{
  "schema_version": 1,
  "batch": "2026-10-25-google-emnlp25-closure",
  "date": "2026-10-25",
  "scope_kind": "explicitly-approved-historical-google-conference-leads",
  "selected_scope_complete": true,
  "delivery_status": "bounded_scope_complete",
  "review_scope_closed": true,
  "all_sources_exhaustive": false,
  "review_watermark_advanced": false,
  "review_watermark": {
    "before": "2026-10-15",
    "after": "2026-10-15"
  },
  "input_artifact": "docs/research-audits/oct25-priority-components/google-emnlp-title-identity-audit.json",
  "input_sha256": "20c113f56ff8d78dfe24f9ea4b9330d7dc9e4316712c3cfaa8703e21e8e28dba",
  "fixed_input_identities": [
    "2503.08669",
    "2505.20417",
    "2507.23220",
    "2509.26600",
    "2510.00263",
    "2510.01637",
    "2601.12208",
    "2601.13880",
    "2602.05945",
    "2603.04364",
    "2603.17070",
    "2604.16287",
    "2605.12483",
    "2605.29442",
    "2605.30788",
    "2606.04095",
    "2606.14958",
    "2606.24579",
    "2607.18754",
    "2607.24484",
    "2607.27853",
    "2608.03276",
    "2609.01725",
    "2609.15191",
    "2609.17764"
  ],
  "fixed_component_keys": [
    "benchmark-self-bias",
    "combinatorial-watermark",
    "cue2narrate-drl",
    "dmast",
    "finance-harness-protocol",
    "interpolated-null-calibration",
    "mechanistic-etm",
    "polaris-hri",
    "probabilistic-autorater",
    "rm-memorization-protocol",
    "scar-shapley-credit",
    "sopbench-protocol",
    "taskpress",
    "xlgobench"
  ],
  "counts": {
    "component_only": 14,
    "not_reproduced_high_cost": 8,
    "not_selected": 3,
    "reviewed_input_identities": 25,
    "components_pending_verification": 0,
    "components_accepted": 14,
    "gpu_components_pending": 0,
    "new_first_publications": 0,
    "capability_reproductions_claimed": 0,
    "gpu_validated_components": 10
  },
  "papers": [
    {
      "id": "2503.08669",
      "title": "SOPBench: Evaluating Language Agents at Following Standard Operating Procedures and Constraints",
      "paper_url": "https://arxiv.org/abs/2503.08669v1",
      "reviewed_version_url": "https://arxiv.org/html/2503.08669v2",
      "published_at": "2025-03-11T17:53:02Z",
      "first_author": "Zekun Li",
      "first_author_affiliation": "University of California, Santa Barbara",
      "verified_priority_organizations": [
        "Google DeepMind"
      ],
      "sources": [
        {
          "url": "https://arxiv.org/abs/2503.08669",
          "role": "exact v1UTC, v2 revised2025-06-17"
        },
        {
          "url": "https://arxiv.org/html/2503.08669v2",
          "role": "ChiWang GoogleDeepMind; Section2 full rule protocol, Section3 model/budget matrix"
        },
        {
          "url": "https://github.com/Leezekun/SOPBench/tree/8790ed7406acbd22270f90bd1c4038de3fc5db8f",
          "role": "MIT source, separate DATA_LICENSE; seven domains and procedure dependency evaluation"
        }
      ],
      "upstream_revision": "8790ed7406acbd22270f90bd1c4038de3fc5db8f",
      "code_license": "MIT; data separately licensed",
      "method_review": "Service-to-constraint and constraint-to-helper mappings form directed SOP graphs. Single/And/Or/Chain nested compositions; all invoked services must be permitted, final database must match oracle, and all prerequisite checks must precede the service in correct order. Implemented evaluator consumes argument-qualified helper identities and private oracle verdicts post-trajectory; no policy sees them. Correct refusal can pass only with matching final state. Not the seven-domain tool environment or an LLM agent.",
      "semantic_deduplication": "No ID/title/semantic nestedSOP protocol match. Generic tool-success score lacks all three dimensions and chain/alternative branches.",
      "cost_assessment": {
        "source": "v2 Section2/3",
        "model_scale": "18models including70B/72B and closed reasoningAPIs",
        "data_scale": "903tests/sevendomains/167tools",
        "available_resource": "one A100 or A30, or local Mac",
        "paper_configuration": "up to20turns and5attemptspercase;full/oracle tool and jailbreak comparisons",
        "estimated_cost_basis": "Whole multi-model comparison expensive; deterministic graph protocol low-cost and independent of models",
        "decision": "full experiment not_reproduced_high_cost; CPU protocol component accepted"
      },
      "decision": "component_only",
      "reason": "Faithful structural evaluator is independently testable without faking policy reasoning or exposing gold to a policy.",
      "review_done": true,
      "accepted_for_implementation": true,
      "implemented": true,
      "catalog_key": "sopbench-protocol",
      "domain": "agent-research",
      "local_module": "auto_research.agent_research.sopbench_protocol",
      "requires_gpu_validation": false,
      "verification": "PureCPU;18combined focused tests pass; no model execution",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "SOPBench: Evaluating Language Agents at Following Standard Operating Procedures and Constraints",
      "component_implemented": true,
      "component_accepted": true,
      "implementation_status": "component_only",
      "component_evidence": "docs/agent-research/2503.08669-sopbench-protocol/metrics/component-cpu-seeds42-44.json",
      "status": "implemented",
      "resolution": "bounded_l1_component_accepted",
      "priority": "P1",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2505.20417",
      "title": "SCAR: Shapley Credit Assignment for More Efficient RLHF",
      "paper_url": "https://arxiv.org/abs/2505.20417v1",
      "published_at": "2025-05-26T18:06:52Z",
      "first_author": "Meng Cao",
      "first_author_affiliation": "McGill University",
      "verified_priority_organizations": [
        "Google DeepMind"
      ],
      "sources": [
        "https://arxiv.org/abs/2505.20417v1",
        "https://arxiv.org/html/2505.20417v1",
        "https://research.google/conferences-and-events/google-at-emnlp-2026/"
      ],
      "method_review": "原文Sec.3.2–3.4 Eq.3–6：外部冻结RM对原顺序、未入选位置填空格的文本coalition评分，v(empty)=0；Shapley必须保留负贡献，span credit只给该span末token。完整论文用自适应分段和SHAP/Owen近似。这里只保真实现N<=10的完整coalition精确式及gamma=1的奖励散布；不把未指定解析器、Owen层次结构或合成RM冒充原系统。",
      "cost_assessment": {
        "source_version": "v1",
        "configuration": "原文Sec.4与Appendix B：单A100 80GB，5种子，Pythia-1B TL;DR含2000步SFT与92K偏好RM训练，另有7B instruction tuning；1K步token级SCAR约48GPU小时，span级约7GPU小时。论文注明token/span时间对比是单次代表性运行。完整多任务多种子训练明显超短时预算；不下载/不试跑。",
        "decision": "成本过高，不复现（审查 DONE）",
        "automatic_retry": false
      },
      "decision": "component_only",
      "reason": "定义明确的独立数值/统计协议已实现；完整LLM实验高成本不跑",
      "review_done": true,
      "accepted_for_implementation": true,
      "implemented": true,
      "catalog_key": "scar-shapley-credit",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "domain": "post-training",
      "code": null,
      "upstream_revision": null,
      "code_license": null,
      "code_status": "固定原文仅链接上游SHAP/TRL/基线工具，未找到本论文作者源码仓库；精确标题公开检索也未核实作者发布。未找到不等于断言未发布。",
      "method_source_version": "v1",
      "validation_status": "cpu_tests_passed_gpu_pending",
      "reviewed_version_url": "https://arxiv.org/html/2505.20417v1",
      "detail_path": "post-training/2505.20417-scar-shapley-credit/README.md",
      "original_lead_title": "SCAR: Shapley Credit Assignment for More Efficient RLHF",
      "component_implemented": true,
      "component_accepted": true,
      "implementation_status": "component_only",
      "component_evidence": "docs/post-training/2505.20417-scar-shapley-credit/metrics/component-cuda-seeds42-44.json",
      "status": "implemented",
      "resolution": "bounded_l1_component_accepted",
      "gpu_validation_artifact": "docs/gpu-validations/scar-shapley-credit-a100-20261025.json",
      "priority": "P0",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2507.23220",
      "title": "Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders",
      "paper_url": "https://arxiv.org/abs/2507.23220v1",
      "published_at": "2025-07-31T03:17:43Z",
      "first_author": "Carolina Zheng",
      "first_author_affiliation": "Columbia University",
      "verified_priority_organizations": [
        "Google Research"
      ],
      "priority_organization_evidence": "完整v3 PDF第一页上标†对应Google Research，Sweta Karlekar、Asif Mallik、Amir Feder具有该署名；HTML漏掉了完整机构表。",
      "reviewed_version_url": "https://arxiv.org/html/2507.23220v3",
      "version_note": "规范身份与首次日期保持原始v1；本轮方法和资源核对固定v3，未逐行声称其等同v1。",
      "sources": [
        {
          "url": "https://arxiv.org/abs/2507.23220v1",
          "kind": "primary-arxiv-identity-v1-date",
          "reviewed": true
        },
        {
          "url": "https://arxiv.org/html/2507.23220v3",
          "kind": "primary-method-resource-affiliation",
          "reviewed": true
        },
        {
          "url": "https://arxiv.org/pdf/2507.23220v3#page=1",
          "kind": "primary-pdf-affiliation",
          "reviewed": true
        },
        {
          "url": "https://github.com/blei-lab/mechanistic-topic-models/tree/2faface4bd07985f885935d62a11d5e4bf9c73bf",
          "kind": "fixed-author-code-or-placeholder",
          "reviewed": true
        }
      ],
      "method_review": "固定v3 Section4.2.2 Eqs9–13：δ~N(0,I)、θ=softmaxδ、βk=sigmoid(W_in vk+b)，W_in为冻结SAE ENCODER，不是decoder；每个feature计数~Binomial(Ntokens,[θβ]i)。variational q拓扑与MC noise显式外传，优化NLL+analytic Gaussian KL。Eqs16–19仅steering才用独立SAE DECODER加权单位化，再移除已有parallel分量、注入指定strength。以不同encoder/decoder负控防隐式绑权。",
      "cost_assessment": {
        "source": "固定v3正文方法/实验及附录；https://arxiv.org/html/2507.23220v3",
        "available_resource": "单张NVIDIA A100或A30，或本地Mac",
        "paper_configuration": "Gemma2-9B、layer10的16K GemmaScope SAE、8数据集、LDA/ETM/BERTopic多变种及Bayesian优化，GPT4.1 topic judge每数据集2800成对比较和额外coherence/steering评测。完整矩阵超单卡短时预算；不下载LLM/SAE或调API，不编造未报告总GPU小时。",
        "estimated_cost_basis": "依据原文配置和完整比较矩阵判断，不是本地完整实验测量；无需用长试跑确认成本。",
        "decision": "成本过高，不复现（审查 DONE）；完整实验不下载、不试跑、不自动重试。仅独立低成本机制/协议诊断，不冒充完整模型或正式benchmark。"
      },
      "decision": "component_only",
      "reason": "用SAE特征而非词袋表达主题，mETM以冻结SAE encoder计算sigmoid topic-feature概率和二项式计数似然训练logistic-normal主题比例；steering独立使用SAE decoder方向替换激活的parallel分量。本地严格区分外部encoder/decoder，仅验证ELBO与steering，不伪造SAE或完整topic评测。",
      "review_done": true,
      "accepted_for_implementation": true,
      "implemented": true,
      "catalog_key": "mechanistic-etm",
      "domain": "foundation-models",
      "local_operator": "auto_research.foundation_models.mechanistic_etm.MechanisticETM",
      "local_module": "auto_research/foundation_models/mechanistic_etm.py",
      "device": "cuda",
      "upstream_code": {
        "url": "https://github.com/blei-lab/mechanistic-topic-models",
        "revision": "2faface4bd07985f885935d62a11d5e4bf9c73bf",
        "license": "MIT",
        "status": "作者方法代码公开，固定commit核验MIT；本地独立按固定v3公式实现mETM/steering子组件，不复制上游代码"
      },
      "semantic_deduplication": "9个规范arXiv ID与标题均未发现已有catalog实现；POLARIS写作方法与POLARIS-Project/Polaris-7B-Preview同名不同文，不合并。",
      "component_validation_status": "implementation_and_local_contracts_pending_root_device_binding",
      "full_experiment_status": "not_reproduced_high_cost",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders",
      "component_implemented": true,
      "component_accepted": true,
      "implementation_status": "component_only",
      "component_evidence": "docs/foundation-models/2507.23220-mechanistic-etm/metrics/component-cuda-seeds42-44.json",
      "status": "implemented",
      "resolution": "bounded_l1_component_accepted",
      "gpu_validation_artifact": "docs/gpu-validations/mechanistic-etm-a100-20261025.json",
      "priority": "P1",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2509.26600",
      "title": "When LLMs Benchmark Themselves: Deconstructing Self-Bias in Automated Evaluation",
      "paper_url": "https://arxiv.org/abs/2509.26600v1",
      "published_at": "2025-09-30T17:48:35Z",
      "first_author": "Wenda Xu",
      "first_author_affiliation": "Google",
      "verified_priority_organizations": [
        "Google"
      ],
      "sources": [
        "https://arxiv.org/abs/2509.26600v1",
        "https://arxiv.org/html/2509.26600v3",
        "https://research.google/conferences-and-events/google-at-emnlp-2026/"
      ],
      "method_review": "Sec.2 Eq.1使用[segment,target,benchmark-role]的逐segment排名，先按segment平均，再diag−leave-self-out peers。原始分数方向必须明确，MetricX-QE越低越好。论文未公开ties细节，本地平均tie是显式约定；不把peer consensus当绝对真值，不在合成表上宣称真实模型有偏。",
      "cost_assessment": {
        "source_version": "v3",
        "configuration": "Sec.3 / Appendices：Gemini2.5Pro、GPT4.1、ClaudeOpus4，六个翻译方向×三种角色设置，每方向200来源文本与全部三模型输出，10K bootstrap/permutation；还含ChatbotArena及27B/32B/更大开放模型。完整API/模型多语矩阵高成本，不运行。",
        "decision": "成本过高，不复现（审查 DONE）",
        "automatic_retry": false
      },
      "decision": "component_only",
      "reason": "定义明确的独立数值/统计协议已实现；完整LLM实验高成本不跑",
      "review_done": true,
      "accepted_for_implementation": true,
      "implemented": true,
      "catalog_key": "benchmark-self-bias",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "domain": "post-training",
      "code": null,
      "upstream_revision": null,
      "code_license": null,
      "code_status": "固定v3正文、摘要与精确标题检索未核实作者源码链接；不把第三方评测框架当作者实现。",
      "method_source_version": "v3",
      "validation_status": "cpu_tests_passed_gpu_pending",
      "reviewed_version_url": "https://arxiv.org/html/2509.26600v3",
      "detail_path": "post-training/2509.26600-benchmark-self-bias/README.md",
      "original_lead_title": "When LLMs Benchmark Themselves: Deconstructing Self-Bias in Automated Evaluation",
      "component_implemented": true,
      "component_accepted": true,
      "implementation_status": "component_only",
      "component_evidence": "docs/post-training/2509.26600-benchmark-self-bias/metrics/component-cuda-seeds42-44.json",
      "status": "implemented",
      "resolution": "bounded_l1_component_accepted",
      "gpu_validation_artifact": "docs/gpu-validations/benchmark-self-bias-a100-20261025.json",
      "priority": "P0",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2510.00263",
      "title": "Judging with Confidence: Calibrating Autoraters to Preference Distributions",
      "paper_url": "https://arxiv.org/abs/2510.00263v1",
      "published_at": "2025-09-30T20:36:41Z",
      "first_author": "Zhuohang Li",
      "first_author_affiliation": "Google",
      "verified_priority_organizations": [
        "Google",
        "Google DeepMind"
      ],
      "sources": [
        "https://arxiv.org/abs/2510.00263v1",
        "https://arxiv.org/html/2510.00263v1",
        "https://research.google/conferences-and-events/google-at-emnlp-2026/"
      ],
      "method_review": "Sec.3 Eq.1–2 piecewise奖励和实验用epsilon=.001可精确实现。解析成功必须在[0,1]；失败Brier=0、Log=log(epsilon)。本地严格tag parser是明确的本地约定，不是作者未公开parser；数学梯度诊断不代表可以对生成文本float解析反传，更不等于GRPO训练。",
      "cost_assessment": {
        "source_version": "v1",
        "configuration": "Sec.4 / Appendix C：Gemma-2-9B与Qwen2.5-7B全参数SFT/RL；约110K Gemini-2.5-Flash偏好标注，RL每步32/64 prompts、每prompt32 generations、batch512，maxseq2048。单卡完整大模型与付费标注矩阵高成本，不运行。",
        "decision": "成本过高，不复现（审查 DONE）",
        "automatic_retry": false
      },
      "decision": "component_only",
      "reason": "定义明确的独立数值/统计协议已实现；完整LLM实验高成本不跑",
      "review_done": true,
      "accepted_for_implementation": true,
      "implemented": true,
      "catalog_key": "probabilistic-autorater",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "domain": "post-training",
      "code": null,
      "upstream_revision": null,
      "code_license": null,
      "code_status": "固定摘要/正文未给本方法作者源码链接；精确标题检索未核实作者仓库。原文parser实现未公开可核验，本地独立实现并标明约定。",
      "method_source_version": "v1",
      "validation_status": "cpu_tests_passed_gpu_pending",
      "reviewed_version_url": "https://arxiv.org/html/2510.00263v1",
      "detail_path": "post-training/2510.00263-probabilistic-autorater/README.md",
      "original_lead_title": "Judging with Confidence: Calibrating Autoraters to Preference Distributions",
      "component_implemented": true,
      "component_accepted": true,
      "implementation_status": "component_only",
      "component_evidence": "docs/post-training/2510.00263-probabilistic-autorater/metrics/component-cuda-seeds42-44.json",
      "status": "implemented",
      "resolution": "bounded_l1_component_accepted",
      "gpu_validation_artifact": "docs/gpu-validations/probabilistic-autorater-a100-20261025.json",
      "priority": "P0",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2510.01637",
      "title": "Detecting Post-generation Edits to Watermarked LLM Outputs via Combinatorial Watermarking",
      "paper_url": "https://arxiv.org/abs/2510.01637v1",
      "published_at": "2025-10-02T03:33:12Z",
      "first_author": "Liyan Xie",
      "first_author_affiliation": "University of Minnesota",
      "verified_priority_organizations": [
        "Google"
      ],
      "priority_organization_evidence": "固定v1作者栏共同一作Muhammad Siddeek署名Google；第一作者首机构为Minnesota。",
      "reviewed_version_url": "https://arxiv.org/html/2510.01637v1",
      "version_note": "本地核对固定v1。",
      "sources": [
        {
          "url": "https://arxiv.org/abs/2510.01637v1",
          "kind": "primary-arxiv-identity-v1-date",
          "reviewed": true
        },
        {
          "url": "https://arxiv.org/html/2510.01637v1",
          "kind": "primary-method-resource-affiliation",
          "reviewed": true
        }
      ],
      "method_review": "固定v1 Section3 Eqs2–5：调用方提供真实secret分区与base logits；bias δ到当前pattern tag后softmax。Iw(t)对所有cyclic phases检查，global mean；local在完整interior t=w−1..T−w平均w个covering windows。全局>=threshold、局部<threshold，阈值需在calibration控制TypeI。binary周期整周期删除可以不可检测，不能承诺完美定位。",
      "cost_assessment": {
        "source": "固定v1正文方法/实验及附录；https://arxiv.org/html/2510.01637v1",
        "available_resource": "单张NVIDIA A100或A30，或本地Mac",
        "paper_configuration": "完整实验Llama2-7B/OPT1.3B，WikiText多水印强度/模式/1–6token编辑×三种编辑方式，1000条64token、阈值校准及PPL矩阵；原文RTX6000Ada48GB。完整模型数据构造与全矩阵不跑，不编造总耗时；只保留CPU公式和负控。",
        "estimated_cost_basis": "依据原文配置和完整比较矩阵判断，不是本地完整实验测量；无需用长试跑确认成本。",
        "decision": "成本过高，不复现（审查 DONE）；完整实验不下载、不试跑、不自动重试。仅独立低成本机制/协议诊断，不冒充完整模型或正式benchmark。"
      },
      "decision": "component_only",
      "reason": "按周期性词表子集模式给下一token logits加偏置，全局统计任意循环相位的窗口匹配率；局部统计覆盖某token的全部完整窗口，定位可能编辑。CPU实现式2–5，不用固定绝对相位、边界补零或随机词表规则冒充论文定义。",
      "review_done": true,
      "accepted_for_implementation": true,
      "implemented": true,
      "catalog_key": "combinatorial-watermark",
      "domain": "foundation-models",
      "local_operator": "auto_research.foundation_models.combinatorial_watermark.detect",
      "local_module": "auto_research/foundation_models/combinatorial_watermark.py",
      "device": "cpu",
      "upstream_code": {
        "url": null,
        "revision": null,
        "license": "未找到作者源码；论文CC BY4.0",
        "status": "固定论文/精确标题检索未找到原作者方法仓库，源码SHA和许可证不可核验；本地独立公式实现"
      },
      "semantic_deduplication": "9个规范arXiv ID与标题均未发现已有catalog实现；POLARIS写作方法与POLARIS-Project/Polaris-7B-Preview同名不同文，不合并。",
      "component_validation_status": "implementation_and_local_contracts_pending_root_device_binding",
      "full_experiment_status": "not_reproduced_high_cost",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "Detecting Post-generation Edits to Watermarked LLM Outputs via Combinatorial Watermarking",
      "component_implemented": true,
      "component_accepted": true,
      "implementation_status": "component_only",
      "component_evidence": "docs/foundation-models/2510.01637-combinatorial-watermark/metrics/component-cpu-seeds42-44.json",
      "status": "implemented",
      "resolution": "bounded_l1_component_accepted",
      "priority": "P1",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2601.12208",
      "title": "CoReflect: A Reflective Co-Evolution Framework for Improving Conversational Evaluation",
      "paper_url": "https://arxiv.org/abs/2601.12208v1",
      "published_at": "2026-01-18T00:56:06Z",
      "first_author": "Yunzhe Li",
      "first_author_affiliation": "University of Illinois Urbana-Champaign",
      "verified_priority_organizations": [
        "Google DeepMind"
      ],
      "sources": [
        "https://arxiv.org/abs/2601.12208v1",
        "https://arxiv.org/html/2601.12208v2",
        "https://research.google/conferences-and-events/google-at-emnlp-2026/"
      ],
      "method_review": "Sec.2.1–2.5 / Appendix G：953条通过真实LLM一致性检查的persona-scenario；planner/simulator/evaluator/analyzer是核心语义模型。只写JSON状态机、对给定数字求均值或手工替换rubric都不能复现其核心，因此不强建低成本heuristic adapter。",
      "cost_assessment": {
        "source_version": "v2",
        "configuration": "Sec.3.1 / Appendix G：30persona×40scenario筛得953对，平均6.85turn，T=3，约7种大型被测模型；Gemini2.5Flash担当simulator/judge，Gemini2.5Pro做reflective analyzer，经AIStudio/VertexAI。完整API多轮矩阵明显高成本，不运行。",
        "decision": "成本过高，不复现（审查 DONE）",
        "automatic_retry": false
      },
      "decision": "not_reproduced_high_cost",
      "reason": "核心依赖真实大型模型/API多轮实验；通用小算子或heuristic不能替代，完整实验高成本不复现",
      "review_done": true,
      "accepted_for_implementation": false,
      "implemented": false,
      "catalog_key": "coreflect",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "domain": "agent-research",
      "code": null,
      "upstream_revision": null,
      "code_license": null,
      "code_status": "固定v2全文与摘要未提供可核验作者代码；精确名称检索仅发现第三方综述引用，不作为源码证据。",
      "method_source_version": "v2",
      "validation_status": "not_applicable_no_implementation",
      "reviewed_version_url": "https://arxiv.org/html/2601.12208v2",
      "detail_path": "agent-research/2601.12208-coreflect/README.md",
      "original_lead_title": "CoReflect: A Reflective Co-Evolution Framework for Improving Conversational Evaluation",
      "priority": "P0",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2601.13880",
      "title": "LifeAgentBench: Benchmarking LLMs for Long-Horizon, Cross-Dimensional Lifestyle Health Reasoning",
      "paper_url": "https://arxiv.org/abs/2601.13880v1",
      "reviewed_version_url": "https://arxiv.org/html/2601.13880v2",
      "published_at": "2026-01-20T11:51:58Z",
      "first_author": "Ye Tian",
      "first_author_affiliation": "University of California San Diego",
      "verified_priority_organizations": [
        "Google Research"
      ],
      "sources": [
        {
          "url": "https://arxiv.org/abs/2601.13880",
          "role": "exact v1 publication timestamp and revised identity"
        },
        {
          "url": "https://arxiv.org/html/2601.13880v2",
          "role": "GoogleResearch XiaoranFan; Sections3/4/5/6 and AppendixE/F model/cost/license"
        },
        {
          "url": "https://gdfwj.github.io/LifeAgentBench/",
          "role": "official project connects code/data and distinguishes LifeAgent baseline from benchmark"
        },
        {
          "url": "https://github.com/gdfwj/LifeAgentBench/tree/cfc227736965f4cd42261ee780f41050ae44cc67",
          "role": "author-linked CP/SQL evaluation and dataset construction; no root license file, GitHub license=null"
        }
      ],
      "upstream_revision": "cfc227736965f4cd42261ee780f41050ae44cc67",
      "code_license": "not_found; paperAppendixF research/noncommercial restrictions",
      "method_review": "Executable-program-derived QA across four lifestyle dimensions, long windows and users. CP embeds prefiltered records; DP generates validated read-only SQL then answers from returned evidence. Distinct VA, completenessEX conditioned onVA, and answerAcc conditioned onEX. LifeAgent is adaptive decomposition/evidence cache/deterministic-tool synthesis, not a scorer. Numeric/pair/multi-item normalization is separate from agent intelligence. Research benchmark explicitly not clinical diagnosis/treatment.",
      "semantic_deduplication": "No ID/title record. Other SQL/health/protocol modules do not establish LifeAgent's model workflow; no cheap scorer claimed as agent reproduction.",
      "cost_assessment": {
        "source": "fixedv2 AppendixE explicit resource statement",
        "model_scale": "13models including70B and proprietaryAPIs",
        "data_scale": "22573QA/100participants and separate hard-subset tool baselines",
        "available_resource": "one A100 or A30, or local Mac",
        "paper_configuration": "CP/DP4096context, open-model inference300–500GPUhours; larger models4A6000",
        "estimated_cost_basis": "Uses author explicit whole-study GPU-hours rather than invented runtime. No records/checkpoints downloaded or inference attempted",
        "decision": "not_reproduced_high_cost"
      },
      "decision": "not_reproduced_high_cost",
      "reason": "Known high-cost full evaluation; preserve faithful method/protocol documentation and official code link, no synthetic capability claims.",
      "review_done": true,
      "accepted_for_implementation": false,
      "implemented": false,
      "catalog_key": "lifeagentbench",
      "domain": "agent-research",
      "documentation_created": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "LifeAgentBench: Benchmarking LLMs for Long-Horizon, Cross-Dimensional Lifestyle Health Reasoning",
      "priority": "P1",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2602.05945",
      "title": "AgenticTagger: Structured Item Representation for Recommendation with LLM Agents",
      "paper_url": "https://arxiv.org/abs/2602.05945v1",
      "reviewed_version_url": "https://arxiv.org/html/2602.05945v2",
      "published_at": "2026-02-05T18:01:37Z",
      "first_author": "Zhouhang Xie",
      "first_author_affiliation": "University of California San Diego",
      "verified_priority_organizations": [
        "Google",
        "Google DeepMind"
      ],
      "sources": [
        {
          "url": "https://arxiv.org/abs/2602.05945",
          "role": "exact v1 UTC and arXiv title"
        },
        {
          "url": "https://arxiv.org/html/2602.05945v2",
          "role": "UCSD first affiliation and Google student researcher footnote; architect/annotator methodSection4, full experimentSection5/6 and online/A-B/deployment keyword context scan"
        }
      ],
      "upstream_revision": null,
      "code_license": null,
      "code_status": "No author-code link found in fixed paper; not inventing one from conference membership",
      "method_review": "Architect LLM distills hierarchical descriptors from item subcorpora; annotator LLM applies vocabulary and returns error suggestions, refine/distill loop constructs structured categorical IDs. Amazon leave-last-out generative-recommendation evaluation; private ranking set47358users/67097items/206345interactions with100samplednegatives. Critique simulation deliberately leaks target-item descriptors and is interpretability evidence, not fair next-item capability gain.",
      "industrial_gate_review": "Section5.2 private ranking comparison is OFFLINE and merely aligned with deployed backbone. No quantified production A/B or clear full-traffic deployment result found in the fixed full text; generic production/deployed-method references do not qualify. No user-named classic exception exists for AgenticTagger.",
      "semantic_deduplication": "No ID/title match; existing semantic-ID/operator methods are different but do not justify bypassing industrial evidence gate.",
      "cost_assessment": {
        "source": "v2 Section4/5/6",
        "model_scale": "Gemini2.5Flash architect and FlashLite annotators plus recommenders",
        "data_scale": "multiple Amazon item corpora and private newsfeed data",
        "available_resource": "one A100 or A30, or local Mac",
        "estimated_cost_basis": "Gate rejection precedes reproduction; no download, API call or compute probe attempted",
        "decision": "not_selected_not_cost_based"
      },
      "decision": "not_selected",
      "reason": "Does not satisfy industrial recommendation online-evidence gate; Google collaboration is priority, not an exemption. Also do not advertise target-item critique leakage as capability.",
      "review_done": true,
      "accepted_for_implementation": false,
      "implemented": false,
      "catalog_key": null,
      "domain": "recommendation",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "AgenticTagger: Structured Item Representation for Recommendation with LLM Agents",
      "priority": "P1",
      "priority_basis": "本轮明确批准的历史线索审查优先级；不代表通过收录门槛或接收实现"
    },
    {
      "id": "2603.04364",
      "title": "Dual-Modality Multi-Stage Adversarial Safety Training: Robustifying Multimodal Web Agents Against Cross-Modal Attacks",
      "paper_url": "https://arxiv.org/abs/2603.04364v1",
      "reviewed_version_url": "https://arxiv.org/html/2603.04364v2",
      "published_at": "2026-03-04T18:29:54Z",
      "first_author": "Haoyu Liu",
      "first_author_affiliation": "University of California, Berkeley",
      "verified_priority_organizations": [
        "Google",
        "Google DeepMind"
      ],
      "sources": [
        {
          "url": "https://arxiv.org/abs/2603.04364",
          "role": "v1 timestamp; v2 revised2026-10-05"
        },
        {
          "url": "https://arxiv.org/html/2603.04364v2",
          "role": "first UC-Berkeley affiliation/Google internship; DingchengLi GoogleDeepMind and LukasRutishauser Google; Sections3/4/5 AppendixB/C/F"
        },
        {
          "url": "https://github.com/huajianduzhuo-code/DMAST_official/tree/e11444116fe26d94d46b8f90ef6bebd8cb728040",
          "role": "author-linked three-stage training source and fixed README; no root LICENSE, GitHub license=null"
        }
      ],
      "upstream_revision": "e11444116fe26d94d46b8f90ef6bebd8cb728040",
      "code_license": "not_found",
      "method_review": "Three-stage teacher imitation, action-preserving oracle SFT, shared-weight agent/attacker self-play. Implemented only Stage3: normalization key(task,attackmode,envseed), pool both role terminal RAW rewards in attacked groups; clean agent-only groups separate; raw-zero-variance excluded. Success-only length penalty enters advantage numerator after raw mean/std. Eq4 clipped GRPO uses[0.9,1.3], kappaKL and capped old/sampling IS correction on PG only, global valid-token denominator. std correction and external L_t explicit because exact conventions/cosine function unspecified. No attacker generation, browser, model loading, SFT or oracle simulation.",
      "semantic_deduplication": "No ID/title match. Generic GRPO already exists, but does not establish this role-pooled group key, raw-before-shaping and clean/attack isolation; new component tests these defining differences.",
      "cost_assessment": {
        "source": "fixedv2 Section5 and AppendixB",
        "model_scale": "12Bstudent and27Bteacher/oracle/evaluation attacker",
        "data_scale": "multiple4000episode collections,42RLtasks,28heldoutMiniWobtasks,100VWA",
        "available_resource": "one A100 or A30, or local Mac",
        "paper_configuration": "three training stages;64/128parallel collectors;25iteration schedule/48updatesperiteration;32768context",
        "estimated_cost_basis": "Full visual collection and multi-stage training exceeds low-cost budget; exact small shared-policy gradient diagnostic needs no pretrained resources",
        "decision": "full experiment not_reproduced_high_cost; component_only accepted"
      },
      "decision": "component_only",
      "reason": "Faithful standalone Stage3 tensor component is directly testable without replacing core browser/policy intelligence with heuristics; full experiment cost-closed.",
      "review_done": true,
      "accepted_for_implementation": true,
      "implemented": true,
      "catalog_key": "dmast",
      "domain": "agent-research",
      "local_module": "auto_research.agent_research.dmast",
      "requires_gpu_validation": true,
      "gpu_validation_artifact": "docs/gpu-validations/dmast-a100-20261025.json",
      "verification": "18 combined focused tests passed; root performs final real A100 execution; no pretrained data/model downloaded",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "Dual-Modality Multi-Stage Adversarial Safety Training: Robustifying Multimodal Web Agents Against Cross-Modal Attacks",
      "component_implemented": true,
      "component_accepted": true,
      "implementation_status": "component_only",
      "component_evidence": "docs/agent-research/2603.04364-dmast/metrics/component-cuda-seeds42-44.json",
      "status": "implemented",
      "resolution": "bounded_l1_component_accepted",
      "priority": "P1",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2603.17070",
      "title": "Large Reasoning Models Struggle to Transfer Parametric Knowledge Across Scripts",
      "paper_url": "https://arxiv.org/abs/2603.17070v1",
      "published_at": "2026-03-17T18:57:26Z",
      "first_author": "Lucas Bandarkar",
      "first_author_affiliation": "Google Research",
      "verified_priority_organizations": [
        "Google"
      ],
      "sources": [
        "https://arxiv.org/abs/2603.17070v1",
        "https://arxiv.org/html/2603.17070v2",
        "https://research.google/conferences-and-events/google-at-emnlp-2026/"
      ],
      "method_review": "Sec.4–6与Appendix A/H：观察性回归、训练专用entity提示干预和真实LoRA是定义性路径。不能把正则音译、手工插提示或小分类器当成其LLM知识检索机制。本文不是新闭式RL loss；本轮仅文献收录，不强造operator。",
      "cost_assessment": {
        "source_version": "v2",
        "configuration": "Sec.6 / Appendix H：GPT-OSS-120B与Qwen3-30B-A3B，599/387样本、8epochs、LoRA rank32，Gemini生成并过滤+模型风格迁移，另有12/31语言评测。模型参数和多语言长推理成本高，不下载、不试跑。",
        "decision": "成本过高，不复现（审查 DONE）",
        "automatic_retry": false
      },
      "decision": "not_reproduced_high_cost",
      "reason": "核心依赖真实大型模型/API多轮实验；通用小算子或heuristic不能替代，完整实验高成本不复现",
      "review_done": true,
      "accepted_for_implementation": false,
      "implemented": false,
      "catalog_key": "cross-script-knowledge-transfer",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "domain": "post-training",
      "code": null,
      "upstream_revision": null,
      "code_license": null,
      "code_status": "固定v2正文与摘要未找到作者代码链接；Llama-Factory为上游训练工具，不记为该论文作者实现。",
      "method_source_version": "v2",
      "validation_status": "not_applicable_no_implementation",
      "reviewed_version_url": "https://arxiv.org/html/2603.17070v2",
      "detail_path": "post-training/2603.17070-cross-script-knowledge-transfer/README.md",
      "original_lead_title": "Large Reasoning Models Struggle to Transfer Parametric Knowledge Across Scripts",
      "priority": "P0",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2604.16287",
      "title": "NaijaS2ST: A Multi-Accent Benchmark for Speech-to-Speech Translation in Low-Resource Nigerian Languages",
      "paper_url": "https://arxiv.org/abs/2604.16287v1",
      "published_at": "2026-04-17T17:49:46Z",
      "first_author": "Marie Maltais",
      "first_author_affiliation": "Mila - Quebec AI Institute",
      "verified_priority_organizations": [
        "Google DeepMind"
      ],
      "priority_organization_evidence": "固定v3作者栏Min Ma署名Google DeepMind。",
      "reviewed_version_url": "https://arxiv.org/html/2604.16287v3",
      "version_note": "规范身份与首次日期保持原始v1；本轮方法和资源核对固定v3，未逐行声称其等同v1。",
      "sources": [
        {
          "url": "https://arxiv.org/abs/2604.16287v1",
          "kind": "primary-arxiv-identity-v1-date",
          "reviewed": true
        },
        {
          "url": "https://arxiv.org/html/2604.16287v3",
          "kind": "primary-method-resource-affiliation",
          "reviewed": true
        },
        {
          "url": "https://github.com/McGill-NLP/naijas2st/tree/02b60538e621d1afe33178a9ea60a86ed7e8f674",
          "kind": "fixed-author-code-or-placeholder",
          "reviewed": true
        }
      ],
      "method_review": "核固定v3 Sections3–4及AppendixC：284说话人、约6K平行句、质控后约269h；5shot来自DEV，方向分开；Seamless不支持的Hausa映射Arabic、Naijá映射French仅为fine-tuned extension。S2TT SSA-COMET/ChrF++，S2ST先OmnilingualASR转写再SSA-COMET，不能用文本exactmatch替代。",
      "cost_assessment": {
        "source": "固定v3正文方法/实验及附录；https://arxiv.org/html/2604.16287v3",
        "available_resource": "单张NVIDIA A100或A30，或本地Mac",
        "paper_configuration": "每语言40余小时语音，Omnilingual-ASR1B、NLLB3.3B/TinyAya3B、Seamless2.3B mono/multi多个方向3–10epochs，另有Gemini/GPT-Audio推理、TTS和人工评审。整套明显超单卡短时预算，不下载音频/模型、不调用付费API。",
        "estimated_cost_basis": "依据原文配置和完整比较矩阵判断，不是本地完整实验测量；无需用长试跑确认成本。",
        "decision": "成本过高，不复现（审查 DONE）；完整实验不下载、不试跑、不自动重试。无本地指标或GPU收据，不以通用统计/字段校验冒充论文评测。"
      },
      "decision": "not_reproduced_high_cost",
      "reason": "构建Hausa、Igbo、Naijá、Yorùbá与英语平行、多口音语音，比较级联ASR→MT→TTS、Seamless端到端及AudioLLM。完整训练/语音与人工评测高成本，审查DONE不复现；不另造通用字段校验器冒充已接入官方benchmark。",
      "review_done": true,
      "accepted_for_implementation": false,
      "implemented": false,
      "catalog_key": "naijas2st",
      "domain": "multimodal-models",
      "local_operator": null,
      "local_module": null,
      "device": null,
      "upstream_code": {
        "url": "https://github.com/McGill-NLP/naijas2st",
        "revision": "02b60538e621d1afe33178a9ea60a86ed7e8f674",
        "license": "论文AppendixC声称资源CC-BY；固定仓库未找到明确许可证文件，GitHub license=null",
        "status": "作者公开语音训练/评估脚本，固定commit；仓库授权与论文CC-BY资源说明分开记录，不擅自指定源码SPDX"
      },
      "semantic_deduplication": "9个规范arXiv ID与标题均未发现已有catalog实现；POLARIS写作方法与POLARIS-Project/Polaris-7B-Preview同名不同文，不合并。",
      "component_validation_status": "not_applicable_reference_only",
      "full_experiment_status": "not_reproduced_high_cost",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "NaijaS2ST: A Multi-Accent Benchmark for Speech-to-Speech Translation in Low-Resource Nigerian Languages",
      "priority": "P1",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2605.12483",
      "title": "Beyond GRPO and On-Policy Distillation: An Empirical Sparse-to-Dense Reward Principle for Language-Model Post-Training",
      "paper_url": "https://arxiv.org/abs/2605.12483v1",
      "published_at": "2026-05-12T17:57:48Z",
      "first_author": "Hejian Sang",
      "first_author_affiliation": "LinkedIn",
      "verified_priority_organizations": [
        "Google"
      ],
      "sources": [
        "https://arxiv.org/abs/2605.12483v1",
        "https://arxiv.org/html/2605.12483v5",
        "https://research.google/conferences-and-events/google-at-emnlp-2026/",
        "https://github.com/HJSang/OPSD_OnPolicyDistillation/tree/39b3b0a5dccbc02b8f1908468bb685d2ea386d9a"
      ],
      "method_review": "Sec.2–3四阶段工作流和等标签预算经验对照是本论文核心；FKL在teacher occupancy、OPD在student occupancy，教师冻结且需共享tokenizer。库中已有GRPO/OPD/KL算子不等于本文完整流程；不重复包装通用KL或把四个状态标签称为复现。",
      "cost_assessment": {
        "source_version": "v5",
        "configuration": "固定v5 Sec.4：Qwen3-1.7B deployment学生配8B/14B教师，另有Llama8B/70B、代码7B/32B；直接学生GRPO约50.2 H200 GPUh，teacher-first约4–6倍且1.50M teacher rollouts、272K logit queries。论文明确研究的是labeled-data allocation而非最低算力。完整流程超预算，不运行。",
        "decision": "成本过高，不复现（审查 DONE）",
        "automatic_retry": false
      },
      "decision": "not_reproduced_high_cost",
      "reason": "核心依赖真实大型模型/API多轮实验；通用小算子或heuristic不能替代，完整实验高成本不复现",
      "review_done": true,
      "accepted_for_implementation": false,
      "implemented": false,
      "catalog_key": "sparse-to-dense-reward-principle",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "domain": "post-training",
      "code": "https://github.com/HJSang/OPSD_OnPolicyDistillation",
      "upstream_revision": "39b3b0a5dccbc02b8f1908468bb685d2ea386d9a",
      "code_license": "not_confirmed_first_party",
      "code_status": "v5正文直接链接HJSang/OPSD_OnPolicyDistillation，固定SHA已核验；根树只有README/scripts/src，未见根LICENSE，GitHub license=null。不能把该OPD子实现等同四阶段完整流程，本地不复制源码。",
      "method_source_version": "v5",
      "validation_status": "not_applicable_no_implementation",
      "first_author_source_version": "v1_pdf_order_v5_affiliation",
      "author_order_randomized": true,
      "reviewed_version_url": "https://arxiv.org/html/2605.12483v5",
      "detail_path": "post-training/2605.12483-sparse-to-dense-reward-principle/README.md",
      "original_lead_title": "Beyond GRPO and On-Policy Distillation: An Empirical Sparse-to-Dense Reward Principle for Language-Model Post-Training",
      "priority": "P0",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2605.29442",
      "title": "How Coding Agents Fail Their Users: A Large-Scale Analysis of Developer-Agent Misalignment in 20,574 Real-World Sessions",
      "paper_url": "https://arxiv.org/abs/2605.29442v1",
      "reviewed_version_url": "https://arxiv.org/html/2605.29442v2",
      "published_at": "2026-05-28T06:35:39Z",
      "first_author": "Ningzhi Tang",
      "first_author_affiliation": "University of Notre Dame",
      "verified_priority_organizations": [
        "Google"
      ],
      "sources": [
        {
          "url": "https://arxiv.org/abs/2605.29442",
          "role": "exact first publication and v2 revision identity"
        },
        {
          "url": "https://arxiv.org/html/2605.29442v2",
          "role": "author header identifies TaoDong Google; full methodology Section3 and AppendixA/D validation protocols"
        },
        {
          "url": "https://github.com/ND-SaNDwichLAB/coding-agent-misalignment/tree/c887cc920129b2db0bdf218c8550e98f74f42f7c",
          "role": "author-linked MIT release; extraction/validation/annotation/aggregation source and reviewed records"
        }
      ],
      "upstream_revision": "c887cc920129b2db0bdf218c8550e98f74f42f7c",
      "code_license": "MIT",
      "method_review": "Observational human-AI study of20574sessions/1639repos. Episodes grounded in developer pushback, verbatim turn quotes and precision-first extraction; separate validation removes normative-prior and absent-context false positives. Multi-axis human-developed codebook with GPT5.4 annotations and expert quality checks;16118validated episodes from29896extracted. No new executable coding policy, training objective or gold-hidden standardized capability benchmark.",
      "semantic_deduplication": "No ID/title record found. Existing agent coding/repair methods address execution rather than this observational qualitative study.",
      "cost_assessment": {
        "source": "v2 Section3/AppendixA/D",
        "model_scale": "GPT5.4 annotation plus human reviewers",
        "data_scale": "20574sessions,29896extracted/16118validated episodes, multiple expert samples",
        "available_resource": "one A100 or A30, or local Mac",
        "estimated_cost_basis": "Not a cost rejection: reproducing empirical claims requires original conversational corpus, human coding/validation and API analysis, not a toy agent fixture",
        "decision": "not_selected"
      },
      "decision": "not_selected",
      "reason": "Human-factors observational meta-analysis rather than a new implementable agent/learning method. Retain audit evidence, do not fabricate an agent or deterministic classifier and call it a reproduction.",
      "review_done": true,
      "accepted_for_implementation": false,
      "implemented": false,
      "catalog_key": null,
      "domain": "agent-research",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "How Coding Agents Fail Their Users: A Large-Scale Analysis of Developer-Agent Misalignment in 20,574 Real-World Sessions",
      "priority": "P1",
      "priority_basis": "本轮明确批准的历史线索审查优先级；不代表通过收录门槛或接收实现"
    },
    {
      "id": "2605.30788",
      "title": "XLGoBench: Detecting cross-lingual skill gaps with algorithmic tasks",
      "paper_url": "https://arxiv.org/abs/2605.30788v1",
      "published_at": "2026-05-29T03:25:32Z",
      "first_author": "Purvam Jain",
      "first_author_affiliation": "Google DeepMind",
      "verified_priority_organizations": [
        "Google DeepMind"
      ],
      "priority_organization_evidence": "固定v1 PDF第一页Purvam Jain上标1=Google DeepMind。",
      "reviewed_version_url": "https://arxiv.org/pdf/2605.30788v1",
      "version_note": "本地核对固定v1。",
      "sources": [
        {
          "url": "https://arxiv.org/abs/2605.30788v1",
          "kind": "primary-arxiv-identity-v1-date",
          "reviewed": true
        },
        {
          "url": "https://arxiv.org/pdf/2605.30788v1",
          "kind": "primary-method-resource-affiliation",
          "reviewed": true
        },
        {
          "url": "https://arxiv.org/pdf/2605.30788v1#page=1",
          "kind": "primary-pdf-affiliation",
          "reviewed": true
        },
        {
          "url": "https://zenodo.org/records/20423110",
          "kind": "fixed-author-supplement-metadata",
          "reviewed": true
        }
      ],
      "method_review": "固定v1 Section2+AppendixC：SLT财富/系数整数−9..9并jumbled chain；DSA状态/动作数相同；PrDSA15动作、传播完整状态概率、top8集合顺序无关；GraphSP唯一最短路（本地rho0 backbone+bypass固定例）；T-RankHamiltonian DAG+transitive edges、nodes:edges1:3。callback只收task/problem白名单深拷贝，score完整JSON，无LLMjudge。",
      "cost_assessment": {
        "source": "固定v1正文方法/实验及附录；https://arxiv.org/pdf/2605.30788v1",
        "available_resource": "单张NVIDIA A100或A30，或本地Mac",
        "paper_configuration": "原文4模型×5算法×7语言×20复杂度×每级50题，包含frontierAPI和Gemma31B，另有重复/置信区间拟合。完整矩阵和人工验证翻译不跑；本地仅5任务结构化诊断、没有代造已验证翻译或跨语言accuracy。",
        "estimated_cost_basis": "依据原文配置和完整比较矩阵判断，不是本地完整实验测量；无需用长试跑确认成本。",
        "decision": "成本过高，不复现（审查 DONE）；完整实验不下载、不试跑、不自动重试。仅独立低成本机制/协议诊断，不冒充完整模型或正式benchmark。"
      },
      "decision": "component_only",
      "reason": "用同一个抽象算法实例在不同语言表述，扫描复杂度来暴露跨语言差距；本地保留SLT、DSA、概率DSA、唯一最短路和唯一锦标赛排序的结构化小题与exact JSON scorer。gold仅scorer计算，不发给model；未运行官方七语模板、LLM或SMD能力矩阵。",
      "review_done": true,
      "accepted_for_implementation": true,
      "implemented": true,
      "catalog_key": "xlgobench",
      "domain": "foundation-models",
      "local_operator": "auto_research.evaluation.xlgobench.evaluate",
      "local_module": "auto_research/evaluation/xlgobench.py",
      "device": "cpu",
      "upstream_code": {
        "url": "https://zenodo.org/records/20423110",
        "revision": null,
        "license": "Zenodo页面未给清晰license，未下载82.2MB包核验内部许可证；论文arXiv non-exclusive license",
        "status": "作者补充材料v1公开：Zenodo20423110，文件MD5=1fd5dd56c216a407dd5a13f08b0f1065；无Git SHA，不伪造。只核正文与公开元数据，独立实现结构化scorer，不复制官方多语模板"
      },
      "semantic_deduplication": "9个规范arXiv ID与标题均未发现已有catalog实现；POLARIS写作方法与POLARIS-Project/Polaris-7B-Preview同名不同文，不合并。",
      "component_validation_status": "implementation_and_local_contracts_pending_root_device_binding",
      "full_experiment_status": "not_reproduced_high_cost",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "XLGoBench: Detecting cross-lingual skill gaps with algorithmic tasks",
      "component_implemented": true,
      "component_accepted": true,
      "implementation_status": "component_only",
      "component_evidence": "docs/foundation-models/2605.30788-xlgobench/metrics/component-cpu-seeds42-44.json",
      "status": "implemented",
      "resolution": "bounded_l1_component_accepted",
      "priority": "P1",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2606.04095",
      "title": "POLARIS: Guiding Small Models to Write Long Stories",
      "paper_url": "https://arxiv.org/abs/2606.04095v1",
      "published_at": "2026-06-02T18:00:07Z",
      "first_author": "Rishanth Rajendhran",
      "first_author_affiliation": "University of Maryland",
      "verified_priority_organizations": [
        "Google DeepMind"
      ],
      "priority_organization_evidence": "固定v2作者栏John Wieting署名Google DeepMind。",
      "reviewed_version_url": "https://arxiv.org/html/2606.04095v2",
      "version_note": "规范身份与首次日期保持原始v1；本轮方法和资源核对固定v2，未逐行声称其等同v1。",
      "sources": [
        {
          "url": "https://arxiv.org/abs/2606.04095v1",
          "kind": "primary-arxiv-identity-v1-date",
          "reviewed": true
        },
        {
          "url": "https://arxiv.org/html/2606.04095v2",
          "kind": "primary-method-resource-affiliation",
          "reviewed": true
        },
        {
          "url": "https://github.com/RishanthRajendhran/POLARIS/tree/50177c1854d83ef9d93339ba7d87e95583e73340",
          "kind": "fixed-author-code-or-placeholder",
          "reviewed": true
        }
      ],
      "method_review": "核固定v2 Section3.2 Eq5和作者fixed verl core：policy-only mean/sample-std，Aref=0.4 min(1,t/15)(rref−μ)/(σ+1e-6)，step0为0；K5+1，clip0.2、KL关闭，token-mean。相同external rubric给policy和ref评分。按adv符号在log-ratio域选精确clipped分支，避免exp溢出的伪有限forward/NaN backward，不改原目标。",
      "cost_assessment": {
        "source": "固定v2正文方法/实验及附录；https://arxiv.org/html/2606.04095v2",
        "available_resource": "单张NVIDIA A100或A30，或本地Mac",
        "paper_configuration": "Qwen3.5-9B、1388 copyrighted prompt-story pairs、4×A10080GB FSDP、batch8、160steps，每run原文约48h并约60美元judge API。完整实验明确超单卡短时预算；人类故事不公开，不能用自造故事冒充原训练数据。",
        "estimated_cost_basis": "依据原文配置和完整比较矩阵判断，不是本地完整实验测量；无需用长试跑确认成本。",
        "decision": "成本过高，不复现（审查 DONE）；完整实验不下载、不试跑、不自动重试。仅独立低成本机制/协议诊断，不冒充完整模型或正式benchmark。"
      },
      "decision": "component_only",
      "reason": "使用结构化Story Quality LLM judge为长篇写作提供软奖励，并加入一个teacher-forced人类参考轨迹；均值/样本标准差只看5个policy输出，参考advantage按15步warmup到0.4后使用相同clipped surrogate。这里只实现HRI/GRPO更新核心，不生成故事或评估写作能力。",
      "review_done": true,
      "accepted_for_implementation": true,
      "implemented": true,
      "catalog_key": "polaris-hri",
      "domain": "post-training",
      "local_operator": "auto_research.post_training.polaris_hri.hri_grpo_loss",
      "local_module": "auto_research/post_training/polaris_hri.py",
      "device": "cuda",
      "upstream_code": {
        "url": "https://github.com/RishanthRajendhran/POLARIS",
        "revision": "50177c1854d83ef9d93339ba7d87e95583e73340",
        "license": "根目录MIT；vendored verl/修改代码由其Apache-2.0许可治理",
        "status": "作者训练/奖励/评估代码公开，固定commit；HRI人类故事与推理因版权不发布、主checkpoint gated，不把no-HRI公开权重当主方法完整数据"
      },
      "semantic_deduplication": "9个规范arXiv ID与标题均未发现已有catalog实现；POLARIS写作方法与POLARIS-Project/Polaris-7B-Preview同名不同文，不合并。",
      "component_validation_status": "implementation_and_local_contracts_pending_root_device_binding",
      "full_experiment_status": "not_reproduced_high_cost",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "POLARIS: Guiding Small Models to Write Long Stories",
      "component_implemented": true,
      "component_accepted": true,
      "implementation_status": "component_only",
      "component_evidence": "docs/post-training/2606.04095-polaris-hri/metrics/component-cuda-seeds42-44.json",
      "status": "implemented",
      "resolution": "bounded_l1_component_accepted",
      "gpu_validation_artifact": "docs/gpu-validations/polaris-hri-a100-20261025.json",
      "priority": "P1",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2606.14958",
      "title": "MVEB: Massive Video Embedding Benchmark",
      "paper_url": "https://arxiv.org/abs/2606.14958v1",
      "published_at": "2026-06-12T21:06:12Z",
      "first_author": "Adnan El Assadi",
      "first_author_affiliation": "Harvard University",
      "verified_priority_organizations": [
        "Google"
      ],
      "priority_organization_evidence": "固定v1作者栏Deep Shah署名Google LLC。",
      "reviewed_version_url": "https://arxiv.org/html/2606.14958v1",
      "version_note": "本地核对固定v1。",
      "sources": [
        {
          "url": "https://arxiv.org/abs/2606.14958v1",
          "kind": "primary-arxiv-identity-v1-date",
          "reviewed": true
        },
        {
          "url": "https://arxiv.org/html/2606.14958v1",
          "kind": "primary-method-resource-affiliation",
          "reviewed": true
        },
        {
          "url": "https://github.com/embeddings-benchmark/mteb/tree/acbcb191eb24ad1929f1bbcf23b90e982e9da62e",
          "kind": "fixed-author-code-or-placeholder",
          "reviewed": true
        }
      ],
      "method_review": "固定v1 Section3：MVEB23任务从184任务池缩减；var-length fps2/cap64frames，固定帧模型遵循其训练配置，音频按model采样率/长度；无per-task finetune。分类8shot frozen logistic probe，retrieval cosine/nDCG10，VCQA候选池accuracy，视频/音频版本配对且只跑模型支持modality。",
      "cost_assessment": {
        "source": "固定v1正文方法/实验及附录；https://arxiv.org/html/2606.14958v1",
        "available_resource": "单张NVIDIA A100或A30，或本地Mac",
        "paper_configuration": "33个公开checkpoint约200M–10.7B，23主任务及184池，多模型视频+音频/8方向矩阵；Table1单H100测LCO7B：MVEB31.9h、MVEB+302.7h。远超单A100/A30短时预算，不下载模型或视频、不构造假的通用指标代理。",
        "estimated_cost_basis": "依据原文配置和完整比较矩阵判断，不是本地完整实验测量；无需用长试跑确认成本。",
        "decision": "成本过高，不复现（审查 DONE）；完整实验不下载、不试跑、不自动重试。无本地指标或GPU收据，不以通用统计/字段校验冒充论文评测。"
      },
      "decision": "not_reproduced_high_cost",
      "reason": "将分类、零样本分类、聚类、配对分类、八方向跨模态检索及视频QA纳入统一视频嵌入评测，按model声明的帧/音频预算运行并成对比较video-only与audio+video。33模型/184任务完整矩阵高成本，审查DONE不复现，不伪造已接入结果。",
      "review_done": true,
      "accepted_for_implementation": false,
      "implemented": false,
      "catalog_key": "mveb",
      "domain": "multimodal-models",
      "local_operator": null,
      "local_module": null,
      "device": null,
      "upstream_code": {
        "url": "https://github.com/embeddings-benchmark/mteb",
        "revision": "acbcb191eb24ad1929f1bbcf23b90e982e9da62e",
        "license": "Apache-2.0；各上游数据集许可证独立",
        "status": "作者在MTEB生态公开视频任务实现，固定commit核验；不把仅配置/字段校验标记成已跑官方评测"
      },
      "semantic_deduplication": "9个规范arXiv ID与标题均未发现已有catalog实现；POLARIS写作方法与POLARIS-Project/Polaris-7B-Preview同名不同文，不合并。",
      "component_validation_status": "not_applicable_reference_only",
      "full_experiment_status": "not_reproduced_high_cost",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "MVEB: Massive Video Embedding Benchmark",
      "priority": "P1",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2606.24579",
      "title": "Cross-Lingual Exploration for Parametric Knowledge",
      "paper_url": "https://arxiv.org/abs/2606.24579v1",
      "published_at": "2026-06-23T13:42:40Z",
      "first_author": "Elisha Diskind",
      "first_author_affiliation": "The Hebrew University of Jerusalem",
      "verified_priority_organizations": [
        "Google Research"
      ],
      "sources": [
        "https://arxiv.org/abs/2606.24579v1",
        "https://arxiv.org/pdf/2606.24579v1",
        "https://research.google/conferences-and-events/google-at-emnlp-2026/"
      ],
      "method_review": "Sec.3.1–4.4与Appendix C/E：完整提示保持原问题语言，majority/minority-aware实际使用语言模型提示聚合而非简单字符串计数，语义正确性和跨语言一致性也由Gemini判定。把counter或字符串翻译器当主模型会破坏核心；本轮不造heuristic组件。",
      "cost_assessment": {
        "source_version": "v1",
        "configuration": "Sec.4 / Appendix C：CLIKE约2K实体、18K多语query，multipath约234K探索路径；ECLeKTic与CLIKE共17语言，6种前沿模型含Qwen3-235B、Gemini、GPT、Grok，经BatchAPI。完整多模型推理与LLM聚合/判定超预算，不运行。",
        "decision": "成本过高，不复现（审查 DONE）",
        "automatic_retry": false
      },
      "decision": "not_reproduced_high_cost",
      "reason": "核心依赖真实大型模型/API多轮实验；通用小算子或heuristic不能替代，完整实验高成本不复现",
      "review_done": true,
      "accepted_for_implementation": false,
      "implemented": false,
      "catalog_key": "cross-lingual-exploration",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "domain": "foundation-models",
      "code": null,
      "upstream_revision": null,
      "code_license": null,
      "code_status": "固定v1 PDF/摘要未找到作者源码链接；作者个人页面只链接论文与模型/论文索引，不当作作者代码。",
      "method_source_version": "v1",
      "validation_status": "not_applicable_no_implementation",
      "reviewed_version_url": "https://arxiv.org/pdf/2606.24579v1",
      "detail_path": "foundation-models/2606.24579-cross-lingual-exploration/README.md",
      "original_lead_title": "Cross-Lingual Exploration for Parametric Knowledge",
      "priority": "P0",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2607.18754",
      "title": "AgentDebugX: An Open-Source Toolkit for Failure Observability, Attribution, and Recovery in LLM Agents",
      "paper_url": "https://arxiv.org/abs/2607.18754v1",
      "published_at": "2026-07-21T06:21:13Z",
      "first_author": "Kunlun Zhu",
      "first_author_affiliation": "University of Illinois Urbana-Champaign",
      "verified_priority_organizations": [
        "Google"
      ],
      "sources": [
        {
          "url": "https://arxiv.org/abs/2607.18754",
          "role": "v1 exact UTC and author identity"
        },
        {
          "url": "https://arxiv.org/html/2607.18754v1",
          "role": "author header; Google coauthor Xiangru Tang; Sections3/4 and AppendicesB/C/D/E"
        },
        {
          "url": "https://github.com/AgentDebugX/AgentDebugX/tree/2437c73c3183e6b6495259dfbb12464a7f07f6b8",
          "role": "author-linked release; MIT; trace and debugger code exists"
        }
      ],
      "upstream_revision": "2437c73c3183e6b6495259dfbb12464a7f07f6b8",
      "code_license": "MIT",
      "method_review": "Immutable AgentTrajectory/AgentEvent evidence with Detect-Attribute-Recover-Rerun outputs. DeepDebug independently reads global and structure-guided views, adjudicates disagreements, emits evidence-backed causal diagnosis and one repair. It inspects but does not execute original tools; suggested repairs require policy/human gate. ErrorHub memory benefit is not evaluated. Attribution receives reference task answers under Who&When with-ground-truth protocol, but not gold responsible agent/step; cannot advertise as answer-free agent capability.",
      "semantic_deduplication": "No exact ID/title/catalog key found before addition. Existing reflection/agent-memory methods are not the same four-stage debugger, but adding a schema or heuristic would not reproduce DeepDebug.",
      "cost_assessment": {
        "source": "v1 Section4 and AppendixC",
        "model_scale": "Qwen3.5-9B/Qwen3.6-27B plus Gemini diagnosis/search and hosted backbones",
        "data_scale": "184Who&When traces,165GAIA tasks,73failed-task reruns and four repair strategies",
        "available_resource": "one A100 or A30, or local Mac",
        "paper_configuration": "DeepDebug roughly5LLM calls;25trace sample12.8Ktokens/diagnosis; long-agent search/tool reruns",
        "estimated_cost_basis": "Full multi-model attribution/recovery is costly; no GPU/API estimate invented and no run attempted",
        "decision": "not_reproduced_high_cost"
      },
      "decision": "not_reproduced_high_cost",
      "reason": "Cost-closed research record; no cheap rule/schema surrogate for causal diagnosis Agent. Privileged-answer protocol is explicitly disclosed.",
      "review_done": true,
      "accepted_for_implementation": false,
      "implemented": false,
      "catalog_key": "agentdebugx",
      "domain": "agent-research",
      "documentation_created": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "AgentDebugX: An Open-Source Toolkit for Failure Observability, Attribution, and Recovery in LLM Agents",
      "priority": "P1",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2607.24484",
      "title": "What do Reward Models Memorize?",
      "paper_url": "https://arxiv.org/abs/2607.24484v1",
      "published_at": "2026-07-27T14:20:59Z",
      "first_author": "Ivo Verhoeven",
      "first_author_affiliation": "ILLC, University of Amsterdam",
      "verified_priority_organizations": [
        "Google DeepMind"
      ],
      "sources": [
        "https://arxiv.org/abs/2607.24484v1",
        "https://arxiv.org/html/2607.24484v2",
        "https://research.google/conferences-and-events/google-at-emnlp-2026/",
        "https://github.com/ioverho/rm-shortcuts/tree/e737a3b26393b59ab354f7bbac8a4744ea441904"
      ],
      "method_review": "Sec.3 Eq.1–2：对每个真实run的reward差先sigmoid，再分别对训练内20次和留出5次求均值，CM=TM−TG。必须按用户划分，不能同用户不同偏好对跨train/test。组件只聚合调用方提供的真实run分数/manifest并检验用户一致；不制造25个RM、不实现描述不充分的cooccurrence搜索、不把sigmoid(mean margin)冒充mean(sigmoid margin)。",
      "cost_assessment": {
        "source_version": "v2",
        "configuration": "Sec.3 / Appendix B–D：PRISM约57.9K与COMMUNITY约137K偏好对；每数据集25个Llama-3.2-1B LoRA训练，先1epoch SFT再3epochs RM，另有3B 8+2run稳健性与10epoch延长训练。累积数十训练+注释/SAE分析高成本，完整实验不跑。",
        "decision": "成本过高，不复现（审查 DONE）",
        "automatic_retry": false
      },
      "decision": "component_only",
      "reason": "定义明确的独立数值/统计协议已实现；完整LLM实验高成本不跑",
      "review_done": true,
      "accepted_for_implementation": true,
      "implemented": true,
      "catalog_key": "rm-memorization-protocol",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "domain": "post-training",
      "code": "https://github.com/ioverho/rm-shortcuts",
      "upstream_revision": "e737a3b26393b59ab354f7bbac8a4744ea441904",
      "code_license": "not_confirmed_first_party",
      "code_status": "作者正文直接链接ioverho/rm-shortcuts，固定SHA已核验；根树无LICENSE，GitHub license=null，新增作者源码授权未确认。本地据公开论文独立重写，不复制作者代码。",
      "method_source_version": "v2",
      "validation_status": "cpu_tests_passed_gpu_pending",
      "reviewed_version_url": "https://arxiv.org/html/2607.24484v2",
      "detail_path": "post-training/2607.24484-rm-memorization-protocol/README.md",
      "original_lead_title": "What do Reward Models Memorize?",
      "component_implemented": true,
      "component_accepted": true,
      "implementation_status": "component_only",
      "component_evidence": "docs/post-training/2607.24484-rm-memorization-protocol/metrics/component-cuda-seeds42-44.json",
      "status": "implemented",
      "resolution": "bounded_l1_component_accepted",
      "gpu_validation_artifact": "docs/gpu-validations/rm-memorization-protocol-a100-20261025.json",
      "priority": "P0",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2607.27853",
      "title": "FinanceHarness: Autonomous Financial Deep Research Framework",
      "paper_url": "https://arxiv.org/abs/2607.27853v1",
      "reviewed_version_url": "https://arxiv.org/html/2607.27853v2",
      "published_at": "2026-07-30T08:32:04Z",
      "first_author": "Yijia Xiao",
      "first_author_affiliation": "Google Cloud AI Research",
      "verified_priority_organizations": [
        "Google Cloud AI Research"
      ],
      "sources": [
        {
          "url": "https://arxiv.org/abs/2607.27853",
          "role": "v1 exact timestamp; revisedv2 identity"
        },
        {
          "url": "https://arxiv.org/html/2607.27853v2",
          "role": "GoogleCloudAI first-listed affiliation/UCLA second; Sections3/4/5/6.2 PIT, rubric, resource/release policy"
        },
        {
          "url": "https://github.com/Yijia-Xiao/FinanceHarness/tree/1163c5e7a8f35ff6217c3a480f7b04562355e0cc",
          "role": "author-linked Apache2 harness; README points to Google official benchmark"
        },
        {
          "url": "https://github.com/google-research/google-research/tree/b24f2136e8ef405b900b5619760126304f190941/finance_harness",
          "role": "Google official FinanceGym/source directory, fixed path commit and FinanceGym subdirectory verified"
        }
      ],
      "upstream_revision": "1163c5e7a8f35ff6217c3a480f7b04562355e0cc",
      "code_license": "Apache-2.0",
      "method_review": "Layered finance harness overPIT corpus: agent sees only question/cutoff; rubric/thesis/future evidence private to evaluator. Retrieval and direct reads requirepublication<=cutoff; detectedPIT leakage invalidates run. Eq2 scores0..4rubrics, normalizes per question then averages questions equally; pre/post-type breakdown separate. Core report orchestration and private corpus not implemented. CPU protocol supplies explicit external-ranked IDs, metadata and post-submission grades, not a financial agent/judge/security sandbox.",
      "semantic_deduplication": "No ID/title match. Existing Diviner point-in-time incident protocol is different from FinanceGym date-bound corpus reads and question-equal rubrics.",
      "cost_assessment": {
        "source": "v2 Section3/5/6.2 and release policy",
        "model_scale": "27Borchestrator,4Bembedding and multiple hosted backbones",
        "data_scale": "100M+articles,4.37Mgraph edges,500runs/400expertquestions/2464rubrics",
        "available_resource": "one A100 or A30, or local Mac",
        "paper_configuration": "private corpus/rubric,500question multisystem evaluation,172separateGRPOtraininginstances",
        "estimated_cost_basis": "Full data/index/model/LLM-grade pipeline expensive and private inputs not released; no large resource downloaded or full run attempted",
        "decision": "full experiment not_reproduced_high_cost; faithful PIT/rubric protocol accepted"
      },
      "decision": "component_only",
      "reason": "Date/invalid-run and rubric-normalization invariants have exact low-cost checks; do not replace research intelligence or claim original scores.",
      "review_done": true,
      "accepted_for_implementation": true,
      "implemented": true,
      "catalog_key": "finance-harness-protocol",
      "domain": "agent-research",
      "local_module": "auto_research.agent_research.finance_harness_protocol",
      "requires_gpu_validation": false,
      "verification": "PureCPU;18combined focused tests pass including live-web invalidation negative control",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "FinanceHarness: Autonomous Financial Deep Research Framework",
      "component_implemented": true,
      "component_accepted": true,
      "implementation_status": "component_only",
      "component_evidence": "docs/agent-research/2607.27853-finance-harness-protocol/metrics/component-cpu-seeds42-44.json",
      "status": "implemented",
      "resolution": "bounded_l1_component_accepted",
      "priority": "P1",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2608.03276",
      "title": "TaskPress: Query-Agnostic KV Cache Compression via Task-Guided Pruning",
      "paper_url": "https://arxiv.org/abs/2608.03276v1",
      "published_at": "2026-08-04T07:51:37Z",
      "first_author": "Wonpyo Park",
      "first_author_affiliation": "Google",
      "verified_priority_organizations": [
        "Google"
      ],
      "priority_organization_evidence": "固定v2作者栏一作Wonpyo Park署名Google。",
      "reviewed_version_url": "https://arxiv.org/html/2608.03276v2",
      "version_note": "规范身份与首次日期保持原始v1；本轮方法和资源核对固定v2，未逐行声称其等同v1。",
      "sources": [
        {
          "url": "https://arxiv.org/abs/2608.03276v1",
          "kind": "primary-arxiv-identity-v1-date",
          "reviewed": true
        },
        {
          "url": "https://arxiv.org/html/2608.03276v2",
          "kind": "primary-method-resource-affiliation",
          "reviewed": true
        }
      ],
      "method_review": "固定v2 Section3 Eqs1–3：sum over guide→odd-K零padding平均池化；γ=max|v|/(2^(b−1)−1)，scores=keyimportance×γ，各head同budget独立topk。输入显式是task-to-context sliced attention，不看真实下游query；真实K/V按时间排序gather，guideKV不保留，原position不能重编号。K/GQA聚合由调用方显式给出，tie按低index为本地约定。",
      "cost_assessment": {
        "source": "固定v2正文方法/实验及附录；https://arxiv.org/html/2608.03276v2",
        "available_resource": "单张NVIDIA A100或A30，或本地Mac",
        "paper_configuration": "Llama3.1-8B/Qwen3-8B/32B，LongBench/RULER16K至64K上下文多压缩比例/方法矩阵；guide演化单A10040GB原文约25.2min加Gemini调用，完整多任务矩阵更大。只做算子，不跑全量加载/演化、无速度计时复现。",
        "estimated_cost_basis": "依据原文配置和完整比较矩阵判断，不是本地完整实验测量；无需用长试跑确认成本。",
        "decision": "成本过高，不复现（审查 DONE）；完整实验不下载、不试跑、不自动重试。仅独立低成本机制/协议诊断，不冒充完整模型或正式benchmark。"
      },
      "decision": "component_only",
      "reason": "prefill时用task guide对context的注意力求和并局部均值池化，与每token量化value scale相乘，在各KV head独立top-k保留context；task guide随后丢弃。本地实际收集小K/V张量及原position IDs，不加载LLM或声称端到端速度收益。",
      "review_done": true,
      "accepted_for_implementation": true,
      "implemented": true,
      "catalog_key": "taskpress",
      "domain": "foundation-models",
      "local_operator": "auto_research.foundation_models.taskpress.prune_taskpress",
      "local_module": "auto_research/foundation_models/taskpress.py",
      "device": "cuda",
      "upstream_code": {
        "url": null,
        "revision": null,
        "license": "未找到TaskPress专属方法源码或可核验源码许可证；论文arXiv non-exclusive license",
        "status": "正文只说明构建在kvpress之上；未提供TaskPress专属仓库，不把第三方kvpress仓库或许可证当作者方法开源"
      },
      "semantic_deduplication": "9个规范arXiv ID与标题均未发现已有catalog实现；POLARIS写作方法与POLARIS-Project/Polaris-7B-Preview同名不同文，不合并。",
      "component_validation_status": "implementation_and_local_contracts_pending_root_device_binding",
      "full_experiment_status": "not_reproduced_high_cost",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "TaskPress: Query-Agnostic KV Cache Compression via Task-Guided Pruning",
      "component_implemented": true,
      "component_accepted": true,
      "implementation_status": "component_only",
      "component_evidence": "docs/foundation-models/2608.03276-taskpress/metrics/component-cuda-seeds42-44.json",
      "status": "implemented",
      "resolution": "bounded_l1_component_accepted",
      "gpu_validation_artifact": "docs/gpu-validations/taskpress-a100-20261025.json",
      "priority": "P1",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2609.01725",
      "title": "From Visual Cues to Spoken Narration: Rethinking Audio Description",
      "paper_url": "https://arxiv.org/abs/2609.01725v1",
      "published_at": "2026-09-01T18:00:50Z",
      "first_author": "Akshita Gupta",
      "first_author_affiliation": "TU Darmstadt & hessian.AI",
      "verified_priority_organizations": [
        "Google Research"
      ],
      "priority_organization_evidence": "固定v1作者栏Federico Tombari署名Google Research / TU Munich。",
      "reviewed_version_url": "https://arxiv.org/html/2609.01725v1",
      "version_note": "本地核对固定v1。",
      "sources": [
        {
          "url": "https://arxiv.org/abs/2609.01725v1",
          "kind": "primary-arxiv-identity-v1-date",
          "reviewed": true
        },
        {
          "url": "https://arxiv.org/html/2609.01725v1",
          "kind": "primary-method-resource-affiliation",
          "reviewed": true
        },
        {
          "url": "https://github.com/multimodal-ai-lab/Cue2Narrate/tree/08c89ea63281b819a0fd73c060d8c954f9e696ef",
          "kind": "fixed-author-code-or-placeholder",
          "reviewed": true
        }
      ],
      "method_review": "固定v1 Section4.3 Eqs5–8：SFT按每个AD目标token求和；positive/negative评分均为同一当前模型和prompt上的token-mean logp，prompt/padding排除；DRL=-logsigmoid((s+−s−)/0.5)，全loss加0.2DRL。离线负caption生成冻结，但其当前模型likelihood仍反传。",
      "cost_assessment": {
        "source": "固定v1正文方法/实验及附录；https://arxiv.org/html/2609.01725v1",
        "available_resource": "单张NVIDIA A100或A30，或本地Mac",
        "paper_configuration": "6.64M双头定位器，Qwen2.5-VL-7B/Qwen3-VL-8B LoRA，训练32帧/推理64帧，多阶段长电影训练；LSMDC研究限制授权、MAD-Eval NDA。完整数据/定位+VLM训练超单卡短时预算，不下载视频或权重；本文未给总GPU小时，不编造。",
        "estimated_cost_basis": "依据原文配置和完整比较矩阵判断，不是本地完整实验测量；无需用长试跑确认成本。",
        "decision": "成本过高，不复现（审查 DONE）；完整实验不下载、不试跑、不自动重试。仅独立低成本机制/协议诊断，不冒充完整模型或正式benchmark。"
      },
      "decision": "component_only",
      "reason": "先定位视觉线索与 spoken narration 两种不同时间窗，再对同一视觉证据的简洁AD和普通caption计算当前模型长度平均likelihood，以DRL抑制冗长caption先验。本地仅保留Stage2 masked SFT+DRL；不运行定位器、VLM或LongLSMDC能力评测。",
      "review_done": true,
      "accepted_for_implementation": true,
      "implemented": true,
      "catalog_key": "cue2narrate-drl",
      "domain": "multimodal-models",
      "local_operator": "auto_research.multimodal.cue2narrate_drl.description_ranking_loss",
      "local_module": "auto_research/multimodal/cue2narrate_drl.py",
      "device": "cuda",
      "upstream_code": {
        "url": "https://github.com/multimodal-ai-lab/Cue2Narrate",
        "revision": "08c89ea63281b819a0fd73c060d8c954f9e696ef",
        "license": "MIT（占位仓库）；论文CC BY-NC-SA 4.0",
        "status": "作者公开仓库固定commit仅含.gitignore、LICENSE和两行README，未发布方法源码；不把仓库存在当作完整实现已开源"
      },
      "semantic_deduplication": "9个规范arXiv ID与标题均未发现已有catalog实现；POLARIS写作方法与POLARIS-Project/Polaris-7B-Preview同名不同文，不合并。",
      "component_validation_status": "implementation_and_local_contracts_pending_root_device_binding",
      "full_experiment_status": "not_reproduced_high_cost",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "From Visual Cues to Spoken Narration: Rethinking Audio Description",
      "component_implemented": true,
      "component_accepted": true,
      "implementation_status": "component_only",
      "component_evidence": "docs/multimodal-models/2609.01725-cue2narrate-drl/metrics/component-cuda-seeds42-44.json",
      "status": "implemented",
      "resolution": "bounded_l1_component_accepted",
      "gpu_validation_artifact": "docs/gpu-validations/cue2narrate-drl-a100-20261025.json",
      "priority": "P1",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    },
    {
      "id": "2609.15191",
      "title": "What Limits Us? Analyzing Self-Reported Limitations in NLP Research",
      "paper_url": "https://arxiv.org/abs/2609.15191v1",
      "published_at": "2026-09-14T08:11:57Z",
      "first_author": "Tawan Thaepprasit",
      "first_author_affiliation": "Chulalongkorn University",
      "verified_priority_organizations": [
        "Google Research"
      ],
      "sources": [
        {
          "url": "https://arxiv.org/abs/2609.15191",
          "role": "v1 exactUTC and arXiv author spelling; PDF renders first surname Thaeprasit"
        },
        {
          "url": "https://arxiv.org/html/2609.15191v1",
          "role": "author header GoogleResearch DingWang/RemiDenton/PiyawatLertvittayakumjorn; Section3 full hybrid-coding procedure and AppendixK prompts"
        },
        {
          "url": "https://github.com/Sundione/nlp-self-reported-limitations/tree/5f7ffe3cf43aa9b1c97d336d6fc530147cd29db7",
          "role": "author-linked CC-BY4 release; human_coded_dataset/llm_coded_dataset and license verified"
        }
      ],
      "upstream_revision": "5f7ffe3cf43aa9b1c97d336d6fc530147cd29db7",
      "code_license": "CC-BY-4.0",
      "method_review": "Meta-research of ACL/EMNLP2020–2025 Limitations sections. Semantic segmentation(Gemini2.5Flash), hybrid deductive/inductive coding(Gemini3.1Pro), emerging-code consolidation/recommendations, then REQUIRED human approval of add/merge/expand/reject before codebook update. Evaluated on150papers with humanconsensus and qualitative metrics. It is not an autonomous RSI update rule or a language-model training algorithm; replacing mandatory human coding with toy automatic rules would alter the method.",
      "semantic_deduplication": "No exact record; existing RSI taxonomy/agent-memory updates are not this human-reviewed scholarly limitation-analysis study.",
      "cost_assessment": {
        "source": "v1 Section3 and AppendixK",
        "model_scale": "Gemini2.5Flash/Gemini3.1Pro plus human annotators",
        "data_scale": "ACL/EMNLP2020–2025 corpus and150paper evaluation",
        "available_resource": "one A100 or A30, or local Mac",
        "estimated_cost_basis": "Not a cost rejection: topic is scholarly meta-analysis, not repository's model/posttraining/agent algorithm reproduction scope. No API or corpus run attempted",
        "decision": "not_selected"
      },
      "decision": "not_selected",
      "reason": "Out-of-scope NLP research meta-analysis. Preserve review evidence; do not conflate human-approved qualitative codebook maintenance with autonomous agent RSI.",
      "review_done": true,
      "accepted_for_implementation": false,
      "implemented": false,
      "catalog_key": null,
      "domain": "outside_algorithm_scope",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "What Limits Us? Analyzing Self-Reported Limitations in NLP Research",
      "priority": "P1",
      "priority_basis": "本轮明确批准的历史线索审查优先级；不代表通过收录门槛或接收实现"
    },
    {
      "id": "2609.17764",
      "title": "How Calibration Content Shapes Attention-Based Reranking",
      "paper_url": "https://arxiv.org/abs/2609.17764v1",
      "published_at": "2026-09-15T19:12:58Z",
      "first_author": "Petros Karypis",
      "first_author_affiliation": "UC San Diego",
      "verified_priority_organizations": [
        "Google",
        "Google DeepMind"
      ],
      "priority_organization_evidence": "固定v1 PDF/正文一作邮箱pkarypis@ucsd.edu、脚注Google internship；Hossein Rajaby Faghihi/Yan Zhu为Google，Peter Chen/Rui Zhu/Noveen Sachdeva为Google DeepMind。",
      "reviewed_version_url": "https://arxiv.org/html/2609.17764v1",
      "version_note": "本地核对固定v1。",
      "sources": [
        {
          "url": "https://arxiv.org/abs/2609.17764v1",
          "kind": "primary-arxiv-identity-v1-date",
          "reviewed": true
        },
        {
          "url": "https://arxiv.org/html/2609.17764v1",
          "kind": "primary-method-resource-affiliation",
          "reviewed": true
        },
        {
          "url": "https://arxiv.org/pdf/2609.17764v1#page=1",
          "kind": "primary-pdf-affiliation",
          "reviewed": true
        }
      ],
      "method_review": "固定v1 Section3 Eq4–7：ΔN=s(empty,I)−s(empty,empty)，Sλ=s(q,I)−[λs(empty,I)+(1−λ)s(empty,empty)]；λ1标准ICR、λ0bare-null。调用方真实三pass提分数并保证doc identity对齐；λ只能validation选。论文oracle λ对全数据查询调参，不能当独立test结果。",
      "cost_assessment": {
        "source": "固定v1正文方法/实验及附录；https://arxiv.org/html/2609.17764v1",
        "available_resource": "单张NVIDIA A100或A30，或本地Mac",
        "paper_configuration": "Gemma3-4B/Qwen3-4B/Llama3.1-8B/Qwen3-14B，top100每doc130tokens、16Kcontext三prefill，多benchmark/5fold/λ网格/QRhead选择；单A10040GB配置。完整模型加载和广泛矩阵超短时组件预算；不宣称真实retrieval改善。",
        "estimated_cost_basis": "依据原文配置和完整比较矩阵判断，不是本地完整实验测量；无需用长试跑确认成本。",
        "decision": "成本过高，不复现（审查 DONE）；完整实验不下载、不试跑、不自动重试。仅独立低成本机制/协议诊断，不冒充完整模型或正式benchmark。"
      },
      "decision": "component_only",
      "reason": "将query+instruction、instruction-inclusive null、bare null三次读出分解，按λ插值应移除的instruction footprint；λ0仍是bare-null校准，不等于无校准。本地只实现对齐document分数的式4–7，不跑LLM注意力/正式排序benchmark，也不作为有线上AB的工业adapter。",
      "review_done": true,
      "accepted_for_implementation": true,
      "implemented": true,
      "catalog_key": "interpolated-null-calibration",
      "domain": "foundation-models",
      "local_operator": "auto_research.foundation_models.interpolated_null_calibration.interpolated_null_calibration",
      "local_module": "auto_research/foundation_models/interpolated_null_calibration.py",
      "device": "cuda",
      "upstream_code": {
        "url": null,
        "revision": null,
        "license": "未找到作者方法源码；论文CC BY4.0",
        "status": "固定v1及精确标题检索未找到作者方法仓库，源码SHA/许可证不可核验；不是把ICR/QRHeads基线仓库当本论文方法开源"
      },
      "semantic_deduplication": "9个规范arXiv ID与标题均未发现已有catalog实现；POLARIS写作方法与POLARIS-Project/Polaris-7B-Preview同名不同文，不合并。",
      "component_validation_status": "implementation_and_local_contracts_pending_root_device_binding",
      "full_experiment_status": "not_reproduced_high_cost",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "original_lead_title": "How Calibration Content Shapes Attention-Based Reranking",
      "component_implemented": true,
      "component_accepted": true,
      "implementation_status": "component_only",
      "component_evidence": "docs/foundation-models/2609.17764-interpolated-null-calibration/metrics/component-cuda-seeds42-44.json",
      "status": "implemented",
      "resolution": "bounded_l1_component_accepted",
      "gpu_validation_artifact": "docs/gpu-validations/interpolated-null-calibration-a100-20261025.json",
      "priority": "P1",
      "priority_basis": "采用已核查catalog优先级；优先级不等于实现或能力验收"
    }
  ],
  "limitations": [
    "Exactly the approved25 historical identities, not allGoogle publications or the four prior unresolved event-page links.",
    "High-cost review DONE does not mean implemented; accepted L1 components require separate frozen-source execution.",
    "Source exceptions, if any, are explicit archived dispositions, not successful full-text reviews.",
    "No pretrained capability comparison or evolve promotion; global review watermark is unchanged."
  ]
}