing Self-Reported Limitations in NLP Research", "organization": "Google collaboration", "id": "2609.15191", "source_url": "https://research.google/pubs/what-limits-us-analyzing-self-reported-limitations-in-nlp-research/", "decision": "already_reviewed_historical_identity", "review_done": true, "previous_record": "docs/research-audits/google-emnlp25-closure/review-decisions.json", "reason": "Known human-AI qualitative research meta-analysis; new directory URL is not a new autonomous RSI implementation." }, { "title": "Hierarchical Functional Fault Grading: A Chiplet-Based Methodology to Supplement Structural ATPG Coverage", "organization": "Google", "source_url": "https://research.google/pubs/hierarchical-functional-fault-grading-a-chiplet-based-methodology-to-supplement-structural-atpg-coverage-3/", "decision": "outside_current_topic_scope", "review_done": true, "reason": "Manufacturing-defect fault simulation for chiplets,not ML model training/inference or an agent algorithm;no claimed new AI implementation." }, { "title": "EmbeddingGemma 2: Unified and Efficient Multimodal Representations", "organization": "Google DeepMind", "source_url": "https://deepmind.google/research/publications/279443/", "decision": "already_catalogued_official_url_addition", "review_done": true, "catalog_key": "embedding-gemma2", "reason": "Known model/paper and existing checkpoints;directory URL delta is not a second model or new-first-publication count." } ] }