{
  "schema_version": 1,
  "reviewed_at": "2026-10-12",
  "domain": "post-training",
  "discovery_origin": "explicit_frozen_implementation_batch_supplement",
  "discovery_scope": "四篇来自用户明确冻结的实施清单，作为补充发现独立登记；不是八份查询的新召回，也不并入 fresh 578 的审查数量。",
  "counted_in_eight_query_retrieval": false,
  "counted_in_fresh_578": false,
  "review_scope": "实施阶段的官方原文方法/评测审查，加本次官方 arXiv v1 日期与缓存 PDF 哈希、精确页码复核；只完成有界相关原文审查，不声称逐页阅读全文。",
  "acceptance_scope": "accepted 表示纳入用户冻结批次及完成论文相关方法/实验审查；不表示当前源码已通过 GPU gate，不表示论文规模能力复现或稳定收益。",
  "summary": {
    "supplementary_records": 4,
    "abstracts_reviewed": 4,
    "bounded_primary_fulltext_reviews": 4,
    "accepted": 4,
    "full_articles_read": 0,
    "verified_pdf_sha256": 4,
    "verified_v1_submission_dates": 4,
    "fulltext_unresolved": 0,
    "unfinished_review_records": [],
    "new_implementation_claims": 0,
    "new_gpu_pass_claims": 0
  },
  "records": [
    {
      "arxiv_id": "2610.13481",
      "title": "Branching On-Policy Distillation",
      "adapter_key": "branch-opd",
      "published": "2026-10-09T12:13:34Z",
      "status": "accepted",
      "discovery_origin": "explicit_frozen_implementation_batch_supplement",
      "matched_queries": [],
      "verified_organization": "Tsinghua University / Peking University / SenseTime Research",
      "organization_evidence": "Official PDF v1, page 1 author affiliations",
      "source": {
        "url": "https://arxiv.org/pdf/2610.13481v1",
        "sha256": "501993b6c44418347f45a2cd280cf987308bb525dc1a90ee90855f0ffa61a042",
        "local_path": "tmp/pdfs/2610.13481.pdf",
        "local_pdf_sha256_matches": true,
        "pdf_pages": 33,
        "version_evidence": "PDF page 1 identifies arXiv:2610.13481v1 and 9 Oct 2026"
      },
      "date_proof": {
        "url": "https://arxiv.org/abs/2610.13481v1",
        "location": "Submission history, v1 entry",
        "v1_submission_utc": "2026-10-09T12:13:34Z",
        "observed_at": "2026-10-12",
        "basis": "independently read official arXiv submission history; not inferred from identifier or project README"
      },
      "reviewed_abstract": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "read_scope": "official abstract; PDF selected method/experiment pages and generation-cost appendix, not references or all appendices",
      "evidence_locations": [
        {"kind": "method", "heading": "4.1 Selecting Fork Points; 4.2 Conditional Resampling and Coverage", "pdf_pages": [4, 5], "equations": [6, 7, 8], "scope": "PDF pages 4 and 5 read in full; student uncertainty/disagreement ranking, frozen student-policy suffix sampling"},
        {"kind": "method", "heading": "4.3 Training Objective", "pdf_pages": [5], "equations": [9, 10, 11], "scope": "shared-prefix mask, token-mean PPO OPD objective; not teacher-forced branches or reward-selected suffixes"},
        {"kind": "method", "heading": "4.4 Protocol-Boundary Repair", "pdf_pages": [6], "equations": [12, 13], "characters_read": [0, 3500], "scope": "positive/negative protocol-event supervision and task-specific detector; empty constraints mean zero auxiliary loss"},
        {"kind": "evaluation", "heading": "5.1 Experimental Setup; 5.2 Main Results", "pdf_pages": [6, 7], "tables": [1, 2], "scope": "page 6 first 3500 characters and full page 7; mathematics avg@16 and single-trajectory agent evaluation, seven-rollout/PBR controls"},
        {"kind": "evaluation", "heading": "B.4 Generation and Environment Costs", "pdf_pages": [22], "tables": [11, 12], "scope": "full PDF page 22; measured tokens distinguished from position estimates and extra environment replay cost"}
      ],
      "method_evidence": "学生策略采样主轨迹；以 uncertainty 与 teacher-student disagreement 排序 fork，同一冻结学生独立采样共享前缀后的 suffix，前缀只计一次；教师对生成 token 评分。PBR 用协议事件正/负约束修复结束与重复行为，不以 gold reward 选分枝。",
      "evaluation_evidence": "Table 1 的数学宏平均为 BranchOPD 44.37%、Vanilla OPD+PBR 38.94%、七独立 rollout+PBR 39.09%；Table 2 ALFWorld ID/OOD 为 57.86%/55.22%。Table 11 的 19.1% ALFWorld、31.2% WebShop 节约是测得生成 token，数学/Search-QA 则是位置估算；Table 12 同时披露 replay 增加环境步骤。上述全为原文结果。",
      "paper_reported_results_only": true,
      "production_online_ab_claimed": false,
      "decision_rationale": "符合冻结批次的后训练/蒸馏核心算法，相关方法、对照与实际 token 成本已审；用户明确授权实施，不将离线 agent benchmark 冒称 production A/B。",
      "upstream_code_release_status_at_implementation_review": "作者链接 https://github.com/chendy25/BranchOPD，当时仅 LICENSE/README，未提供可用算法实现；不将仓库存在等同代码发布。",
      "local_implementation_boundary": "已实现有界数学单 turn 学生分枝、真实 teacher scoring、PBR 与 LoRA 更新；未实现 ALFWorld/WebShop 环境快照和多 turn 分枝，未复现原模型/数据/长预算结果。",
      "current_source_gpu_validation_claimed": false,
      "new_implementation_claimed": false
    },
    {
      "arxiv_id": "2610.13257",
      "title": "CAST: Credit Alignment for Cross-Tokenizer On-Policy Distillation",
      "adapter_key": "cast-opd",
      "published": "2026-10-09T09:03:37Z",
      "status": "accepted",
      "discovery_origin": "explicit_frozen_implementation_batch_supplement",
      "matched_queries": [],
      "verified_organization": "Taobao & Tmall Group of Alibaba / Singapore Management University",
      "organization_evidence": "Official PDF v1, page 1 author affiliations",
      "source": {
        "url": "https://arxiv.org/pdf/2610.13257v1",
        "sha256": "a930beee09025e2abb6a7c82ca412c2cdbdc7fd9aba16eedfcdee19af7eb6e85",
        "local_path": "tmp/pdfs/2610.13257.pdf",
        "local_pdf_sha256_matches": true,
        "pdf_pages": 22,
        "version_evidence": "PDF page 1 identifies arXiv:2610.13257v1 and 9 Oct 2026"
      },
      "date_proof": {
        "url": "https://arxiv.org/abs/2610.13257",
        "location": "Submission history, v1 entry",
        "v1_submission_utc": "2026-10-09T09:03:37Z",
        "observed_at": "2026-10-12",
        "basis": "versioned abs transport cache miss; unversioned official page independently retrieved and its exact v1 history read"
      },
      "reviewed_abstract": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "read_scope": "official abstract; full PDF pages 4-8 and page 18 relevant method/setup/results/selection protocol; not every appendix or qualitative example",
      "evidence_locations": [
        {"kind": "method", "heading": "3.1 Preliminaries", "pdf_pages": [4], "figure": 3, "scope": "teacher retokenization of the actual student-decoded text, alignment chunks, projected scores and separate six-token credit chunks"},
        {"kind": "method", "heading": "3.2 Uncertainty-Smoothness Credit Assignment", "pdf_pages": [5], "equations": [1, 2, 3, 4, 5, 6, 7], "scope": "full page; entropy times inverse local teacher-signal variation, stop-gradient weights and conserved chunk advantage"},
        {"kind": "method", "heading": "3.3 Selective Token Anchoring and Optimization", "pdf_pages": [6], "equations": [8, 9, 10, 11], "scope": "unaligned-token initial-policy KL, aligned PPO objective and independent multiple rollouts"},
        {"kind": "evaluation", "heading": "4.1 Experimental Setup; 4.2 Main Results; 4.3 Discussions", "pdf_pages": [6, 7, 8], "tables": [1, 2, 3, 4, 5, 6], "scope": "heterogeneous/shared tokenizer settings, matched uniform/shuffled credit, anchoring and rollout controls; three evaluated seed summaries"},
        {"kind": "evaluation_integrity", "heading": "Appendix C: Checkpoint selection; Evaluation protocol and run variability", "pdf_pages": [18], "table": 9, "characters_read": [0, 2983], "scope": "all six benchmarks used to choose highest-Avg-6 checkpoints/seed-checkpoint pairs; original paper explicitly warns optimistic selection"}
      ],
      "method_evidence": "先对实际解码文本做跨 tokenizer alignment，再将教师 span log probability 投影并构成 credit chunks；以 uncertainty×smoothness 分配 token credit，保持每 chunk 总 advantage；未对齐位置用初始学生 selective KL，而不是给无对应 token 编造教师分数。",
      "evaluation_evidence": "Table 1 异构 tokenizer Mistral/Llama Avg-6 为 47.22/43.87；Table 4 matched uniform control 为 45.07，CAST 为 47.22。Tables 5/6 给出三 seed 47.01±0.28 与 43.72±0.16，但 page 18 明确在六个评测 benchmark 上挑 checkpoint，部分主表又挑 seed-checkpoint 最大 Avg-6，不能把三 seed 解释为干净隔离 test 的稳定收益。",
      "paper_reported_results_only": true,
      "evaluation_limitations": ["原文 checkpoint/seed 选择使用报告中的 benchmark，不等于本地 train/validation-only selection 契约。", "部分任务分数下降，不能只以六项均值宣称所有能力提升。"],
      "production_online_ab_claimed": false,
      "decision_rationale": "跨 tokenizer 文本对齐后的 credit 与未对齐 KL 是明确核心机制，主节方法和控制实验已审，纳入用户冻结批次；保留原论文评测选择偏差，不移植其 test-selection 做法。",
      "upstream_code_release_status_at_implementation_review": "未找到原作者可用公开实现；本地不是作者代码移植。",
      "local_implementation_boundary": "已实现真实 heterogeneous tokenizer alignment、credit、initial-policy reference 与 LoRA 更新；规范 key cast-opd 区别于旧 Agent CAST，CLI cast 仅兼容别名。原旧-source A100 诊断不证明本次 namespace 源码已经执行；全对齐旧 run 也不证明未对齐 KL 已有真实 GPU 覆盖。",
      "current_source_gpu_validation_claimed": false,
      "new_implementation_claimed": false
    },
    {
      "arxiv_id": "2610.13521",
      "title": "Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation",
      "adapter_key": "road-mopd",
      "published": "2026-10-09T12:41:51Z",
      "status": "accepted",
      "discovery_origin": "explicit_frozen_implementation_batch_supplement",
      "matched_queries": [],
      "verified_organization": "Nanjing University / JD.com",
      "organization_evidence": "Official PDF v1, page 1 author affiliations",
      "source": {
        "url": "https://arxiv.org/pdf/2610.13521v1",
        "sha256": "464e401ddde486ea2ee908b07f4d359d4591e5136e2ba5f2dc02dbd7089443a6",
        "local_path": "tmp/pdfs/2610.13521.pdf",
        "local_pdf_sha256_matches": true,
        "pdf_pages": 30,
        "version_evidence": "PDF page 1 identifies arXiv:2610.13521v1 and 9 Oct 2026"
      },
      "date_proof": {
        "url": "https://arxiv.org/abs/2610.13521",
        "location": "Submission history, v1 entry",
        "v1_submission_utc": "2026-10-09T12:41:51Z",
        "observed_at": "2026-10-12",
        "basis": "versioned abs transport cache miss; independently retrieved unversioned official page and exact v1 history"
      },
      "reviewed_abstract": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "read_scope": "official abstract; full PDF pages 3-8 plus first 4000 characters of page 9 method/evaluation/ablation content; not all references/appendices",
      "evidence_locations": [
        {"kind": "method", "heading": "3.1 Problem Setup; 3.2 Cross-Teacher-Guided Credit Assignment", "pdf_pages": [3, 4, 5], "equations": [1, 2, 3, 4, 5], "figure": 2, "scope": "routed and non-routed teacher score the same student trajectory against a frozen reference; agreement modulates target token credit"},
        {"kind": "method", "heading": "3.3 Preventive Multi-Teacher Gradient Projection", "pdf_pages": [5, 6], "equations": [6, 7, 8, 9], "scope": "actual auxiliary OPD gradient on identical stored target-domain trajectory; per-parameter-tensor conflict projection with norm-retention floor"},
        {"kind": "method", "heading": "3.4 Overall Objective and Training", "pdf_pages": [6], "equations": [10, 11], "scope": "sequence mean then domain mean; only corrected target gradients are averaged/applied, auxiliary gradients are constraints rather than extra optimizer updates"},
        {"kind": "evaluation", "heading": "4.1 Experimental Setup; 4.2 Main Results", "pdf_pages": [6, 7, 8], "tables": [1, 2, 3], "scope": "Code/Math/Medical public benchmarks, Avg@4/32/1 protocols, two student sizes, low-conflict and three-domain controls"},
        {"kind": "evaluation", "heading": "4.3 Further Analysis", "pdf_pages": [8, 9], "table": 4, "figures": [3, 4], "page_9_characters_read": [0, 4000], "scope": "same-trajectory versus PCGrad compatibility, credit/projection/adaptive-retention ablations; local gradient statements are not capability guarantees"}
      ],
      "method_evidence": "非路由教师作为同轨迹 critic，不额外生成辅助域 rollout；相对于共享冻结 reference 的一致性调制 token credit。目标梯度与真实 auxiliary OPD 梯度逐 tensor 检查冲突，尽量投影但保留目标范数下限；只有修正后的目标域梯度进入 AdamW。",
      "evaluation_evidence": "Table 1 Qwen3-4B Math×Medical 两域平均 Vanilla MOPD 42.71%、Road 47.92%，Medical 平均39.45%→48.23%。Tables 2/3 分别报告低冲突 Math×Code 与三域结果，Table 4 区分 credit、gradient protection 和 adaptive retention；更小学生的部分 Math/Code 分数仍低于 ExOPD，不能声称所有域都最优。",
      "paper_reported_results_only": true,
      "production_online_ab_claimed": false,
      "decision_rationale": "多教师 token credit 加同轨迹 preventive gradient projection 两部分均有原文方法与对照证据，纳入冻结批次；投影数学性质不充当生产/能力收益证明。",
      "upstream_code_release_status_at_implementation_review": "未找到该论文自己的可用作者实现；公共 baseline/dependency 链接不等于作者开源。",
      "local_implementation_boundary": "已实现两域、不同 teacher、真实 per-tensor auxiliary-gradient projection 与 LoRA 更新；GSM monetary/nonmonetary 子域仅 routing 诊断，不是原论文 Math×Medical。Unicode JSONL 运输修复不改变问题/答案/分区；实际失败与重跑由 root 独立归档。",
      "current_source_gpu_validation_claimed": false,
      "new_implementation_claimed": false
    },
    {
      "arxiv_id": "2610.13285",
      "title": "BoundaryEvo: Data-Model Co-Evolution at the Capability Boundary for Long-Chain Reasoning Distillation",
      "adapter_key": "boundary-evo",
      "published": "2026-10-09T09:37:33Z",
      "status": "accepted",
      "discovery_origin": "explicit_frozen_implementation_batch_supplement",
      "matched_queries": [],
      "verified_organization": "Shanghai Artificial Intelligence Laboratory / Nanjing University / Tongji University / The Chinese University of Hong Kong",
      "organization_evidence": "Official PDF v1, page 1 author affiliations",
      "source": {
        "url": "https://arxiv.org/pdf/2610.13285v1",
        "sha256": "d5e5f36ff36642a787638f484af3f54a4d25cc25db85c5bbde082fb6c25230a0",
        "local_path": "tmp/pdfs/2610.13285.pdf",
        "local_pdf_sha256_matches": true,
        "pdf_pages": 29,
        "version_evidence": "PDF page 1 identifies arXiv:2610.13285v1 and 9 Oct 2026"
      },
      "date_proof": {
        "url": "https://arxiv.org/abs/2610.13285",
        "location": "Submission history, v1 entry",
        "v1_submission_utc": "2026-10-09T09:37:33Z",
        "observed_at": "2026-10-12",
        "basis": "versioned abs transport cache miss; independently read exact v1 history on retrieved official unversioned page"
      },
      "reviewed_abstract": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "read_scope": "official abstract; PDF pages 3-8 method/main experimental sections and F.1-F.3 raw prompt/threshold evidence on pages 26-27; not every appendix or reference",
      "evidence_locations": [
        {"kind": "method", "heading": "3.1 Stage 1: Judge-Free Direct Distillation for Capability Priming", "pdf_pages": [3, 4], "equations": [1], "scope": "teacher behavior after basic validity checks, without correctness/capability-gap filtering"},
        {"kind": "method", "heading": "3.2 Stage 2: Dynamic Capability-Boundary Co-Evolution", "pdf_pages": [4, 5], "equations": [2, 3, 4, 5, 6, 7, 8, 9], "figure": 2, "scope": "independent current-student/fixed-expert rollouts, analyzer quality/validity, rechecked replay plus novel acquisition, fresh multiple teacher trajectories then independent verification"},
        {"kind": "method", "heading": "3.3 Training Objective", "pdf_pages": [5], "equations": [10], "scope": "response-token-normalized SFT after priming or verified boundary selection"},
        {"kind": "evaluation", "heading": "4.1-4.6 Experiments", "pdf_pages": [6, 7, 8], "tables": [1, 2, 3], "scope": "same-priming/1K-budget random and LLM2LLM controls, three turns, 11 benchmarks; Qwen3.5-2B gains need not be monotonic"},
        {"kind": "implementation_protocol", "heading": "F.1 Stage 1 Teacher Prompt; F.2 Student/Expert Prompt; F.3 Capability-Boundary Analyzer Prompt", "pdf_pages": [26, 27], "scope": "full selected pages: no gold to solvers, reference-aware external analyzer, strict JSON and score thresholds; reference-correctness/leakage hard rejection switches disabled in original reported route"}
      ],
      "method_evidence": "先少量 judge-free capability priming，再以当前学生和固定 expert 的能力边界更新训练数据。每轮重验 replay、获取 novel QAs；选择后再生成至少两个 fresh teacher trajectories，并对最终答案独立验证后做 response-only SFT。reference 只属于训练侧外部 analyzer/verification，不输入被评估 solver。",
      "evaluation_evidence": "Qwen3.5-4B 使用100 priming+3×300 evolution共1K；Tables 1/2 IMO-AnswerBench 62.9%→75.3%、AMO-Bench26.8%→46.0%。随机/LLM2LLM沿同一priming checkpoint与预算比较，Table 3 给出2B/Gemma迁移且2B Turn2优于Turn3的非单调现象。原训练依赖DeepSeek/GPT/Gemini多角色，不等于本地小公开模型已重现同样数据质量或收益。",
      "paper_reported_results_only": true,
      "production_online_ab_claimed": false,
      "decision_rationale": "能力 priming 后随学生变化而重估边界和训练分布是明确核心机制，相关方法/控制实验及原 prompt 契约已读；纳入冻结实施，不以接口/解析测试代替真实 co-evolution。",
      "upstream_code_release_status_at_implementation_review": "原文承诺未来公开代码与数据，本次实施审查未找到可用作者仓库。",
      "local_implementation_boundary": "已实现真实 priming、每轮 solver/analyzer、边界选择、fresh teacher 多rollout、验证与SFT路径；analyzer解析失败或没有Stage2参数更新必须保留负结果，不称作有效边界演化验证。当前 GPU结果及任何协议诊断由 root 独立验证。",
      "current_source_gpu_validation_claimed": false,
      "new_implementation_claimed": false
    }
  ]
}