{
  "schema_version": "widedeep_public_related_work_v1",
  "verified_on": "2026-10-31",
  "status": "primary_sources_checked",
  "positioning": {
    "headline": "针对具体能力扩展错题、长期调整难度，并把成功参考带回原题验证局部纠错。",
    "wide": "只从出现失败的题出发，扩展仍考同一具体能力的相似题；全对题不作为 Wide 来源。",
    "deep": "全错时寻找较易练习，全对时继续挑战；保持同一能力的题目关系与长期记录，简单题成功不等于原难题解决。",
    "combined": "同一轮内可以同时安排 Wide 与 Deep，而不是把两者拆成先后独立的进化阶段。",
    "correction": "成功参考用于解释和修正原题失败；回到原材料、原规则执行成功后，只学习与目标有关的改正动作。",
    "established_ideas": [
      "自生成材料后继续训练、借助反馈纠错、保留失败记忆、选择局部学习位置，都已有直接先例。",
      "相关工作并非都只关注当前错题，也不能统称为没有 Deep；本项目的术语是自己的实验划分。",
      "现候选没有完整复现 HINT-SD、AgentRefine 或 Agent-R，借鉴部分思想不等于实现论文全部方法。"
    ],
    "hypothesis_to_test": "同一具体能力的横向扩展与纵向长期追踪，能否共同提高有效学习材料的产出，并使原难题和未见任务在相同预算下受益。",
    "evidence_needed": [
      "用相同起点、环境、学习方式、题量和总成本，分开比较 Wide、Deep、两者组合及加入纠错。",
      "同时检查改正材料是否真实成功、原难题能否无帮助完成、已会任务是否保持、未见任务是否改善。",
      "与可靠的成功材料自训练、语言反思及直接相关纠错方法比较，并运行独立训练种子。",
      "学生完成分析、出题和纠错的版本与老师参与的版本分开报告。"
    ],
    "current_evidence": "这是文献和方法定位核对，不是新增 GPU 实验结果，也不支持首次提出或已超越已有方法的结论。"
  },
  "papers": [
    {
      "id": "hint_sd",
      "short_name": "HINT-SD",
      "title": "HINT-SD: Targeted Hindsight Self-Distillation for Long-Horizon Agents",
      "authors": [
        "Woongyeong Yeo",
        "Yumin Choi",
        "Taekyung Ki",
        "Sung Ju Hwang"
      ],
      "year": 2026,
      "first_submitted": "2026-05-18",
      "last_revised": "2026-10-02",
      "publication_note": "arXiv 摘要页标注 EMNLP Findings 2026。",
      "urls": {
        "abstract": "https://arxiv.org/abs/2605.17873",
        "full_text": "https://arxiv.org/html/2605.17873v3",
        "code": "https://github.com/wgcyeo/HINT-SD"
      },
      "verification": {
        "status": "verified_live_primary_sources",
        "read_scope": "摘要页；全文第 2、3 节及附录 A.2。"
      },
      "question": "任务最后失败了，哪些早先动作才值得纠正？",
      "method_points": [
        "回看整条失败过程，由模型选出与失败相关的少数动作，并分别给出反馈。",
        "看到反馈的同规模模型指导只看到正常历史的模型；训练集中在选中的动作片段，不给每一步都生成反馈。",
        "正文把模型自身作为分析者和指导者；附录实际使用从学生初始化、缓慢随学生更新的指导模型。",
        "它不以先修成一条完整成功轨迹为必要条件，因此能从失败过程获得额外训练信号；反馈本身仍可能出错。"
      ],
      "experiments": {
        "environments": [
          "BFCL v3 的多轮 Base、Long Context 类别",
          "AppWorld"
        ],
        "main_model": "Qwen3-4B-Instruct-2507",
        "setup": "每题 4 次作答，15 个训练周期；每条失败过程最多选 3 个位置；报告 Avg@4 和 Best@4。",
        "reported_findings": [
          {
            "metric": "BFCL v3 Avg@4",
            "method": "HINT-SD-Multi",
            "value": 41.88,
            "comparison": "GRPO",
            "comparison_value": 31.56,
            "unit": "百分制分数",
            "source": "https://arxiv.org/html/2605.17873v3#S3.T1"
          },
          {
            "metric": "AppWorld Avg@4",
            "method": "HINT-SD-Multi",
            "value": 18.46,
            "comparison": "SDPO",
            "comparison_value": 9.74,
            "unit": "百分制分数",
            "source": "https://arxiv.org/html/2605.17873v3#S3.T1"
          }
        ],
        "efficiency_note": "论文在自身设置中报告每个训练步 84.76 秒降至 37.45 秒；不能拿这个比例估算本项目的耗时。"
      },
      "comparison_to_project": {
        "useful_prior_idea": "把反馈集中到与失败有关的少数动作，避免整段无差别学习和逐步生成反馈。",
        "present_difference": "本项目当前候选先把修正路线放回原失败题执行，整题成功后才学习相关改正动作；并未照搬论文的反馈条件化分布学习。",
        "unanswered_by_this_paper": "它的实验没有直接回答：在同一具体能力的难易题目之间长期追踪和回考，是否能稳定提高这类纠错信号的产出。",
        "public_summary": "借鉴“找准该改的动作”，再检验同能力难易题能否提供更有用的修正参考。"
      },
      "evidence": [
        {
          "claim": "完整轨迹定位、只在选定动作上训练",
          "url": "https://arxiv.org/html/2605.17873v3#S2",
          "locator": "Section 2, Hindsight Feedback Generation / Targeted Self-Distillation"
        },
        {
          "claim": "同规模指导模型随学生更新",
          "url": "https://arxiv.org/html/2605.17873v3#A1.SS2",
          "locator": "Appendix A.2, teacher initialization and EMA update rate"
        },
        {
          "claim": "模型、任务与主要成绩",
          "url": "https://arxiv.org/html/2605.17873v3#S3",
          "locator": "Sections 3.1–3.2, Table 1"
        }
      ]
    },
    {
      "id": "agent_refine",
      "short_name": "AgentRefine",
      "title": "AgentRefine: Enhancing Agent Generalization through Refinement Tuning",
      "authors": [
        "Dayuan Fu",
        "Keqing He",
        "Yejie Wang",
        "Wentao Hong",
        "Zhuoma Gongque",
        "Weihao Zeng",
        "Wei Wang",
        "Jingang Wang",
        "Xunliang Cai",
        "Weiran Xu"
      ],
      "year": 2025,
      "first_submitted": "2025-01-03",
      "last_revised": "2025-02-24",
      "publication_note": "arXiv 摘要页标注 ICLR 2025。",
      "urls": {
        "abstract": "https://arxiv.org/abs/2501.01702",
        "full_text": "https://arxiv.org/html/2501.01702v2"
      },
      "verification": {
        "status": "verified_live_primary_sources",
        "read_scope": "摘要页；全文第 2、3、4 节，尤其 3.3 的训练目标和表 2 的删减对照。"
      },
      "question": "模型是在记住熟悉环境里的答案，还是能根据新反馈改正错误？",
      "method_points": [
        "用较强模型构造多样化环境、任务以及带错误和修正的交互过程，再通过规则检查筛选。",
        "把错误动作及其反馈保留在后续上下文里，但错误轮次本身不作为学习目标。",
        "学习其余正确轮次的思考和动作，包括根据反馈做出的修正；并不是只学习与旧答案文字不同的动作。",
        "论文将这种修正学习与跨环境表现、面对动作表述变化时的表现联系起来。"
      ],
      "experiments": {
        "environments": [
          "ScienceWorld",
          "ALFWorld",
          "BabyAI",
          "PDDL",
          "Jericho"
        ],
        "models": [
          "Llama 3 系列",
          "Mistral-v0.3"
        ],
        "data_generation_model": "主要使用 GPT-4o-2024-05-13；另有 DeepSeek-v2.5 数据生成实验。",
        "reported_findings": [
          "在未用于收集各自训练数据的任务上，与 Agent-FLAN、AgentGym、AgentGen 等进行比较。",
          "表 2 在 8000 条数据设置下比较去掉修正学习、去掉带修正的样本、以及连错误轮次一起学习；这些处理都损害论文中的跨任务表现。",
          "另做 ALFWorld 动作表述扰动和多次尝试实验。不同基线的训练环境不同，表中已区分见过与未见过的环境。"
        ]
      },
      "comparison_to_project": {
        "useful_prior_idea": "保留错误及反馈作为上下文，避免把错误输出当正确目标学习。",
        "present_difference": "本项目当前侧重原题执行成功后的局部改正动作。撤回错误前重写，不自动等于在错误已经发生后还能补救；两种能力需要分别测试。",
        "unanswered_by_this_paper": "多样化修正数据的收益，不能直接替代同能力难度谱系长期调度的收益验证。",
        "public_summary": "借鉴“错误留作上下文、正确修正用于学习”，同时明确避免犯错与现场补救的区别。"
      },
      "evidence": [
        {
          "claim": "数据生成与验证",
          "url": "https://arxiv.org/html/2501.01702v2#S3.SS1",
          "locator": "Sections 3.1–3.2"
        },
        {
          "claim": "错误轮次不参与损失，正确轮次参与",
          "url": "https://arxiv.org/html/2501.01702v2#S3.SS3",
          "locator": "Section 3.3, Equation 1"
        },
        {
          "claim": "五个环境及修正学习删减对照",
          "url": "https://arxiv.org/html/2501.01702v2#S4",
          "locator": "Section 4.1 and Table 2"
        }
      ]
    },
    {
      "id": "agent_r",
      "short_name": "Agent-R",
      "title": "Agent-R: Training Language Model Agents to Reflect via Iterative Self-Training",
      "authors": [
        "Siyu Yuan",
        "Zehui Chen",
        "Zhiheng Xi",
        "Junjie Ye",
        "Zhengyin Du",
        "Jiecao Chen"
      ],
      "year": 2025,
      "first_submitted": "2025-01-20",
      "last_revised": "2025-03-24",
      "urls": {
        "abstract": "https://arxiv.org/abs/2501.11425",
        "full_text": "https://arxiv.org/html/2501.11425"
      },
      "verification": {
        "status": "verified_live_primary_sources",
        "read_scope": "摘要页当前 v3；实际读取无版本号 HTML 的第 3、4 节和表 2。"
      },
      "question": "能否让模型在出错时及时转回正确路线，而非一直重复错误？",
      "method_points": [
        "用树搜索从同一前缀探索不同路线，收集质量较差与较好的分支。",
        "让当前模型在失败路线中找出它能够识别的第一个错误位置，构造带反思衔接和正确后续的训练材料。",
        "修正材料与较好路线、通用对话数据混合训练；随着模型更新，重新生成修正材料。",
        "论文既比较直接在失败末尾接正确答案，也比较由模型选择更及时的转折位置。"
      ],
      "experiments": {
        "environments": [
          "WebShop",
          "ScienceWorld",
          "TextCraft"
        ],
        "main_model": "Llama-3.1-8B-Instruct",
        "setup": "3 次自训练迭代；用于搜索的任务分别为 300、200、200；测试任务分别为 200、200、100。搜索每层 4 个候选、深度 20，每次估计采 8 份作答；最终评测交互上限为 100 轮。",
        "reported_findings": [
          {
            "metric": "TextCraft 成功率",
            "method": "Agent-R",
            "value": 78,
            "comparison": "同表 ETO",
            "comparison_value": 75,
            "unit": "百分比",
            "source": "https://arxiv.org/html/2501.11425#S4.SS2.SSS0.Px5.tab1.1"
          },
          {
            "metric": "三个环境的论文平均分",
            "method": "Agent-R",
            "value": 70.71,
            "comparison": "同表 ETO",
            "comparison_value": 65.12,
            "unit": "分",
            "source": "https://arxiv.org/html/2501.11425#S4.SS2.SSS0.Px5.tab1.1"
          }
        ],
        "metric_note": "WebShop、ScienceWorld 使用平均最终奖励，TextCraft 使用成功率；三个环境的平均分不是统一的成功率。"
      },
      "comparison_to_project": {
        "useful_prior_idea": "在可对应的失败与成功路线间构造修正，并随当前模型能力重新收集。",
        "present_difference": "本项目当前不进行这种完整树搜索；它尝试从同能力较易题的成功参考出发，在原失败题条件下重新验证修正。",
        "unanswered_by_this_paper": "它是 TextCraft 上很直接的比较对象，但不同模型、题量、交互上限和搜索成本必须统一后再比较。",
        "public_summary": "它已在 TextCraft 验证过从错误路线学习修正，适合成为本项目的重要对照。"
      },
      "evidence": [
        {
          "claim": "树搜索和由当前模型选择修正位置",
          "url": "https://arxiv.org/html/2501.11425#S3.SS1",
          "locator": "Section 3.1"
        },
        {
          "claim": "修正、正确路线、通用数据混合并迭代更新",
          "url": "https://arxiv.org/html/2501.11425#S3.SS2",
          "locator": "Section 3.2, Equation 6"
        },
        {
          "claim": "数据规模、100 轮上限、三轮迭代及成绩",
          "url": "https://arxiv.org/html/2501.11425#S4.SS2",
          "locator": "Section 4.2, Table 2"
        }
      ]
    },
    {
      "id": "star",
      "short_name": "STaR",
      "title": "STaR: Bootstrapping Reasoning With Reasoning",
      "authors": [
        "Eric Zelikman",
        "Yuhuai Wu",
        "Jesse Mu",
        "Noah D. Goodman"
      ],
      "year": 2022,
      "first_submitted": "2022-03-28",
      "last_revised": "2022-05-20",
      "urls": {
        "abstract": "https://arxiv.org/abs/2203.14465",
        "full_text": "https://arxiv.org/html/2203.14465v2"
      },
      "verification": {
        "status": "verified_live_primary_sources",
        "read_scope": "摘要页；全文引言、方法和实验任务说明。"
      },
      "question": "能否用少量示例，让模型用自己生成的正确推理继续改进？",
      "method_points": [
        "先让模型自己生成推理和答案，保留答案正确的推理。",
        "对答错的问题，给出已知正确答案，让模型再生成能够解释它的推理，再用于学习。",
        "重复生成和训练；依赖任务的正确答案，并不等于没有外部正确性依据。"
      ],
      "experiments": {
        "environments": [
          "算术",
          "CommonsenseQA",
          "GSM8K"
        ],
        "reported_findings": [
          "论文在算术、常识问答与数学文字题上报告自生成推理的迭代训练收益。"
        ]
      },
      "comparison_to_project": {
        "useful_prior_idea": "让模型自己产生学习材料，利用正确性筛选逐步扩充可学内容。",
        "present_difference": "当前项目的任务是多步环境交互，正确性由环境执行决定，重点是同能力新题、难度变化和回到原题修正。",
        "unanswered_by_this_paper": "已有“错误后借助额外信息再生成正确材料”的先例，因此不能把这一思路单独当作本项目新贡献。",
        "public_summary": "自生成正确材料再学习，是项目沿用的基础思想。"
      },
      "evidence": [
        {
          "claim": "生成、按正确答案再解释、筛选并反复训练",
          "url": "https://arxiv.org/abs/2203.14465",
          "locator": "Abstract"
        },
        {
          "claim": "算术、CommonsenseQA 与 GSM8K 实验",
          "url": "https://arxiv.org/html/2203.14465v2#S4",
          "locator": "Section 4"
        }
      ]
    },
    {
      "id": "reflexion",
      "short_name": "Reflexion",
      "title": "Reflexion: Language Agents with Verbal Reinforcement Learning",
      "authors": [
        "Noah Shinn",
        "Federico Cassano",
        "Edward Berman",
        "Ashwin Gopinath",
        "Karthik Narasimhan",
        "Shunyu Yao"
      ],
      "year": 2023,
      "first_submitted": "2023-03-20",
      "last_revised": "2023-10-10",
      "urls": {
        "abstract": "https://arxiv.org/abs/2303.11366",
        "full_text": "https://arxiv.org/html/2303.11366v4"
      },
      "verification": {
        "status": "verified_live_primary_sources",
        "read_scope": "摘要页；全文第 3、4 节。"
      },
      "question": "不重新训练模型，能否记住失败教训并在下一次尝试改正？",
      "method_points": [
        "由任务反馈生成语言反思，并保存在记忆中，供后续尝试使用。",
        "核心机制不更新模型权重；改进来自反馈、反思记忆和再次尝试。",
        "包含行动者、评价者和反思者；不同任务使用不同反馈来源。"
      ],
      "experiments": {
        "environments": [
          "ALFWorld",
          "HotpotQA",
          "HumanEval",
          "MBPP 及其他代码任务"
        ],
        "reported_findings": [
          "在决策、检索推理和编程任务上比较带反思记忆的重试与原基线。"
        ],
        "budget_note": "使用反馈后再尝试的结果，不能当成没有反馈、没有额外尝试的单次作答成绩。"
      },
      "comparison_to_project": {
        "useful_prior_idea": "失败记录应帮助后续决策，而不是每次从头遗忘。",
        "present_difference": "本项目还要把经过原题验证的改正动作训练进模型，并验证不再提供参考时是否仍会做。",
        "unanswered_by_this_paper": "记住语言反思与沿同能力难度谱系长期训练是不同机制，应通过同预算对照分别测量。",
        "public_summary": "反思记忆能帮助再次尝试；本项目进一步检验能否把可靠改正学进模型。"
      },
      "evidence": [
        {
          "claim": "不更新权重，而以语言反馈和记忆改进",
          "url": "https://arxiv.org/abs/2303.11366",
          "locator": "Abstract"
        },
        {
          "claim": "三组件和记忆机制",
          "url": "https://arxiv.org/html/2303.11366v4#S3",
          "locator": "Section 3"
        },
        {
          "claim": "决策、问答、编程实验",
          "url": "https://arxiv.org/html/2303.11366v4#S4",
          "locator": "Section 4"
        }
      ]
    },
    {
      "id": "rest",
      "short_name": "ReST",
      "title": "Reinforced Self-Training (ReST) for Language Modeling",
      "authors": [
        "Caglar Gulcehre",
        "Tom Le Paine",
        "Srivatsan Srinivasan",
        "Ksenia Konyushkova",
        "Lotte Weerts",
        "Abhishek Sharma",
        "Aditya Siddhant",
        "Alex Ahern",
        "Miaosen Wang",
        "Chenjie Gu",
        "Wolfgang Macherey",
        "Arnaud Doucet",
        "Orhan Firat",
        "Nando de Freitas"
      ],
      "year": 2023,
      "first_submitted": "2023-08-17",
      "last_revised": "2023-08-21",
      "urls": {
        "abstract": "https://arxiv.org/abs/2308.08998",
        "full_text": "https://arxiv.org/html/2308.08998v2"
      },
      "verification": {
        "status": "verified_live_primary_sources",
        "read_scope": "摘要页；全文引言、第 3 节方法和实验任务说明。"
      },
      "question": "怎样重复利用模型生成的较好答案，让自训练成本更可控？",
      "method_points": [
        "生成阶段用当前模型扩充答案；改进阶段按评分筛选数据，再反复学习这批数据。",
        "改进后的模型再生成下一批数据，生成和训练分开进行。",
        "原论文实验的评分主要来自学习得到的偏好评分模型，不是 TextCraft 环境的二元成功判定。"
      ],
      "experiments": {
        "environments": [
          "IWSLT 2014 德语到英语翻译",
          "WMT 2020 中文到英语翻译",
          "内部 Web Domain 翻译任务"
        ],
        "reported_findings": [
          "论文报告自动评分和人工翻译质量的改善，并研究同批数据多次学习的成本收益。"
        ]
      },
      "comparison_to_project": {
        "useful_prior_idea": "批量收集、可靠筛选、统一学习，以及对实际生成和学习成本分别记账。",
        "present_difference": "ReST 的基础循环不是本项目定义的“针对错题生成相似新题”；它不能被简单改名为 Wide。",
        "unanswered_by_this_paper": "需要分别比较普通成功材料自训练、错误扩展、长期难度追踪和局部纠错，而不是把所有自训练都归入同一类。",
        "public_summary": "批量生成、筛选和学习已有成熟先例；项目新增的研究问题在于如何持续选择值得学的题。"
      },
      "evidence": [
        {
          "claim": "Grow / Improve 的分工及反复利用数据",
          "url": "https://arxiv.org/html/2308.08998v2#S3",
          "locator": "Section 3 and Figure 1"
        },
        {
          "claim": "翻译任务与评测",
          "url": "https://arxiv.org/html/2308.08998v2#S4",
          "locator": "Section 4"
        }
      ]
    }
  ]
}