[ { "id": "2609.14858", "title": "Dream-RSI: Recursive Self-Improvement through Evolving Worlds", "date": "2026-09-14", "dateLabel": "首次提交", "category": "methods", "methodType": "harness", "priority": "C", "contentType": "paper", "url": "https://arxiv.org/abs/2609.14858", "references": { "main": { "label": "arXiv 摘要与首次提交日期", "url": "https://arxiv.org/abs/2609.14858" }, "intro": { "label": "§1 Introduction", "url": "https://arxiv.org/html/2609.14858#S1" }, "method": { "label": "§3 探索策略、回放与选版本", "url": "https://arxiv.org/html/2609.14858#S3" }, "exp": { "label": "§4 Experiments", "url": "https://arxiv.org/html/2609.14858#S4" }, "appendix": { "label": "附录 A 任务定义及数值检查", "url": "https://arxiv.org/html/2609.14858#A1" }, "prompt": { "label": "附录 B.2 策略修改提示", "url": "https://arxiv.org/html/2609.14858#A2.SS2" } }, "tldr": { "gap": { "value": "固定搜索策略不会随经验改善；直接在线试验新策略又需要大量完整搜索才能知道好坏，导致改进搜索策略本身非常昂贵。", "refs": [ "intro" ] }, "position": { "value": "研究能否利用已有发现过程，低成本改进“下一轮怎样探索”的策略，而不更换或训练底层 coding agent。", "refs": [ "intro", "method" ] }, "conclusion": { "value": "在算法、数学优化和 GPU kernel 三类任务中，以较少发现成本达到相当或更好的结果。Lasso 另有未参与搜索的数据集测试；其他任务主要衡量同一问题上的搜索质量。", "refs": [ "exp" ] } }, "dimensions": { "object": { "value": "改可执行的探索策略代码：选哪些分支、开多少并行工作区、每条分支继续多久、何时停止。生成的解程序也会更新；底层 coding agent、模型参数及判分器固定。", "refs": [ "method" ] }, "executor": { "value": "Lasso：Gemini-3.1 Pro 或 Gemini-3.7-Flash，经 Gemini CLI 执行代码搜索；数学与 KernelBench：Gemini-3.1 Pro。", "refs": [ "exp" ] }, "modifier": { "value": "固定的策略开发 agent 阅读历史回放分数与策略源码，再改探索策略。正文单列了这一角色，但没有单独给出其模型型号，不将执行模型型号自动当作修改者型号。", "refs": [ "method", "prompt" ] }, "verdict": { "value": "Lasso：对照 sklearn 检查目标函数误差，通过后按运行时间评分；数学：程序计算集合、圆布局或自相关的目标值与约束；KernelBench：与参考实现比数值结果，通过后测速度。策略候选在历史树上回放，直接复用已保存的这些结果。", "refs": [ "method", "exp", "appendix" ] }, "seed": { "value": "Gemini CLI 外加工作区调度层，支持并行生成、逐步修订和执行评测。初始探索策略固定分支数与步数；Dream-RSI 将这层调度代码开放给修改者。", "refs": [ "method", "exp" ] }, "cycle": { "value": "真实搜索并记录发现树 → 把树变成可回放的历史环境 → agent 修改策略并在同一历史池比较候选 → 部署最高分策略 → 新搜索继续扩充历史池。回放不能评价历史中从未出现的新分支。", "refs": [ "method" ] }, "train": { "value": "不训练模型参数。Lasso 使用 SimpleTES 的 17 个合成实例搜索；数学搜索 Sum–Difference、Circle Packing、Autocorrelation 三题；GPU 搜索 KernelBench 的 VGG16、LayerNorm、ConvDiv、ConvMax 四项。", "refs": [ "exp" ] }, "debug": { "value": "策略选版本使用已完成搜索生成的发现树，候选在同一历史池回放比较。Lasso 的正确性检查实例与计时实例另行生成、互不相同。", "refs": [ "method", "appendix" ] }, "test": { "value": "Lasso 另测 Gisette、RCV1、DNA、Leukemia、Colon、Duke Breast 六个留出数据集。数学三题与 KernelBench 四项报告各自搜索所得解的质量和成本，不是另一个未见任务集。", "refs": [ "exp" ] }, "isolation": { "value": "Lasso 的留出结果支持解程序的跨数据泛化；不能据此声称探索策略已通过独立跨任务测试。历史回放用于选策略，不能同时当成独立测试。", "refs": [ "method", "exp" ] }, "novelty": { "value": "将历史发现树变成可运行的策略试验环境，反复比较调度代码；不同于只把历史轨迹总结成提示词或记忆。", "refs": [ "intro", "method" ] } }, "tags": [ "HarnessCode", "Improver", "Workflow", "OfflineSearch" ], "links": [ { "label": "代码", "url": "https://github.com/zhengkid/Dream-RSI" }, { "label": "项目", "url": "https://dream-rsi.com/" } ], "visibleKeys": [ "object", "executor", "modifier", "verdict", "seed" ], "reviewedAt": "2026-09-15", "attributions": [ { "tag": "org:google-deepmind", "label": "Google DeepMind", "kind": "institution", "sources": [ { "label": "论文作者机构署名", "url": "https://arxiv.org/html/2609.14858" } ] } ], "priorityBasis": "Google DeepMind 参与署名；直接研究探索策略自改进，列为重点代表作。" }, { "id": "2608.26530", "title": "PILOT in the Loop: Live Self-Improvement for Long-Horizon Agents", "date": "2026-08-27", "dateLabel": "首次提交", "category": "methods", "methodType": "harness", "priority": "K", "contentType": "paper", "url": "https://arxiv.org/abs/2608.26530", "references": { "main": { "label": "arXiv 摘要与首次提交日期", "url": "https://arxiv.org/abs/2608.26530" }, "intro": { "label": "§1 Introduction", "url": "https://arxiv.org/html/2608.26530#S1" }, "method": { "label": "§2 双向监督与 Pi 实现", "url": "https://arxiv.org/html/2608.26530#S2" }, "setup": { "label": "§3.1 实验协议", "url": "https://arxiv.org/html/2608.26530#S3.SS1" }, "result": { "label": "§3.2–3.3 结果", "url": "https://arxiv.org/html/2608.26530#S3.SS2" }, "skills": { "label": "附录 A.3 经验保存指令", "url": "https://arxiv.org/html/2608.26530#A1.SS3" } }, "tldr": { "gap": { "value": "事后反思只能帮助后续任务，救不了仍在进行的失败；单 agent 又要执行又要诊断，而普通委派往往要等 worker 结束才能干预。", "refs": [ "intro" ] }, "position": { "value": "研究把任务执行与监督分开，能否在运行过程中纠偏，并把当场发现的经验变成以后可用的技能。", "refs": [ "intro" ] }, "conclusion": { "value": "两个冻结模型、三个 benchmark 的六组组合中五组最好；重复 Terminal-Bench 2.0 的实验也有提升，但使用同一批题反复进化，不能解释为未见任务的泛化提升。", "refs": [ "setup", "result" ] } }, "dimensions": { "object": { "value": "持久变化的是技能文件和长期记忆；运行中的纠偏消息影响当前 worker。模型参数不变，论文没有让 agent 重写整个 Pi runtime。", "refs": [ "method", "skills" ] }, "executor": { "value": "分别用 GLM-5.1、Kimi-K2.6；每组 supervisor 与 worker 都使用该组同一个冻结模型，运行在独立会话上下文中。", "refs": [ "method", "setup" ] }, "modifier": { "value": "supervisor 查看 worker 的通知、问题和必要的轨迹，发送纠偏消息、终止无效 worker，并写入可复用技能或记忆。", "refs": [ "method" ] }, "verdict": { "value": "运行中只看工具输出、错误、进度及轨迹，不给 benchmark 分数。每轮结束后用 Terminal-Bench verifier 的成功／失败决定保留哪些运行产生的技能；SWE-bench 两组使用任务测试评价修复结果。", "refs": [ "setup", "result" ] }, "seed": { "value": "在 Pi coding-agent runtime 上增加 supervisor–worker 双向通道，可通知、提问、指导和终止；另有技能库与长期记忆。进化比较从共同的初始技能库开始。", "refs": [ "method", "setup" ] }, "cycle": { "value": "各题从同一轮共享技能库的隔离副本启动 → 运行中写经验 → 全轮结束后，仅合并 verifier 判成功的运行所产经验 → 下一轮再加载。verifier 不参与撰写经验。", "refs": [ "setup" ] }, "train": { "value": "不训练参数。持续进化实验反复运行 Terminal-Bench 2.0 的 89 道任务，共 20 次改进迭代；经验来自这些任务的实时轨迹。", "refs": [ "setup", "result" ] }, "debug": { "value": "每轮结束后按该轮题目的 verifier 结果筛选经验；GLM-5.1 的 PILOT／Pi／OpenCode 使用相同初始库和经验保存指令，各自独立更新。", "refs": [ "setup", "skills" ] }, "test": { "value": "一次执行比较：Terminal-Bench 2.0、SWE-bench Multilingual、SWE-bench Pro，各题从新状态开始。持续进化结果：同一 Terminal-Bench 2.0，报告历轮最高表现。", "refs": [ "setup", "result" ] }, "isolation": { "value": "同一轮各任务隔离；跨轮重复相同题目，因此进化曲线不是 train/test 隔离实验。SWE-bench 结果是一次执行比较，不是把 Terminal-Bench 进化后的技能拿去做迁移测试。", "refs": [ "setup" ] }, "novelty": { "value": "监督者可以在 worker 尚未结束时指导下一步，同时将经验持久化；区别于只在任务结束后做反思。", "refs": [ "method" ] } }, "tags": [ "Skill", "MemoryContent", "SeparateEvolver", "Online", "SameSet" ], "links": [ { "label": "代码", "url": "https://github.com/XiaoYang66/Pilot" } ], "visibleKeys": [ "object", "executor", "modifier", "verdict", "seed" ], "reviewedAt": "2026-09-15" }, { "id": "2607.12227", "title": "Rethinking the Evaluation of Harness Evolution for Agents", "date": "2026-07-14", "dateLabel": "首次提交", "category": "evaluation", "methodType": "", "priority": "C", "contentType": "paper", "url": "https://arxiv.org/abs/2607.12227", "references": { "main": { "label": "arXiv 摘要与首次提交日期", "url": "https://arxiv.org/abs/2607.12227" }, "intro": { "label": "§1 Introduction", "url": "https://arxiv.org/html/2607.12227#S1" }, "method": { "label": "§3 四种方法及反馈权限", "url": "https://arxiv.org/html/2607.12227#S3" }, "setup": { "label": "§4.1 初始框架、模型与预算", "url": "https://arxiv.org/html/2607.12227#S4.SS1" }, "no": { "label": "§4.2 无测试反馈", "url": "https://arxiv.org/html/2607.12227#S4.SS2" }, "yes": { "label": "§4.3 有测试反馈", "url": "https://arxiv.org/html/2607.12227#S4.SS3" }, "split": { "label": "§4.4 45/10/34 划分", "url": "https://arxiv.org/html/2607.12227#S4.SS4" }, "discussion": { "label": "§5.1–5.2 修改分析与限制", "url": "https://arxiv.org/html/2607.12227#S5" } }, "tldr": { "gap": { "value": "进化时反复用 benchmark 反馈，最后又在同一批题上报分，可能把更多尝试或记住题目带来的收益当成框架改进；缺少相同预算的简单基线也会放大这种误读。", "refs": [ "intro" ] }, "position": { "value": "检验 harness 进化是否优于相同反馈与尝试预算下的直接重试，以及进化后的框架能否迁移到未参与搜索的任务。", "refs": [ "intro", "method" ] }, "conclusion": { "value": "在 Terminal-Bench 2.1 上，harness 进化没有稳定胜过并行采样或逐次修订；45/10/34 划分下，留出测试平均只提高 0.6 个百分点。结论受所测模型、框架和任务范围限制。", "refs": [ "yes", "split", "discussion" ] } }, "dimensions": { "object": { "value": "对比四种改进对象：独立生成多个答案、根据旧轨迹修订答案、跨任务共享的 harness、每题单独定制的 harness。模型参数全部固定。", "refs": [ "method" ] }, "executor": { "value": "无测试反馈：Claude Opus 4.6、GPT-5.4、GPT-5.4 mini；有测试反馈及留出测试：Claude Opus 4.6、GPT-5.4。", "refs": [ "setup", "no", "yes", "split" ] }, "modifier": { "value": "每组采用对应模型生成轨迹总结；meta-agent 据此修改提示、工具说明、记忆或 middleware。直接重试基线只改答案／轨迹，不改共享框架。", "refs": [ "method", "discussion" ] }, "verdict": { "value": "无测试反馈时，修改只看轨迹，并行候选由模型自己选；有测试反馈时，Terminal-Bench 的可执行单元测试返回对错，供修订和选结果。最终成功率都由 benchmark 测试判定。", "refs": [ "method", "no", "yes" ] }, "seed": { "value": "所有方法共用 AHE 的初始 harness：基础提示与 shell 执行工具。进化允许增加提示规则、工具层或运行时检查；直接重试保持该框架固定。", "refs": [ "setup", "discussion" ] }, "cycle": { "value": "统一每题 5 次尝试预算；比较将预算花在并行探索、逐次改解、跨题改框架或单题改框架。候选选择是否可用单元测试，与各组反馈条件保持一致。", "refs": [ "method", "setup" ] }, "train": { "value": "同题比较在 Terminal-Bench 2.1 的 89 题上反复尝试／进化。泛化实验另外把同一 benchmark 分为 45 题进化、10 题验证、34 题测试。", "refs": [ "setup", "split" ] }, "debug": { "value": "泛化实验根据 10 道验证题选最佳 harness；45 道进化题提供单元测试反馈。同题比较直接用当前题的轨迹或测试结果，没有独立验证集。", "refs": [ "method", "split" ] }, "test": { "value": "同题比较仍测这 89 题，分别报告无测试反馈与有测试反馈设置；泛化实验只在另 34 道留出题测选好的框架。", "refs": [ "no", "yes", "split" ] }, "isolation": { "value": "必须分开看两组结论：同题成绩衡量自适应搜索，45/10/34 设置才检验跨题泛化。相同尝试数也不意味着每种方法实际 token 开销完全一致。", "refs": [ "setup", "split" ] }, "novelty": { "value": "同时控制反馈可见性与尝试预算，并把“同题反复搜索”与“留出任务可复用框架”拆开评估。", "refs": [ "intro", "method" ] } }, "tags": [ "HarnessCode", "ExecutableVerifier", "LLMJudge", "SameSet", "HeldOut" ], "links": [ { "label": "代码", "url": "https://github.com/rethinking-harness-evolution" } ], "visibleKeys": [ "object", "executor", "modifier", "verdict", "seed" ], "reviewedAt": "2026-09-15", "attributions": [ { "tag": "org:washington", "label": "University of Washington", "kind": "institution", "sources": [ { "label": "论文作者与机构署名", "url": "https://arxiv.org/html/2607.12227" } ] }, { "tag": "org:allenai", "label": "Allen Institute for AI", "kind": "institution", "sources": [ { "label": "论文作者与机构署名", "url": "https://arxiv.org/html/2607.12227" } ] }, { "tag": "person:hannaneh-hajishirzi", "label": "Hannaneh Hajishirzi", "kind": "scholar", "sources": [ { "label": "论文作者与机构署名", "url": "https://arxiv.org/html/2607.12227" } ] }, { "tag": "person:yulia-tsvetkov", "label": "Yulia Tsvetkov", "kind": "scholar", "sources": [ { "label": "论文作者与机构署名", "url": "https://arxiv.org/html/2607.12227" } ] } ], "priorityBasis": "University of Washington / Allen Institute for AI 参与；直接检验 harness 进化的核心评测问题。" }, { "id": "2609.00768", "title": "DiagEvo: Diagnosis-Guided Self-Evolution via Hierarchical Error Memory", "date": "2026-09-01", "dateLabel": "首次提交", "category": "methods", "methodType": "weights", "priority": "K", "contentType": "paper", "url": "https://arxiv.org/abs/2609.00768", "references": { "main": { "label": "arXiv 摘要与首次提交日期", "url": "https://arxiv.org/abs/2609.00768" }, "intro": { "label": "§1 Introduction", "url": "https://arxiv.org/html/2609.00768#S1" }, "method": { "label": "§3 自博弈、投票与诊断", "url": "https://arxiv.org/html/2609.00768#S3" }, "vote": { "label": "§3.2 伪标签和奖励", "url": "https://arxiv.org/html/2609.00768#S3.SS2" }, "setup": { "label": "§4.1 数据及判分", "url": "https://arxiv.org/html/2609.00768#S4.SS1" }, "result": { "label": "§4.2–4.4 结果与分析", "url": "https://arxiv.org/html/2609.00768#S4.SS2" }, "train": { "label": "附录 E 训练配置", "url": "https://arxiv.org/html/2609.00768#A5" }, "limits": { "label": "附录 A 局限", "url": "https://arxiv.org/html/2609.00768#A1" } }, "tldr": { "gap": { "value": "自主出题能持续训练模型，但仅控制难度和多样性不能告诉出题者“下一轮要补什么能力”；依赖外部题库或教师又削弱了自博弈的独立性。", "refs": [ "intro" ] }, "position": { "value": "研究能否从模型自己的失败历史提取明确的错误原因，并据此生成下一轮训练课程，不使用外部任务材料。", "refs": [ "intro", "method" ] }, "conclusion": { "value": "默认 4B 诊断模型下，三个 solver 的九项评测均分都超过比较方法。Qwen3-8B 的数学均分为 72.3%；训练标签仍来自多数票，共同犯错时可能自我强化。", "refs": [ "result", "limits" ] } }, "dimensions": { "object": { "value": "训练 solver 与 challenger 的参数，并更新指导出题的错误记忆。记忆只给 challenger，不装进 solver 的答题 harness；因此主要归为模型参数进化。", "refs": [ "method" ] }, "executor": { "value": "三个 solver 分别由 Qwen3-4B-Base、Qwen3-8B-Base、OctoThinker-8B-Hybrid-Base 训练得到；测试时由它们直接回答推理题。", "refs": [ "setup" ] }, "modifier": { "value": "challenger 生成数学题；冻结的 Qwen3-4B-Instruct-2507 默认负责诊断和整理错误记忆；训练器交替以 GRPO 更新 challenger 与 solver。另比较 30B-A3B、235B-A22B 诊断模型。", "refs": [ "method", "setup", "train" ] }, "verdict": { "value": "训练：solver 对自生成题采样 12 次，多数答案当伪标签；另采样的训练回答与其一致得 +1，否则 −1。数学测试由 GPT-4o 按 simple-evals 对照标准答案判分；四项通用推理测试用答案 exact match。", "refs": [ "vote", "setup", "train" ] }, "seed": { "value": "R-Zero 式 challenger–solver 自博弈训练循环，新增错误原因诊断、分层记忆和双重投票筛选。起始记忆为空，诊断与检索模型冻结，不是浏览器或 coding-agent 框架。", "refs": [ "method", "train" ] }, "cycle": { "value": "出题 → solver 投票筛题 → 用保留题做 GRPO → 对比与伪标签一致／不一致的轨迹，定位最早推理分歧 → 更新错误原因及掌握状态 → 指导下一轮出题。", "refs": [ "method", "vote" ] }, "train": { "value": "全部训练题由 challenger 在自博弈中生成，内容是数学题，不使用 GSM8K 或 MATH 题目训练。每轮每候选采样 12 次构造标签，优化时另采样 8 次；论文分析五轮演化。", "refs": [ "setup", "train", "result" ] }, "debug": { "value": "按多数答案占比筛选中等难度题，且多数票要明显领先第二名；根据针对某错误的答题一致性更新“待掌握／已掌握”。这些是自生成题上的课程决策，不是外部验证集成绩。", "refs": [ "method", "vote" ] }, "test": { "value": "数学：MATH-500、GSM8K、OlympiadBench、Minerva Math、AMC；AMC 的 40 题各采样 32 次报平均正确率。通用推理：MMLU-Pro、SuperGPQA、GPQA-Diamond、BBEH，贪心解码后 exact match。", "refs": [ "setup" ] }, "isolation": { "value": "测试 benchmark 不供训练循环取题或取标准答案；训练伪标签与测试判分来源不同。自生成题无需人工标签，不等于多数票能证明答案正确。", "refs": [ "setup", "limits" ] }, "novelty": { "value": "把“具体在哪类推理上反复失败”变成有状态的出题依据，同时用第一、第二答案票数差过滤不稳定伪标签。", "refs": [ "method" ] } }, "tags": [ "Weights", "Data", "MemoryContent", "SelfFeedback", "LLMJudge" ], "links": [], "visibleKeys": [ "object", "executor", "modifier", "verdict", "seed" ], "reviewedAt": "2026-09-15", "attributions": [ { "tag": "org:meituan-longcat", "label": "Meituan LongCat", "kind": "institution", "sources": [ { "label": "论文作者机构署名", "url": "https://arxiv.org/html/2609.00768" } ] } ] }, { "id": "2609.08183", "title": "NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness", "date": "2026-09-08", "dateLabel": "首次提交", "category": "methods", "methodType": "weights", "priority": "K", "contentType": "report", "url": "https://arxiv.org/abs/2609.08183", "references": { "main": { "label": "arXiv 摘要与首次提交日期", "url": "https://arxiv.org/abs/2609.08183" }, "intro": { "label": "§1 Introduction", "url": "https://arxiv.org/html/2609.08183#S1" }, "data": { "label": "§3.1–3.4 轨迹、质量与路由", "url": "https://arxiv.org/html/2609.08183#S3" }, "allocation": { "label": "§3.5 按能力调整数据", "url": "https://arxiv.org/html/2609.08183#S3.SS5" }, "train": { "label": "§4 SFT 与 on-policy distillation", "url": "https://arxiv.org/html/2609.08183#S4" }, "eval": { "label": "§5 评测配置与表 1–2", "url": "https://arxiv.org/html/2609.08183#S5" }, "ablation": { "label": "§5.2 数据对比与规模分析", "url": "https://arxiv.org/html/2609.08183#S5.SS2" } }, "tldr": { "gap": { "value": "执行轨迹既记录任务经验，也暴露模型能力缺口；如果只当作固定的训练文本，就无法根据新模型的弱点调整下一轮学习内容。", "refs": [ "intro" ] }, "position": { "value": "探索路由 harness 能否把真实交互中的能力信号转成训练课程和数据配比，持续提升单个 agent 模型。", "refs": [ "intro", "allocation" ] }, "conclusion": { "value": "4B 与 9B 模型在所列十项评测的均分分别提高 5.93、3.44 个百分点。报告把持续多轮更新列为下一步；当前结果不能直接证明长期递归增益。", "refs": [ "eval", "intro" ] } }, "dimensions": { "object": { "value": "更新 Qwen3.5-4B／9B 参数，并按能力缺口调整训练数据配比。Routing harness 负责收集、路由和组织数据；报告没有展示其代码自动进化。", "refs": [ "intro", "allocation", "train" ] }, "executor": { "value": "最终测试用 NeoHorse-1-4B／9B。训练轨迹来自多模型部署系统；模型池与部署配置可变，报告未完整列出所有轨迹生成模型。", "refs": [ "data", "eval" ] }, "modifier": { "value": "固定训练流程执行分阶段 SFT 和 on-policy distillation；路由估计与分层评估调整课程和数据比例。蒸馏教师给学生生成位置的 token 分布，但正文未公开教师具体型号。", "refs": [ "allocation", "train" ] }, "verdict": { "value": "训练入库先用规则检查消息／工具调用完整性，再按目标完成、指令遵循、工具使用、证据一致、错误恢复、终止六项审查轨迹；需语义判断的部分交给 judge。参数监督来自已保留回答和蒸馏教师分布，不是把十个 benchmark 分数直接当逐 token 奖励。", "refs": [ "data", "train" ] }, "seed": { "value": "OpenSquilla 等部署 harness：管理上下文、工具及模型路由，记录请求、推理、工具观察和恢复过程。测试中 QwenClawBench／PinchBench 用 OpenSquilla，WorkBuddy／VitaBench 用各自官方框架。", "refs": [ "intro", "data", "eval" ] }, "cycle": { "value": "收集交互 → 去重、去污染、结构和语义检查 → 按路由估计从低到高组织三阶段 SFT 与学生在线生成的蒸馏 → 分层评估能力缺口 → 调整后续训练数据比例。", "refs": [ "allocation", "train" ] }, "train": { "value": "主要为自建路由 harness 产生的约 10万–100万条轨迹，按用户轮次序列化，保留当前轮推理及历史工具上下文；辅以公开指令、推理、工具和代码数据。受控数据实验另明确比较 Toucan 与自建路由轨迹。", "refs": [ "data", "ablation" ] }, "debug": { "value": "用与训练去重隔离的分层评估集统计能力缺口，指导下一轮数据配比。报告没有公开该分层集的逐项组成，也未说明它与最终十项报告评测是否完全分开，不能称为独立最终测试。", "refs": [ "allocation", "eval" ] }, "test": { "value": "QwenClawBench、WorkBuddy Bench、PinchBench、VitaBench、BFCL V4、τ²-Bench、HumanEval、LiveCodeBench v6、IFBench、IFEval。代码按可执行测试，指令遵循按 benchmark 规则；VitaBench 明确用 DeepSeek-V4-Flash 模拟用户并担任 judge；其他 agent bench 沿用官方评测，报告未逐项展开全部底层裁判。", "refs": [ "eval" ] }, "isolation": { "value": "训练候选与评测题做精确及近似去重。训练隔离不等于评测从未参与调数据；摘要写 eleven benchmarks，但正文表格实际列十项，本站按表格十项记录。", "refs": [ "data", "allocation", "eval" ] }, "novelty": { "value": "区分路由器预测的能力需求、策略调整后的分配和实际服务模型；用预测需求安排课程，避免把“恰好用了强模型”误当成题目难度。", "refs": [ "data", "train" ] } }, "tags": [ "Weights", "Data", "GoldLabel" ], "links": [ { "label": "代码与模型", "url": "https://github.com/TokenRhythm/NeoHorse" } ], "visibleKeys": [ "object", "executor", "modifier", "verdict", "seed" ], "reviewedAt": "2026-09-15" }, { "id": "metarsi-v1", "title": "MetaRSI-v1: A Meta-Recursive Self-Improving System for Recursive Self-Improving Systems Themselves", "date": "2026-09-03", "dateLabel": "报告发布", "category": "methods", "methodType": "joint", "priority": "K", "contentType": "report", "url": "https://www.cosmosmind.ai/research/metarsi-v1", "references": { "main": { "label": "官方技术报告与发布日期", "url": "https://www.cosmosmind.ai/research/metarsi-v1" }, "intro": { "label": "§1，PDF p2–4", "url": "https://www.cosmosmind.ai/research/metarsi-v1.pdf#page=2" }, "method": { "label": "§4 统一循环与三类操作，PDF p9–20", "url": "https://www.cosmosmind.ai/research/metarsi-v1.pdf#page=9" }, "setup": { "label": "§5.1–5.2，PDF p21–22", "url": "https://www.cosmosmind.ai/research/metarsi-v1.pdf#page=21" }, "frontier": { "label": "§5.3–5.4，PDF p23–25", "url": "https://www.cosmosmind.ai/research/metarsi-v1.pdf#page=23" }, "gates": { "label": "附录 D–F，PDF p45–47", "url": "https://www.cosmosmind.ai/research/metarsi-v1.pdf#page=45" } }, "tldr": { "gap": { "value": "现有自改进多依赖容易自动验对的任务，而且数据、harness、参数更新往往分开设计，无法根据当前失败决定该改哪一层。", "refs": [ "intro" ] }, "position": { "value": "研究能否由系统选择并改进数据生成、harness 修改和参数训练的组合顺序，使“怎样改进”本身也能根据经验调整。", "refs": [ "intro", "method" ] }, "conclusion": { "value": "作者报告 Qwen3.5-35B-A3B 四项评测平均提高 10.9 个百分点，比最强固定顺序高 3.6 点。虽以更开放的科学任务为目标，当前实验仍主要是可执行代码与标准答案推理题。", "refs": [ "setup" ] } }, "dimensions": { "object": { "value": "联合更新训练记录、五类 harness 配置和模型 LoRA；调度策略还会改变操作顺序及候选生成指令。API 模型实验只能更新数据与 harness，不能写模型参数。", "refs": [ "method", "setup" ] }, "executor": { "value": "联合更新：自托管 Qwen3.5-35B-A3B。API 路线：GLM-5.2、GPT-5.6 Sol、Kimi K3、Claude Opus 5、DeepSeek V4 Pro、Gemini 3.1 Pro，各自运行自己的改进循环。", "refs": [ "setup", "frontier" ] }, "modifier": { "value": "每组被测模型自己负责诊断、生成、调度和模型式核查，没有外部更强教师；固定代码负责验证、评测及批准发布。LoRA 训练由后端执行。", "refs": [ "setup", "gates" ] }, "verdict": { "value": "代码任务由容器里的可执行 verifier 检查；GPQA-D-hard100 用答案匹配，AIME 用数值匹配。生成训练题另做独立重解、参考解通过／空操作失败、去重和溯源检查；这些入库反馈与最终封存测试分数分开。", "refs": [ "setup", "gates" ] }, "seed": { "value": "自建 RSI-Harness：系统提示、记忆、内置工具、技能、MCP 外接工具五类配置，附固定运行循环与验证接口。推理题关闭工具、技能、MCP，只允许改系统提示；代码题使用完整 agent 配置。", "refs": [ "method", "setup" ] }, "cycle": { "value": "从失败生成结构化诊断 → 选择数据／harness／参数操作及顺序 → 生成候选 → 固定验收程序选择至多一个可发布后继 → 后续阶段重新诊断。参数候选从固定底座按累计数据重新训练，避免混淆旧 adapter 的训练历史。", "refs": [ "method", "gates" ] }, "train": { "value": "Data-RSI 根据失败原因生成、验证训练记录，Model-RSI 用累计保留数据做 LoRA；不是直接拿封存测试题训练。报告给出生成和去重规则，但未公布每个 benchmark 的完整适应题清单与训练样本量。", "refs": [ "method", "gates" ] }, "debug": { "value": "在非封存的 adaptation set 上评估候选并决定保留；共享固定的验收规则。适应集的逐项构成未在报告中完整列出，不能把最终测试集题数当成调试集题数。", "refs": [ "method", "setup" ] }, "test": { "value": "Terminal-Bench 2.1：89 题；SWE-bench Pro：731 题公开测试部分；GPQA-D-hard100：从 Diamond 固定选 100 题，附录列 ID；AIME：2025／2026 的 I、II 共 60 题。最终评价已发布版本，五个随机种子取平均。", "refs": [ "setup", "gates" ] }, "isolation": { "value": "报告规定封存评测在系统冻结后开启，不向中途决策返回成绩；合成数据与封存题做精确、n-gram 和语义去重。声明的隔离规则不能代替尚未公开的逐项运行数据。", "refs": [ "setup", "gates" ] }, "novelty": { "value": "把数据、harness 与参数更新做成可组合的操作，并检查上一阶段产物是否仍适用于当前模型；区别于仅反复调用一个固定修改器。", "refs": [ "method" ] } }, "tags": [ "Weights", "HarnessCode", "Data", "Improver", "JointEvolution" ], "links": [ { "label": "代码", "url": "https://github.com/CosmosMind-ai/RSI-Harness" }, { "label": "PDF", "url": "https://www.cosmosmind.ai/research/metarsi-v1.pdf" } ], "visibleKeys": [ "object", "executor", "modifier", "verdict", "seed" ], "reviewedAt": "2026-09-15" }, { "id": "icoder-27b", "title": "iCoder-27B: Recursive AI-Led Development of Frontier Industrial Coding Model", "date": "2026-09-15", "dateLabel": "收录日期", "category": "methods", "methodType": "weights", "priority": "K", "contentType": "report", "url": "https://huggingface.co/i-Coder/iCoder-27B", "references": { "main": { "label": "官方模型卡及技术报告入口", "url": "https://huggingface.co/i-Coder/iCoder-27B" }, "intro": { "label": "摘要与 §1，PDF p1–3", "url": "https://huggingface.co/i-Coder/iCoder-27B/resolve/main/Coder_Tech_Report.pdf#page=1" }, "prior": { "label": "§3 Research Skills，PDF p5–6", "url": "https://huggingface.co/i-Coder/iCoder-27B/resolve/main/Coder_Tech_Report.pdf#page=5" }, "data": { "label": "§4.1–4.2，PDF p6–8", "url": "https://huggingface.co/i-Coder/iCoder-27B/resolve/main/Coder_Tech_Report.pdf#page=6" }, "training": { "label": "§4.3–4.4，PDF p9–17", "url": "https://huggingface.co/i-Coder/iCoder-27B/resolve/main/Coder_Tech_Report.pdf#page=9" }, "results": { "label": "§6.1–6.3、表 9–10，PDF p18–20", "url": "https://huggingface.co/i-Coder/iCoder-27B/resolve/main/Coder_Tech_Report.pdf#page=18" }, "pool": { "label": "附录 A.1–A.2，PDF p28–29", "url": "https://huggingface.co/i-Coder/iCoder-27B/resolve/main/Coder_Tech_Report.pdf#page=28" }, "eval": { "label": "附录 A.2 逐 benchmark 评测，PDF p29", "url": "https://huggingface.co/i-Coder/iCoder-27B/resolve/main/Coder_Tech_Report.pdf#page=29" } }, "tldr": { "gap": { "value": "在小模型、窄任务上完成一次自动训练，不等于能研发可发布的工业模型；完整开发还要处理数据质量、训练目标失效、verifier 错判和资源约束。", "refs": [ "intro" ] }, "position": { "value": "研究人类把研发经验预先写成 Research Skills 后，agent 能否自主选择和修订实验，开发有竞争力的 RTL／GPU kernel 模型。", "refs": [ "intro", "prior" ] }, "conclusion": { "value": "产出 iCoder-27B，并报告 RTLLM 68.0、KernelBench L1 正确率 61% 等结果。人类仍规定目标、权限和验证标准；报告没有给出完全无人介入研发的对照证据。", "refs": [ "results", "prior" ] } }, "dimensions": { "object": { "value": "主要更新 27B 模型参数，同时生成任务数据、调整 SFT／自蒸馏／RL 的训练方案。人类写的 Research Skills 在运行中固定，不是让 agent 任意重写自身研发框架。", "refs": [ "prior", "data", "training" ] }, "executor": { "value": "训练与最终答题对象为 iCoder-27B，技术报告将起点写为 Qwen3.6-27B。SFT 另用 DeepSeek-V4-Pro 生成并验证教师轨迹。", "refs": [ "data", "results" ] }, "modifier": { "value": "研发 agent 根据 Research Skills 选择数据变换、训练目标和检查点，训练后端执行更新。报告未公开该研发 agent 的具体模型型号；不能把 27B 学生或 SFT 教师当成它。", "refs": [ "prior", "data" ] }, "verdict": { "value": "数据／训练反馈来自编译、仿真和数值检查及其错误记录；SFT 只保留验过的教师解。RTL 用任务 testbench 仿真；GPU kernel 与参考输出比数值，正确后才考虑加速。基础设施故障单列为 unjudged，不当成答错。", "refs": [ "data", "training", "eval" ] }, "seed": { "value": "人类提供 Data→SFT→OPSD→RLVR 的 Research Skills、任务队列、实验日志、资源接口和验证约束。agent 可以回到数据阶段改方案，但不能削弱已批准的正确性标准。", "refs": [ "prior" ] }, "cycle": { "value": "构造可执行任务池 → 找学生失败而教师成功的题做 SFT → 利用失败轨迹与 verifier 诊断做自蒸馏 → 用可执行奖励做 RL；小规模实验的失败可触发数据、奖励或训练目标修订。", "refs": [ "data", "training" ] }, "train": { "value": "自建任务池 55,462 题：RTL 34,972、GPU kernel 20,490。实际各阶段：SFT 28,952 条验过的教师轨迹；OPSD 1,874 道可恢复失败任务；RLVR 13,212 道验证任务。每题带说明、接口、参考实现和验证器；报告未完整列出初始来源题库。", "refs": [ "data", "pool", "results" ] }, "debug": { "value": "阶段内用小规模对照试验及验证结果选择路线和检查点；SFT 选择学生 pass@4 为 0、DeepSeek-V4-Pro 四次至少一次通过的题。文中部分训练诊断也使用命名 benchmark，不能把所有报告结果视为从未用于研发的盲测。", "refs": [ "data", "training" ] }, "test": { "value": "VerilogEval 两种任务各 156 题、RTLLM 50、CVDP 78、ArchXBench 71、RealBench-Module 60；KernelBench 250、TritonBench-G 184。RTL 分别用 Icarus／cocotb／自检 testbench／Verilator；kernel 按参考数值匹配，KernelBench 另报超过 1.05× 的正确加速率。", "refs": [ "eval" ] }, "isolation": { "value": "Research Skills 要求留出隔离和来源谱系划分；报告未给完整 train/validation/test ID 映射。正文将某些 benchmark 用于训练阶段诊断，因此不额外标为严格三阶段隔离。", "refs": [ "prior", "training", "eval" ] }, "novelty": { "value": "把人类输入集中成可复用研发规则，让 agent 根据实测故障改训练决策；贡献是完整工业模型开发过程，不只是自动跑一段固定训练脚本。", "refs": [ "prior", "training" ] } }, "tags": [ "Weights", "Data", "ExecutableVerifier" ], "links": [ { "label": "代码", "url": "https://github.com/bingreeky/iCoder" }, { "label": "PDF", "url": "https://huggingface.co/i-Coder/iCoder-27B/resolve/main/Coder_Tech_Report.pdf" } ], "visibleKeys": [ "object", "executor", "modifier", "verdict", "seed" ], "reviewedAt": "2026-09-15", "attributions": [ { "tag": "org:sjtu", "label": "Shanghai Jiao Tong University", "kind": "institution", "sources": [ { "label": "技术报告 p1 作者机构", "url": "https://huggingface.co/i-Coder/iCoder-27B/resolve/main/Coder_Tech_Report.pdf#page=1" } ] }, { "tag": "org:nus", "label": "National University of Singapore", "kind": "institution", "sources": [ { "label": "技术报告 p1 作者机构", "url": "https://huggingface.co/i-Coder/iCoder-27B/resolve/main/Coder_Tech_Report.pdf#page=1" } ] } ] }, { "id": "2609.11873", "title": "The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement", "date": "2026-09-10", "dateLabel": "首次提交", "category": "overview", "methodType": "", "priority": "C", "contentType": "survey", "url": "https://arxiv.org/abs/2609.11873", "references": { "main": { "label": "arXiv 摘要与首次提交日期", "url": "https://arxiv.org/abs/2609.11873" }, "intro": { "label": "§1 Introduction", "url": "https://arxiv.org/html/2609.11873#S1" }, "roadmap": { "label": "§2 RSI 定义与自主程度", "url": "https://arxiv.org/html/2609.11873#S2" }, "landscape": { "label": "§3–4 方法及场景", "url": "https://arxiv.org/html/2609.11873#S3" }, "industry": { "label": "§5 Industry Practice", "url": "https://arxiv.org/html/2609.11873#S5" }, "limits": { "label": "§6 挑战与未来方向", "url": "https://arxiv.org/html/2609.11873#S6" } }, "tldr": { "gap": { "value": "可自动判分的封闭任务进步很快，但长程真实工作仍受状态跟踪、反馈可靠性和失败累积限制；“自改进”一词又混合了自主程度很不同的系统。", "refs": [ "intro" ] }, "position": { "value": "为 RSI 建立按自主程度递进的发展路线，分析不同任务场景需要什么能力，以及现有系统距离更自主的改进还差什么。", "refs": [ "roadmap" ] }, "conclusion": { "value": "区分自主执行改进、制定改进策略、获取经验、适应环境、递归改进改进机制五层。文章提供路线图和初步产业案例，不代表五层能力都已被实验实现。", "refs": [ "roadmap", "industry", "limits" ] } }, "dimensions": { "scope": { "value": "覆盖方法分类、科学发现、具身智能、软件工程及企业实践；适合作为 RSI 全景阅读入口。", "refs": [ "landscape", "industry" ] }, "framework": { "value": "用五种自主能力比较系统：谁执行改进、谁选策略、谁获取经验、谁适应新环境，以及改进机制能否继续改进自身。", "refs": [ "roadmap" ] }, "feedback": { "value": "比较可执行验证、环境反馈和开放任务评价的不同限制；强调可验证、可迁移、可保留的改进，而非一次分数增加。", "refs": [ "roadmap", "limits" ] }, "evidence": { "value": "结合文献、跨领域 benchmark 趋势和产业案例。HCI 表示相对该 benchmark 首年水平消除了多少分数差距，不是任务准确率，也不能直接混合不同评测协议。", "refs": [ "roadmap", "industry" ] }, "reading": { "value": "先读 §2 建立分类，再按关注场景读 §4；§5 的产业案例用于理解实践路径，不与统一预算的 benchmark 实验混为一谈。", "refs": [ "roadmap", "industry" ] }, "limits": { "value": "路线中的后期自主能力属于研究目标；图中的未来趋势与初步案例不能作为已实现无限递归提升的证据。", "refs": [ "roadmap", "limits" ] } }, "tags": [ "Survey", "RSIRoadmap" ], "links": [ { "label": "项目页", "url": "https://theseus-labs-rsi.github.io/" } ], "visibleKeys": [ "scope", "framework", "feedback" ], "reviewedAt": "2026-09-15", "priorityBasis": "面向 RSI 全景与自主程度的发展路线，作为入门主线阅读。" }, { "id": "data-centric-rsi-survey", "title": "Data-Centric Recursive Improvement for Foundation Models: A Survey", "date": "2026-09-02", "dateLabel": "发布", "category": "overview", "methodType": "", "priority": "K", "contentType": "survey", "url": "https://www.preprints.org/manuscript/202609.0154", "references": { "main": { "label": "Preprints 摘要、首次发布日期", "url": "https://www.preprints.org/manuscript/202609.0154" }, "framework": { "label": "§1–2 引言、定义与分类", "url": "https://www.preprints.org/manuscript/202609.0154" }, "repo": { "label": "作者配套阅读地图：Signal–Decision–Update Loop", "url": "https://github.com/zoezou2015/Awesome-Data-Centric-Recursive-Improvement" } }, "tldr": { "gap": { "value": "数据构建与评测常被分开讨论，难以追踪一次评测发现的弱点到底怎样改变下一轮数据和模型。", "refs": [ "framework" ] }, "position": { "value": "以数据相关对象为中心，整理评价信号如何经过可信度判断和干预决策，形成可持续的模型改进。", "refs": [ "framework" ] }, "conclusion": { "value": "用“信号—决策—更新”连接数据选择、过滤、合成及后训练；同时分析反馈过拟合、错误评价和更新不稳定。它是整理机制的综述，不是一套统一训练实验。", "refs": [ "framework", "repo" ] } }, "dimensions": { "scope": { "value": "聚焦反馈驱动的数据变化，包括训练数据、检索和记忆等影响后续模型行为的对象。", "refs": [ "framework" ] }, "framework": { "value": "三个问题：什么信号发现弱点；谁判断信号可信并决定如何干预；具体更新什么数据对象。", "refs": [ "framework" ] }, "feedback": { "value": "分别讨论可执行结果、标准答案、模型评价和偏好等信号；强调保留监督来源，不能把一次高分自动当作正确更新依据。", "refs": [ "framework" ] }, "reading": { "value": "先读 §1–2 的分类，再用作者 GitHub 阅读地图按信号和数据更新路径查文献，适合与本站的执行者／修改者维度配合。", "refs": [ "framework", "repo" ] }, "limits": { "value": "不同文献的训练数据、反馈和测试协议各不相同；此综述的分类不是统一实验结论。Preprints 页面标注为未同行评审的预印本。", "refs": [ "main" ] } }, "tags": [ "Survey", "Data" ], "links": [ { "label": "配套仓库", "url": "https://github.com/zoezou2015/Awesome-Data-Centric-Recursive-Improvement" } ], "visibleKeys": [ "scope", "framework", "feedback" ], "reviewedAt": "2026-09-15" }, { "id": "ai4ai-survey", "title": "AI4AI Survey: From Long-Horizon Agents to Recursive Self-Improvement", "date": "2026-08", "dateLabel": "发布", "category": "overview", "methodType": "", "priority": "K", "contentType": "survey", "url": "https://simpleagentlab.com/ai4ai/", "references": { "main": { "label": "官方项目页：2026 年 8 月综述", "url": "https://simpleagentlab.com/ai4ai/" }, "paper": { "label": "论文 §2、Appendix A：定义与覆盖范围", "url": "https://www.preprints.org/manuscript/202608.2108" }, "map": { "label": "作者配套文献地图", "url": "https://kaiwu5.github.io/Awesome-AI4AI/" } }, "tldr": { "gap": { "value": "规划、写代码、跑实验等单项能力很强，连成完整科研流程却未必可靠；分散的 agent、AI4AI 与 RSI 术语又让这种差距不易比较。", "refs": [ "main" ] }, "position": { "value": "研究怎样界定和衡量 AI 将改进想法推进到可靠验证结果的能力，并区分一次自改进与改进机制本身的递归变化。", "refs": [ "main", "paper" ] }, "conclusion": { "value": "按目标、规划、执行、反馈、修复的承担者分析系统，强调组件组合后可靠性下降及人类仍掌握的决策权。这些判断来自所审阅系统，不是对所有未来系统的结论。", "refs": [ "main", "paper" ] } }, "dimensions": { "scope": { "value": "覆盖长程 agent、模型与 harness 两条改进路线、自动化研究和自修改系统。", "refs": [ "main" ] }, "framework": { "value": "逐阶段区分人类、系统及双方共同承担的工作；“长程”重点是早期动作影响后续动作与最终结果，不只是 token 多或运行时间长。", "refs": [ "paper" ] }, "feedback": { "value": "检查评价标准是谁定的、结果如何影响修复和后续决策，并提醒匹配预算、验证泄漏和人类干预都会影响成绩解释。", "refs": [ "paper" ] }, "reading": { "value": "先看定义和分类，再读两条改进路线；配套 Awesome-AI4AI 提供文献地图，RSIHub 提供实验框架入口。", "refs": [ "main", "map" ] }, "limits": { "value": "对所审阅系统的归纳受搜索时间与纳入标准限制；不能把“单项能力强”解释为端到端自主研发已解决。项目页和论文的具体定义应分别按原文阅读。", "refs": [ "main", "paper" ] } }, "tags": [ "Survey", "LongHorizon" ], "links": [ { "label": "论文", "url": "https://www.preprints.org/manuscript/202608.2108" }, { "label": "配套文献地图", "url": "https://kaiwu5.github.io/Awesome-AI4AI/" } ], "visibleKeys": [ "scope", "framework", "feedback" ], "reviewedAt": "2026-09-15" }, { "id": "reef", "title": "Reef: Continual Learning Infrastructure for Self-Improving Agents", "date": "2026-09-15", "dateLabel": "收录日期", "category": "resources", "methodType": "", "priority": "K", "contentType": "repository", "url": "https://github.com/Human-Agent-Society/reef", "references": { "main": { "label": "README：How it works / When to use Reef", "url": "https://github.com/Human-Agent-Society/reef" }, "harness": { "label": "README：Harness-evolving deployment", "url": "https://github.com/Human-Agent-Society/reef#harness-evolving-deployment" }, "recipes": { "label": "README：Recipes and examples", "url": "https://github.com/Human-Agent-Society/reef#recipes-and-examples" } }, "tldr": { "gap": { "value": "持续学习不只是训练一次模型，还需要把线上交互、反馈、候选评价和版本交付接在一起，否则更新难以稳定进入实际服务。", "refs": [ "main" ] }, "position": { "value": "提供可复用基础设施，让已有 agent 从交互记录学习，并将通过评价的参数或 harness 更新交付给后续请求。", "refs": [ "main" ] }, "conclusion": { "value": "支持参数训练与提示／规则／技能更新，保留版本历史。具体学什么、怎样评分由学习配置（recipe）决定；安装框架本身不等于 agent 会自动获得可靠提升。", "refs": [ "main", "recipes" ] } }, "dimensions": { "object": { "value": "按学习配置（recipe） 更新模型参数，或 harness 中的提示、规则、技能；两条路线都由版本系统保存与交付。", "refs": [ "main" ] }, "executor": { "value": "由用户配置的模型执行请求；参数训练可接 Slime 训练后端与 SGLang 推理服务，harness 路线可接模型 API，无需本地训练 GPU。", "refs": [ "main" ] }, "modifier": { "value": "学习配置（recipe）定义从交互和反馈产生更新的方法，候选评估模块按配置选择更新，版本模块保存和交付结果。没有一个固定通用的修改模型。", "refs": [ "main", "recipes" ] }, "verdict": { "value": "反馈需关联到具体交互记录，可是分数或结构化说明；例如 coding 教程把测试失败上报后生成候选技能，再与当前版本在三道教学任务上比较，仅胜出才发布。", "refs": [ "main", "harness" ] }, "seed": { "value": "已有推理服务与 agent 客户端之外，补上交互记录、反馈匹配、更新任务、候选评价和版本交付。它是连接组件的基础设施，不是统一的 benchmark 方法。", "refs": [ "main" ] }, "data": { "value": "数据由部署场景与 recipe 决定。README 列出的示例包括 AIME 2025、IMOAnswerBench、Terminal-Bench、GSM8K 任务流和 WildClawBench；这些不是全项目共享的一套训练／测试划分。", "refs": [ "recipes" ] }, "use": { "value": "适合搭建自己的持续学习或 harness 优化实验；先选 recipe，再明确任务集、反馈实现、候选选择和独立测试集。", "refs": [ "main", "recipes" ] }, "limits": { "value": "框架提供接口与版本管理，不替使用者保证反馈正确、数据隔离或改进泛化。示例任务上的候选胜出不能直接当成独立 benchmark 提升。", "refs": [ "main", "harness" ] } }, "tags": [ "Infrastructure", "Weights", "Skill", "Continual" ], "links": [ { "label": "仓库", "url": "https://github.com/Human-Agent-Society/reef" } ], "visibleKeys": [ "object", "executor", "modifier", "verdict", "seed" ], "reviewedAt": "2026-09-15" }, { "id": "openai-research-acceleration", "title": "Research Acceleration: The View Inside OpenAI", "date": "2026-09-06", "dateLabel": "博客发布", "category": "overview", "methodType": "", "priority": "K", "contentType": "blog", "url": "https://openai.com/index/research-acceleration-view-inside-openai/", "references": { "main": { "label": "官方博客：导言与 §1–5", "url": "https://openai.com/index/research-acceleration-view-inside-openai/" }, "methods": { "label": "官方博客：Appendix / Our methods for this post", "url": "https://openai.com/index/research-acceleration-view-inside-openai/" } }, "tldr": { "gap": { "value": "外界难以观察前沿实验室内部 agent 到底承担了多少研究工作；代码和实验数量增加，也未必等同整体科研进展加速。", "refs": [ "main" ] }, "position": { "value": "以内部使用记录说明 coding agent 如何参与 AI 研发，测量使用量、任务类型、完成情况及人类干预。", "refs": [ "main", "methods" ] }, "conclusion": { "value": "报告 agent 使用、代码产出与实验量增加，且承担更复杂任务；人类仍决定研究方向、结果取舍和部署。数据主要是观察性证据，不能单凭相关性证明自主 RSI 已实现。", "refs": [ "main", "methods" ] } }, "dimensions": { "scope": { "value": "OpenAI 内部研究流程中的 agent 使用，涵盖实现、运行、分析等环节；不是公开统一训练 benchmark。", "refs": [ "main" ] }, "actors": { "value": "coding agents 辅助研究人员；人类继续设定优先级、决定追哪些结果及是否扩大训练或部署。文章不提供一个可复现的固定执行模型／harness 配置。", "refs": [ "main" ] }, "feedback": { "value": "按内部使用与任务结果记录统计，由 agent 分类器分析任务类型和可判断结果的成功率；不确定结果排除，另统计人类干预。不是 benchmark verifier 给所有会话逐一验对。", "refs": [ "main", "methods" ] }, "data": { "value": "主要分析 2026 年内部会话、代码和实验运行数据；按月份与估计任务难度分组。文章没有提供一套可下载的 train/test 数据集。", "refs": [ "main", "methods" ] }, "limits": { "value": "算力、模型、使用方式和研究流程同时变化，代码量／实验量不是研究质量的直接度量；内部观察不能等同受控因果实验或多代自改进结果。", "refs": [ "main", "methods" ] } }, "tags": [ "IndustryReport", "ResearchAutomation" ], "links": [], "visibleKeys": [ "scope", "actors", "feedback" ], "reviewedAt": "2026-09-15", "attributions": [ { "tag": "org:openai", "label": "OpenAI", "kind": "institution", "sources": [ { "label": "官方博客：导言与 §1–5", "url": "https://openai.com/index/research-acceleration-view-inside-openai/" } ] } ] }, { "id": "shiyu-rsi-what-evolves", "title": "什么在进化？——Model、Harness 与 Artifact 的三层地图", "url": "https://shiyunee.github.io/blogs/RSI-What-Evolves/", "date": "2026-08-10", "dateLabel": "博客发布", "category": "overview", "methodType": "", "priority": "K", "contentType": "blog", "references": { "main": { "label": "§1–2、§4、§6、§8；写在前面：来源与致谢", "url": "https://shiyunee.github.io/blogs/RSI-What-Evolves/" } }, "tldr": { "gap": { "value": "不同工作都叫 self-evolving，却可能只改答案、改运行框架或训练模型；混在一起比较，容易把产物变好当成系统能力变强。", "refs": [ "main" ] }, "position": { "value": "中文入门与交叉阅读笔记：以 Model、Harness、Artifact 三层地图组织案例，对照 Shilong Liu、Lilian Weng 与知乎文章的不同视角。", "refs": [ "main" ] }, "conclusion": { "value": "把“修改对象”与“反馈是否改进下一轮的方法”分开，说明三层如何相互促进；核心分类和案例来源明确归于原作者。", "refs": [ "main" ] } }, "dimensions": { "author": { "value": "作者：Shiyu Ni；个人研究博客系列。", "refs": [ "main" ] }, "scope": { "value": "中文入门与交叉阅读笔记：以 Model、Harness、Artifact 三层地图组织案例，对照 Shilong Liu、Lilian Weng 与知乎文章的不同视角。", "refs": [ "main" ] }, "reading": { "value": "先读 §2 三层划分，再读 §4 harness 的设计与优化，最后用 §8 的反馈、评估和递归问题检查具体案例。", "refs": [ "main" ] }, "evidence": { "value": "依据 §1–2、§4、§6、§8；写在前面：来源与致谢；简介依据文章原文；文中研究判断与实验宣称归于作者。", "refs": [ "main" ] } }, "tags": [ "RSIRoadmap" ], "links": [ { "label": "系列另一篇", "url": "https://shiyunee.github.io/blogs/RSI-Where-the-Loop-Closes/" } ], "visibleKeys": [ "author", "reading" ], "reviewedAt": "2026-09-16", "reviewStatus": "source-reviewed" }, { "id": "shiyu-rsi-loop-closes", "title": "递归如何闭环？——从可修改范围到 Agent as Service", "url": "https://shiyunee.github.io/blogs/RSI-Where-the-Loop-Closes/", "date": "2026-08-10", "dateLabel": "博客发布", "category": "overview", "methodType": "", "priority": "K", "contentType": "blog", "references": { "main": { "label": "定义；Everything as Service；Scoped → Joint → Recursive；来源与致谢", "url": "https://shiyunee.github.io/blogs/RSI-Where-the-Loop-Closes/" } }, "tldr": { "gap": { "value": "迭代次数多不等于递归；同时开放数据、训练、环境和框架，也容易让能力提升的来源无法区分。", "refs": [ "main" ] }, "position": { "value": "基于 Melon 长文的中文学习笔记，梳理可修改范围，并讨论怎样设计能归因、能复用、能验证递归的研究系统。", "refs": [ "main" ] }, "conclusion": { "value": "先隔离单个研究对象，再测试联合改进，最后检查改进后的系统是否更会改进下一代；服务化的意义是明确实验边界。", "refs": [ "main" ] } }, "dimensions": { "author": { "value": "作者：Shiyu Ni；个人研究博客系列。", "refs": [ "main" ] }, "scope": { "value": "基于 Melon 长文的中文学习笔记，梳理可修改范围，并讨论怎样设计能归因、能复用、能验证递归的研究系统。", "refs": [ "main" ] }, "reading": { "value": "接上篇阅读：重点看各类 scope 的区别，以及如何固定其他组件、隔离评测、记录版本来判断改进来自哪里。", "refs": [ "main" ] }, "evidence": { "value": "依据 定义；Everything as Service；Scoped → Joint → Recursive；来源与致谢；简介依据文章原文；文中研究判断与实验宣称归于作者。", "refs": [ "main" ] } }, "tags": [ "RSIRoadmap" ], "links": [ { "label": "系列另一篇", "url": "https://shiyunee.github.io/blogs/RSI-What-Evolves/" } ], "visibleKeys": [ "author", "reading" ], "reviewedAt": "2026-09-16", "reviewStatus": "source-reviewed" }, { "id": "liu-self-evolving-taxonomy", "title": "A Taxonomy of Self-evolving Agents", "url": "https://lsl.zone/blog/2026/a-taxonomy-of-self-evolving-agents/", "date": "2026-07-08", "dateLabel": "博客发布", "category": "overview", "methodType": "", "priority": "K", "contentType": "blog", "references": { "main": { "label": "Models, Harness, and Artifacts；Agent Harness Self-improvement；A Blurred Boundary", "url": "https://lsl.zone/blog/2026/a-taxonomy-of-self-evolving-agents/" } }, "tldr": { "gap": { "value": "相似的自进化名称涵盖不同修改对象，难以判断工作之间的关系。", "refs": [ "main" ] }, "position": { "value": "按产物优化、harness 自改进、无标准答案的模型学习，给出领域分类地图。", "refs": [ "main" ] }, "conclusion": { "value": "再将 harness 展开为 prompt／memory、tool／skill 和多 agent；强调三层可相互促进，最终价值要看实际产物。", "refs": [ "main" ] } }, "dimensions": { "author": { "value": "作者／发布者：Shilong Liu。", "refs": [ "main" ] }, "scope": { "value": "按产物优化、harness 自改进、无标准答案的模型学习，给出领域分类地图。", "refs": [ "main" ] }, "reading": { "value": "适合作为分类入口；对照本站“什么在变”，再读文末三层边界及现实应用的讨论。", "refs": [ "main" ] }, "evidence": { "value": "依据 Models, Harness, and Artifacts；Agent Harness Self-improvement；A Blurred Boundary；简介依据文章原文；文中研究判断与实验宣称归于作者。", "refs": [ "main" ] } }, "tags": [ "RSIRoadmap" ], "links": [ { "label": "Shiyu 的相关解读", "url": "https://shiyunee.github.io/blogs/RSI-What-Evolves/" } ], "visibleKeys": [ "author", "reading" ], "reviewedAt": "2026-09-16", "reviewStatus": "source-reviewed", "attributions": [ { "tag": "person:shilong-liu", "label": "Shilong Liu", "kind": "scholar", "sources": [ { "label": "原文署名／发布机构", "url": "https://lsl.zone/blog/2026/a-taxonomy-of-self-evolving-agents/" } ] } ] }, { "id": "weng-harness-engineering", "title": "Harness Engineering for Self-Improvement", "url": "https://lilianweng.github.io/posts/2026-07-04-harness/", "date": "2026-07-04", "dateLabel": "博客发布", "category": "overview", "methodType": "", "priority": "K", "contentType": "blog", "references": { "main": { "label": "Harness Design Patterns；Harness Optimization；Harness Layer vs Core Intelligence?", "url": "https://lilianweng.github.io/posts/2026-07-04-harness/" } }, "tldr": { "gap": { "value": "模型能力之外，执行循环、状态管理和评估也决定 agent 能否稳定完成长期任务。", "refs": [ "main" ] }, "position": { "value": "从工程设计和研究路线解释 harness 如何支持自改进，而不只是列举提示技巧。", "refs": [ "main" ] }, "conclusion": { "value": "串起工作流、文件记忆、并行子 agent，以及从上下文到框架代码、修改机制和参数联合优化的路径。", "refs": [ "main" ] } }, "dimensions": { "author": { "value": "作者／发布者：Lilian Weng。", "refs": [ "main" ] }, "scope": { "value": "从工程设计和研究路线解释 harness 如何支持自改进，而不只是列举提示技巧。", "refs": [ "main" ] }, "reading": { "value": "先读三种设计模式，再按 Harness Optimization 对照论文；近期 RSI 路线是作者判断，不是已实现的统一实验结论。", "refs": [ "main" ] }, "evidence": { "value": "依据 Harness Design Patterns；Harness Optimization；Harness Layer vs Core Intelligence?；简介依据文章原文；文中研究判断与实验宣称归于作者。", "refs": [ "main" ] } }, "tags": [ "ResearchAutomation" ], "links": [ { "label": "Shiyu 的相关解读", "url": "https://shiyunee.github.io/blogs/RSI-What-Evolves/" } ], "visibleKeys": [ "author", "reading" ], "reviewedAt": "2026-09-16", "reviewStatus": "source-reviewed", "attributions": [ { "tag": "person:lilian-weng", "label": "Lilian Weng", "kind": "scholar", "sources": [ { "label": "原文署名／发布机构", "url": "https://lilianweng.github.io/posts/2026-07-04-harness/" } ] } ] }, { "id": "zhihu-self-evolving-guide", "title": "自进化（Self-evolving／RSI），一篇就够了", "url": "https://zhuanlan.zhihu.com/p/2065227313973825752", "date": "2026-09-16", "dateLabel": "收录日期（原文发布日期未确认）", "category": "overview", "methodType": "", "priority": "R", "contentType": "blog", "references": { "main": { "label": "上篇：来源与致谢、§3、§4.8、§6（Shiyu 转述）", "url": "https://shiyunee.github.io/blogs/RSI-What-Evolves/" } }, "tldr": { "gap": { "value": "据 Shiyu 的转述，文章通过案例解释产物、harness 和模型如何参与自进化。", "refs": [ "main" ] }, "position": { "value": "作为中文案例综述的原始阅读入口收录；本站简介仅依据 Shiyu 的引用与转述。", "refs": [ "main" ] }, "conclusion": { "value": "转述涉及 AlphaEvolve、Hermes、AIDE² 和联合优化，并保留预算对等、独立评测等反面检查。未直接读取原文，不扩写实验结论。", "refs": [ "main" ] } }, "dimensions": { "author": { "value": "作者／发布者：知乎原作者（当前来源未提供署名）。", "refs": [ "main" ] }, "scope": { "value": "作为中文案例综述的原始阅读入口收录；本站简介仅依据 Shiyu 的引用与转述。", "refs": [ "main" ] }, "reading": { "value": "可先读 Shiyu 上篇对应案例，再打开知乎原文；这是一条引用线索，不是本站已直接核对的原文摘要。", "refs": [ "main" ] }, "evidence": { "value": "依据 上篇：来源与致谢、§3、§4.8、§6；原文当前无法直接读取，只按 Shiyu 博客中明确引用的内容建立阅读入口。", "refs": [ "main" ] } }, "tags": [ "RSIRoadmap" ], "links": [ { "label": "Shiyu 的相关解读", "url": "https://shiyunee.github.io/blogs/RSI-What-Evolves/" } ], "visibleKeys": [ "author", "reading" ], "reviewedAt": "2026-09-16", "reviewStatus": "secondary-reference" }, { "id": "melon-self-evolve-rsi", "title": "万字深度：从 Self-Evolve 到 RSI", "url": "https://www.xiaohongshu.com/discovery/item/6a71f07700000000220146bf", "date": "2026-09-16", "dateLabel": "收录日期（原文发布日期未确认）", "category": "overview", "methodType": "", "priority": "R", "contentType": "blog", "references": { "main": { "label": "下篇：来源与致谢、定义、Everything as Service（Shiyu 转述）", "url": "https://shiyunee.github.io/blogs/RSI-Where-the-Loop-Closes/" } }, "tldr": { "gap": { "value": "据 Shiyu 的转述，只扩大可修改范围不足以解释改进的来源，也不足以证明改进者本身变强。", "refs": [ "main" ] }, "position": { "value": "收录下篇学习笔记的思想来源：按可修改范围组织研究，并讨论服务化的实验架构。", "refs": [ "main" ] }, "conclusion": { "value": "转述的主线是 Everything as Service 与 Scoped → Joint → Recursive；当前简介来自下篇的明确归因，不冒充原文直接核对。", "refs": [ "main" ] } }, "dimensions": { "author": { "value": "作者／发布者：Melon（署名据 Shiyu 的致谢）。", "refs": [ "main" ] }, "scope": { "value": "收录下篇学习笔记的思想来源：按可修改范围组织研究，并讨论服务化的实验架构。", "refs": [ "main" ] }, "reading": { "value": "小红书原文可能需要登录；可结合 Shiyu 下篇阅读，注意区分原作者框架与学习笔记补充的论文说明。", "refs": [ "main" ] }, "evidence": { "value": "依据 下篇：来源与致谢、定义、Everything as Service；原文当前无法直接读取，只按 Shiyu 博客中明确引用的内容建立阅读入口。", "refs": [ "main" ] } }, "tags": [ "RSIRoadmap" ], "links": [ { "label": "Shiyu 的相关解读", "url": "https://shiyunee.github.io/blogs/RSI-Where-the-Loop-Closes/" } ], "visibleKeys": [ "author", "reading" ], "reviewedAt": "2026-09-16", "reviewStatus": "secondary-reference" }, { "id": "yudkowsky-recursive-self-improvement", "title": "Recursive Self-Improvement", "url": "https://www.lesswrong.com/posts/JBadX7rwdcRFzGuju/recursive-self-improvement", "date": "2008-12-01", "dateLabel": "博客发布", "category": "overview", "methodType": "", "priority": "K", "contentType": "blog", "references": { "main": { "label": "正文：optimization slope/resources/efficiency 与递归层次", "url": "https://www.lesswrong.com/posts/JBadX7rwdcRFzGuju/recursive-self-improvement" } }, "tldr": { "gap": { "value": "讨论 AI 能力增长为什么可能不是匀速，以及快速增长需要哪些依赖关系。", "refs": [ "main" ] }, "position": { "value": "RSI 的早期概念论述，分析优化机会、资源、效率以及改进认知机制的反馈关系。", "refs": [ "main" ] }, "conclusion": { "value": "区分解决任务、获得知识、改进发现知识的方法与改进认知机制；属于理论论证，不是现代 agent 的实证报告。", "refs": [ "main" ] } }, "dimensions": { "author": { "value": "作者／发布者：Eliezer Yudkowsky。", "refs": [ "main" ] }, "scope": { "value": "RSI 的早期概念论述，分析优化机会、资源、效率以及改进认知机制的反馈关系。", "refs": [ "main" ] }, "reading": { "value": "适合理解“递归”为什么不等于重复尝试；不要把文中的增长预期当成今天系统已验证的能力。", "refs": [ "main" ] }, "evidence": { "value": "依据 正文：optimization slope/resources/efficiency 与递归层次；简介依据文章原文；文中研究判断与实验宣称归于作者。", "refs": [ "main" ] } }, "tags": [ "RSIRoadmap" ], "links": [ { "label": "Shiyu 的相关解读", "url": "https://shiyunee.github.io/blogs/RSI-What-Evolves/" } ], "visibleKeys": [ "author", "reading" ], "reviewedAt": "2026-09-16", "reviewStatus": "source-reviewed" }, { "id": "weco-aide2-blog", "title": "AIDE²: The First Evidence of Recursive Self-Improvement", "url": "https://www.weco.ai/blog/first-evidence-of-recursive-self-improvement", "date": "2026-07-14", "dateLabel": "博客发布", "category": "overview", "methodType": "", "priority": "K", "contentType": "blog", "references": { "main": { "label": "§1、The inner-loop evaluation、§2.2、§3.1、§4", "url": "https://www.weco.ai/blog/first-evidence-of-recursive-self-improvement" } }, "tldr": { "gap": { "value": "作者希望区分仅自动运行研究，与在固定成本下持续提高研究效率。", "refs": [ "main" ] }, "position": { "value": "双层 autoresearch 案例：外层修改内层 agent 的代码，检验改进能否迁移到未用于优化的任务。", "refs": [ "main" ] }, "conclusion": { "value": "报告内层能力提升和外部任务迁移，但明确未达到改进后的内层也更擅长充当外层的条件；标题中的“首次”是作者主张。", "refs": [ "main" ] } }, "dimensions": { "author": { "value": "作者／发布者：Weco Team。", "refs": [ "main" ] }, "scope": { "value": "双层 autoresearch 案例：外层修改内层 agent 的代码，检验改进能否迁移到未用于优化的任务。", "refs": [ "main" ] }, "reading": { "value": "重点读 public/private 分数、成本约束与 ignition test，区分内层变强和修改者变强。", "refs": [ "main" ] }, "evidence": { "value": "依据 §1、The inner-loop evaluation、§2.2、§3.1、§4；简介依据文章原文；文中研究判断与实验宣称归于作者。", "refs": [ "main" ] } }, "tags": [ "ResearchAutomation" ], "links": [ { "label": "Shiyu 的相关解读", "url": "https://shiyunee.github.io/blogs/RSI-What-Evolves/" } ], "visibleKeys": [ "author", "reading" ], "reviewedAt": "2026-09-16", "reviewStatus": "source-reviewed", "attributions": [ { "tag": "org:weco", "label": "Weco", "kind": "institution", "sources": [ { "label": "原文署名／发布机构", "url": "https://www.weco.ai/blog/first-evidence-of-recursive-self-improvement" } ] } ] }, { "id": "deepmind-alphaevolve-blog", "title": "AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms", "url": "https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/", "date": "2025-05-14", "dateLabel": "博客发布", "category": "overview", "methodType": "", "priority": "K", "contentType": "blog", "references": { "main": { "label": "Designing better algorithms with large language models；导言与应用案例", "url": "https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/" } }, "tldr": { "gap": { "value": "复杂算法优化需要大量候选探索，同时需要可靠地验证候选是否正确、是否更快。", "refs": [ "main" ] }, "position": { "value": "官方案例介绍：Gemini 提出算法代码，自动评估器验证，进化搜索继续改进有潜力的候选。", "refs": [ "main" ] }, "conclusion": { "value": "展示数学、数据中心和 AI 训练相关算法改进；主要修改对象是算法产物，训练基础设施获益与搜索器自身改进应分开理解。", "refs": [ "main" ] } }, "dimensions": { "author": { "value": "作者／发布者：AlphaEvolve team · Google DeepMind。", "refs": [ "main" ] }, "scope": { "value": "官方案例介绍：Gemini 提出算法代码，自动评估器验证，进化搜索继续改进有潜力的候选。", "refs": [ "main" ] }, "reading": { "value": "适合理解“产物进化如何反哺模型训练”；先看代码生成与验证循环，再看实际应用。", "refs": [ "main" ] }, "evidence": { "value": "依据 Designing better algorithms with large language models；导言与应用案例；简介依据文章原文；文中研究判断与实验宣称归于作者。", "refs": [ "main" ] } }, "tags": [ "ResearchAutomation" ], "links": [ { "label": "Shiyu 的相关解读", "url": "https://shiyunee.github.io/blogs/RSI-What-Evolves/" } ], "visibleKeys": [ "author", "reading" ], "reviewedAt": "2026-09-16", "reviewStatus": "source-reviewed", "attributions": [ { "tag": "org:google-deepmind", "label": "Google DeepMind", "kind": "institution", "sources": [ { "label": "原文署名／发布机构", "url": "https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/" } ] } ] }, { "id": "minimax-m27-blog", "title": "MiniMax M2.7: Early Echoes of Self-Evolution", "url": "https://www.minimax.io/news/minimax-m27-en", "date": "2026-03-18", "dateLabel": "博客发布", "category": "overview", "methodType": "", "priority": "K", "contentType": "blog", "references": { "main": { "label": "Building an agent for model self-evolution", "url": "https://www.minimax.io/news/minimax-m27-en" } }, "tldr": { "gap": { "value": "模型研发涉及数据、训练和工程协作，人工研究吞吐限制迭代速度。", "refs": [ "main" ] }, "position": { "value": "官方研发案例：内部 M2.7 参与研究工作，并更新技能、记忆和研究 agent 的运行框架。", "refs": [ "main" ] }, "conclusion": { "value": "介绍实验监控、日志分析、代码修复与框架迭代；人类仍提供方向和关键决策，属于研发流程报告。", "refs": [ "main" ] } }, "dimensions": { "author": { "value": "作者／发布者：MiniMax。", "refs": [ "main" ] }, "scope": { "value": "官方研发案例：内部 M2.7 参与研究工作，并更新技能、记忆和研究 agent 的运行框架。", "refs": [ "main" ] }, "reading": { "value": "读研究 agent 的具体工作流，区分它参与自身研发和完全自主递归；产品 benchmark 分数不能单独归因于自进化。", "refs": [ "main" ] }, "evidence": { "value": "依据 Building an agent for model self-evolution；简介依据文章原文；文中研究判断与实验宣称归于作者。", "refs": [ "main" ] } }, "tags": [ "ResearchAutomation" ], "links": [ { "label": "Shiyu 的相关解读", "url": "https://shiyunee.github.io/blogs/RSI-What-Evolves/" } ], "visibleKeys": [ "author", "reading" ], "reviewedAt": "2026-09-16", "reviewStatus": "source-reviewed", "attributions": [ { "tag": "org:minimax", "label": "MiniMax", "kind": "institution", "sources": [ { "label": "原文署名／发布机构", "url": "https://www.minimax.io/news/minimax-m27-en" } ] } ] }, { "id": "apodex-10-blog", "title": "Apodex-1.0", "url": "https://www.apodex.com/blog/apodex-1.0", "date": "2026-06-08", "dateLabel": "博客发布", "category": "overview", "methodType": "", "priority": "K", "contentType": "blog", "references": { "main": { "label": "Scaling Reasoning with an Agent Team；Where Apodex Stands", "url": "https://www.apodex.com/blog/apodex-1.0" } }, "tldr": { "gap": { "value": "长期研究任务需要检索、工具和验证，单条执行链与单个上下文难以承载全部工作。", "refs": [ "main" ] }, "position": { "value": "介绍并行研究 agent 团队与统一验证的运行设计，对比单 agent 与 heavy-duty 模式。", "refs": [ "main" ] }, "conclusion": { "value": "orchestrator 分配任务，子 agent 并行研究，global verifier 审核证据；这主要说明 harness 设计，不能单凭自检证明跨轮框架进化。", "refs": [ "main" ] } }, "dimensions": { "author": { "value": "作者／发布者：Apodex。", "refs": [ "main" ] }, "scope": { "value": "介绍并行研究 agent 团队与统一验证的运行设计，对比单 agent 与 heavy-duty 模式。", "refs": [ "main" ] }, "reading": { "value": "适合对照“验证答案”和“修改 harness”的区别；看团队协作结构与同系列模式比较。", "refs": [ "main" ] }, "evidence": { "value": "依据 Scaling Reasoning with an Agent Team；Where Apodex Stands；简介依据文章原文；文中研究判断与实验宣称归于作者。", "refs": [ "main" ] } }, "tags": [ "ResearchAutomation" ], "links": [ { "label": "Shiyu 的相关解读", "url": "https://shiyunee.github.io/blogs/RSI-What-Evolves/" } ], "visibleKeys": [ "author", "reading" ], "reviewedAt": "2026-09-16", "reviewStatus": "source-reviewed" }, { "id": "weng-prompt-engineering", "title": "Prompt Engineering", "url": "https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/", "date": "2023-03-15", "dateLabel": "博客发布", "category": "overview", "methodType": "", "priority": "R", "contentType": "blog", "references": { "main": { "label": "正文目录：提示、示例与推理方法", "url": "https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/" } }, "tldr": { "gap": { "value": "模型输出对输入组织敏感，需要区分不同提示方法及它们的适用条件。", "refs": [ "main" ] }, "position": { "value": "提示工程背景读物，为理解 harness 中的 prompt 组件提供基础。", "refs": [ "main" ] }, "conclusion": { "value": "梳理提示与推理方法；属于自进化研究的前置知识，不应把使用提示技巧直接称为 RSI。", "refs": [ "main" ] } }, "dimensions": { "author": { "value": "作者／发布者：Lilian Weng。", "refs": [ "main" ] }, "scope": { "value": "提示工程背景读物，为理解 harness 中的 prompt 组件提供基础。", "refs": [ "main" ] }, "reading": { "value": "作为 Weng harness 博客的背景链接阅读，帮助理解提示设计与让系统自动更新提示的区别。", "refs": [ "main" ] }, "evidence": { "value": "依据 正文目录：提示、示例与推理方法；简介依据文章原文；文中研究判断与实验宣称归于作者。", "refs": [ "main" ] } }, "tags": [ "Prompt" ], "links": [ { "label": "Shiyu 的相关解读", "url": "https://shiyunee.github.io/blogs/RSI-What-Evolves/" } ], "visibleKeys": [ "author", "reading" ], "reviewedAt": "2026-09-16", "reviewStatus": "source-reviewed", "attributions": [ { "tag": "person:lilian-weng", "label": "Lilian Weng", "kind": "scholar", "sources": [ { "label": "原文署名／发布机构", "url": "https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/" } ] } ] }, { "id": "2609.20519", "title": "SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness", "date": "2026-09-17", "dateLabel": "论文首发", "category": "methods", "methodType": "harness", "contentType": "paper", "priority": "C", "url": "https://arxiv.org/abs/2609.20519", "references": { "main": { "label": "§2.1–2.5、§3.1–3.3、§5.1", "url": "https://arxiv.org/html/2609.20519v1" }, "blog": { "label": "作者博客：Humans set priors / Results / Recursive Efficient Improvement", "url": "https://nvlabs.github.io/SoL-Pi/" }, "heat": { "label": "Hugging Face：2026-09-21 热度快照", "url": "https://huggingface.co/papers/2609.20519" } }, "tldr": { "gap": { "value": "长任务反复传递上下文和工具输出，成本不断累积；局部省 token 也可能损害任务完成率。", "refs": [ "main" ] }, "position": { "value": "研究自动搜索能否找到跨任务、跨模型复用的 harness 效率改进，而非只提高搜索题得分。", "refs": [ "main" ] }, "conclusion": { "value": "保留动作合并、上下文压缩、输出归档和证据摘要四种机制。完整组合保留约 94% 的 Pi 得分，API 成本约降三分之一；不是无损提效。", "refs": [ "main" ] } }, "dimensions": { "object": { "value": "修改 Pi 的执行、上下文和工具输出处理代码；模型参数与外层研究程序固定。", "refs": [ "main" ] }, "executor": { "value": "主要对照 GPT-5.6 Sol 与 Opus 5；群体实验由 Sol 协调、20 个 Luna worker 执行。", "refs": [ "main" ] }, "modifier": { "value": "研究 agent 提案、实现并独立审查；论文未逐角色列型号。人类先筛方向，后审阅并整理成功机制代码。", "refs": [ "main", "blog" ] }, "verdict": { "value": "开发用隐藏回归测试或自建可执行 verifier 判成功，再比较 token／成本并检查能力下限；不是模型自称完成。", "refs": [ "main" ] }, "seed": { "value": "Pi：已有模型调用、文件／命令工具和原生上下文压缩的 coding harness；四种机制作为扩展加入。", "refs": [ "main" ] }, "cycle": { "value": "152 个方向各自实现、测试、保留或淘汰；临时修改的研究调度代码不继承。", "refs": [ "main" ] }, "train": { "value": "无参数训练。搜索用 495 个 GitHub issue–PR 仓库环境及 40 个自建 verifier 环境，共 535 个。", "refs": [ "main" ] }, "debug": { "value": "开发环境反馈驱动修改；冻结候选后用 EdgeBench 11 题单向验收，失败不再反馈给搜索。", "refs": [ "main" ] }, "test": { "value": "EdgeBench 另 40 题留作最终泛化，主表汇报全部 51 题；另测 Terminal-Bench 4 的 63 题、IMO 2026 六题（Lean 4 验证）及 kernel 优化。", "refs": [ "main" ] }, "isolation": { "value": "不能把主表全部 51 题都称为最终未参与选择的测试；Performance 配置还按各模型最高单机制得分挑选，与预先固定的完整组合有别。", "refs": [ "main" ] }, "limits": { "value": "Terminal-Bench 4 完成 15 题，Pi 为 18 题；更低成本不等于更高成功率。多代递归提效仍是后续设想。", "refs": [ "main" ] }, "evidence": { "value": "HF 页面 95 赞（2026-09-21）；关注度快照不代表独立复现或同行评审。", "refs": [ "heat" ] } }, "visibleKeys": [ "object", "executor", "modifier", "verdict", "seed" ], "tags": [ "HarnessCode", "Context", "Tool", "SeparateEvolver", "ExecutableVerifier", "Update20260921" ], "links": [ { "label": "作者博客", "url": "https://nvlabs.github.io/SoL-Pi/" } ], "reviewedAt": "2026-09-21", "attributions": [ { "tag": "org:nvidia", "label": "NVIDIA", "kind": "institution", "sources": [ { "label": "§2.1–2.5、§3.1–3.3、§5.1", "url": "https://arxiv.org/html/2609.20519v1" } ] }, { "tag": "org:mit", "label": "MIT", "kind": "institution", "sources": [ { "label": "§2.1–2.5、§3.1–3.3、§5.1", "url": "https://arxiv.org/html/2609.20519v1" } ] }, { "tag": "person:song-han", "label": "Song Han", "kind": "scholar", "sources": [ { "label": "§2.1–2.5、§3.1–3.3、§5.1", "url": "https://arxiv.org/html/2609.20519v1" } ] } ], "priorityBasis": "本期阅读推荐；依据主题相关性、作者来源和公开关注信号，不作为质量排名。" }, { "id": "2609.19656", "title": "Self-Evolving Search Index", "date": "2026-09-17", "dateLabel": "论文首发", "category": "methods", "methodType": "harness", "contentType": "paper", "priority": "K", "url": "https://arxiv.org/abs/2609.19656", "references": { "main": { "label": "§3–4；附录 A.1–A.2、B.1–B.5", "url": "https://arxiv.org/html/2609.19656v1" }, "heat": { "label": "Hugging Face：2026-09-21 热度快照", "url": "https://huggingface.co/papers/2609.19656" } }, "tldr": { "gap": { "value": "固定索引无法适应不同检索需求，失败后通常还要人手动诊断和重写表示。", "refs": [ "main" ] }, "position": { "value": "让索引自动发现检索缺口、生成练习查询并修订索引键，检验对搜索和记忆读取的帮助。", "refs": [ "main" ] }, "conclusion": { "value": "变化的是文档／记忆的检索表示，不是模型参数。逐次核验新键是否忠于原文、能检索回原文并减少混淆。", "refs": [ "main" ] } }, "dimensions": { "object": { "value": "改文档或记忆条目的索引键，即供检索匹配的文本表示；原始内容和下游 agent 固定。", "refs": [ "main" ] }, "executor": { "value": "检索器为 BM25、BGE-Large、Qwen3-Embedding-8B；搜索用 GPT-OSS-120B、GPT-5.4-nano、Gemini-3.7-Flash、Kimi-K2.5；记忆用 Qwen3.5-9B。", "refs": [ "main" ] }, "modifier": { "value": "Qwen3.6-35B-A3B 同时承担索引修改、模拟查询和忠实性／可回答性评审。", "refs": [ "main" ] }, "verdict": { "value": "进化反馈：同一 Qwen 按原文给忠实性／可回答性打 0–3 分，至少 2 分通过；另用检索排名和键间相关度检查区分性。", "refs": [ "main" ] }, "seed": { "value": "在已有检索器上加索引修改循环；搜索／记忆沿用官方运行系统，替换索引但仍向 reader 返回原文。", "refs": [ "main" ] }, "cycle": { "value": "诊断失败键→修订→模型与检索规则验收；每文档最多 10 键，含固定原文键。", "refs": [ "main" ] }, "train": { "value": "不训练参数。在各基准语料／历史记忆上模拟查询，20 轮共 2,560 条查询用于索引进化；不使用正式评测问题。", "refs": [ "main" ] }, "debug": { "value": "按忠实性与检索规则接受局部修改；报告固定第 20 轮，不按测试最高分挑版本。", "refs": [ "main" ] }, "test": { "value": "BRIGHT、Spider 2.0、FIBEN、BEAVER 用标注相关性算 nDCG@10；BrowseComp-Plus 830 题由 Qwen3-32B 对参考答案判分；LongMemEval-V2 Small 451 题用规则与 GPT-5.2 medium 判分。", "refs": [ "main" ] }, "isolation": { "value": "隔离的是评测查询；索引允许访问将被检索的语料或历史记忆，并非语料也完全不可见。", "refs": [ "main" ] }, "evidence": { "value": "HF 页面 46 赞（2026-09-21）；属于检索与记忆组件自改进的相关路线。", "refs": [ "heat" ] } }, "visibleKeys": [ "object", "executor", "modifier", "verdict", "seed" ], "tags": [ "MemoryContent", "MemoryMechanism", "Context", "SeparateEvolver", "LLMJudge", "GoldLabel", "Update20260921" ], "links": [], "reviewedAt": "2026-09-21", "attributions": [], "priorityBasis": "本期阅读推荐；依据主题相关性、作者来源和公开关注信号，不作为质量排名。" }, { "id": "2609.17523", "title": "ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents", "date": "2026-09-15", "dateLabel": "论文首发", "category": "methods", "methodType": "joint", "contentType": "paper", "priority": "K", "url": "https://arxiv.org/abs/2609.17523", "references": { "main": { "label": "§2.2–2.4、§4.2–4.4、附录 §7.1–7.3 / 表1–2", "url": "https://arxiv.org/html/2609.17523v1" }, "heat": { "label": "Hugging Face：2026-09-21 热度快照", "url": "https://huggingface.co/papers/2609.17523" } }, "tldr": { "gap": { "value": "科研对话产生了需求、纠错和执行证据，但如果不转成后续可用的任务与反馈，系统不会持续改进。", "refs": [ "main" ] }, "position": { "value": "研究科研协作能否同时推动工作方式和模型学习：内层改 harness，外层在选定 harness 下做强化学习。", "refs": [ "main" ] }, "conclusion": { "value": "三轮联合实验报告单次答对率由 42.2% 到 73.3%。另做固定模型／固定 harness 对照；单独 harness 实验只改指令和 scoped skill。", "refs": [ "main" ] } }, "dimensions": { "object": { "value": "联合实验更新 harness 与 Qwen 参数；单独 harness 实验只改指令／技能文本，执行循环、工具和上下文处理固定。", "refs": [ "main" ] }, "executor": { "value": "Qwen3.5-4B 执行科研任务；参数学习后使用对应更新的检查点。", "refs": [ "main" ] }, "modifier": { "value": "GPT-6 Astra 诊断并编辑 harness；Qwen3.8-27B 模拟受限用户反馈并解释反馈，训练程序执行 GRPO。", "refs": [ "main" ] }, "verdict": { "value": "harness：私有 verifier 给对错／提交状态，模拟用户只回复允许的确认或检查要求。参数：按固定 rubric 汇总可执行检查与固定 judge 分数，不用模拟用户的三值诊断分替代奖励。", "refs": [ "main" ] }, "seed": { "value": "Biomni 科研工具／数据基础上的单文件 harness；已有 Python REPL、数据读取与答案提交，初始没有新增指令／技能。", "refs": [ "main" ] }, "cycle": { "value": "联合实验每轮先搜索 harness 10 步，再 RL 更新 20 步，共三轮；单独 harness 为 288 次对话、24 次更新。", "refs": [ "main" ] }, "train": { "value": "LAB-Bench／Biomni-Eval1 的 LitQA2、DbQA、ProtocolQA、GWAS；任务库存 895 题不等于训练样本数。RL 使用当前模型新生成轨迹。", "refs": [ "main" ] }, "debug": { "value": "联合实验用固定 validation 选 harness；单独 harness 用 adaptation 选版本，其 validation 仅供事后对照。", "refs": [ "main" ] }, "test": { "value": "四类科学任务上比较联合系统初末版本的单次正确率；固定 harness 的模型对照另用 pass@4。附录列库存，未逐项给出各实验 split 数量。", "refs": [ "main" ] }, "isolation": { "value": "论文声明开发、策略训练、最终测试分离；未完整披露逐任务判分代码及解释型 judge 型号，不能将整套判分笼统写成规则匹配。", "refs": [ "main" ] }, "evidence": { "value": "9 月 15 日首发，本期补录；HF 30 赞（2026-09-21），不算 9 月 16 日后新论文。", "refs": [ "heat", "main" ] } }, "visibleKeys": [ "object", "executor", "modifier", "verdict", "seed" ], "tags": [ "Skill", "Prompt", "Weights", "JointEvolution", "SeparateEvolver", "LLMJudge", "ExecutableVerifier", "Update20260921" ], "links": [], "reviewedAt": "2026-09-21", "attributions": [ { "tag": "org:phai-labs", "label": "PhAI Labs", "kind": "institution", "sources": [ { "label": "§2.2–2.4、§4.2–4.4、附录 §7.1–7.3 / 表1–2", "url": "https://arxiv.org/html/2609.17523v1" } ] } ], "priorityBasis": "本期阅读推荐；依据主题相关性、作者来源和公开关注信号，不作为质量排名。" }, { "id": "2609.19203", "title": "Position: It is Time to Virtualize Foundation Models with a Self-evolving Operating System Layer", "date": "2026-09-16", "dateLabel": "论文首发", "category": "overview", "methodType": "", "contentType": "paper", "priority": "K", "url": "https://arxiv.org/abs/2609.19203", "references": { "main": { "label": "引言、§3–4、附录 A–C、作者机构", "url": "https://arxiv.org/html/2609.19203v1" } }, "tldr": { "gap": { "value": "各 agent 框架重复实现记忆、预算与约束，行为难以迁移，跨系统治理也脆弱。", "refs": [ "main" ] }, "position": { "value": "提出 FMOS：像操作系统管理硬件一样，统一管理模型调用、记忆、资源分配和验证。", "refs": [ "main" ] }, "conclusion": { "value": "主张基于运行经验调整系统策略，让改进跨应用复用；是一篇架构立场论文，不是已跑通多代 RSI 的实证。", "refs": [ "main" ] } }, "dimensions": { "scope": { "value": "基础模型之上的可学习系统层：知识管理、模型选择、计算预算、验证与策略执行。", "refs": [ "main" ] }, "reading": { "value": "结合 harness 服务化阅读；区分作者希望实现的系统能力与已经展示的实验。", "refs": [ "main" ] }, "framework": { "value": "用统一接口隔离应用与底层模型，联合调配知识、计算与验证资源。", "refs": [ "main" ] }, "limits": { "value": "不能从“self-evolving OS”的定位推断出统一的训练／测试数据集或成熟自主闭环。", "refs": [ "main" ] } }, "visibleKeys": [ "scope", "reading" ], "tags": [ "RSIRoadmap", "Infrastructure", "Update20260921" ], "links": [], "reviewedAt": "2026-09-21", "attributions": [ { "tag": "org:hpe", "label": "Hewlett Packard Enterprise", "kind": "institution", "sources": [ { "label": "引言、§3–4、附录 A–C、作者机构", "url": "https://arxiv.org/html/2609.19203v1" } ] }, { "tag": "org:uchicago", "label": "University of Chicago", "kind": "institution", "sources": [ { "label": "引言、§3–4、附录 A–C、作者机构", "url": "https://arxiv.org/html/2609.19203v1" } ] }, { "tag": "person:ian-foster", "label": "Ian Foster", "kind": "scholar", "sources": [ { "label": "引言、§3–4、附录 A–C、作者机构", "url": "https://arxiv.org/html/2609.19203v1" } ] } ], "priorityBasis": "本期阅读推荐；依据主题相关性、作者来源和公开关注信号，不作为质量排名。" }, { "id": "anthropic-when-ai-builds-itself", "title": "When AI builds itself", "date": "2026-09-18", "dateLabel": "博客更新", "category": "overview", "methodType": "", "contentType": "blog", "priority": "C", "url": "https://www.anthropic.com/institute/recursive-self-improvement", "references": { "main": { "label": "Evidence from within Anthropic；Update 9/18/2026；图注与脚注", "url": "https://www.anthropic.com/institute/recursive-self-improvement" } }, "tldr": { "gap": { "value": "代码与实验量增长未必意味着模型已经能独立决定研究方向，需要看内部研发中实际承担了哪些工作。", "refs": [ "main" ] }, "position": { "value": "Anthropic 用研发记录、员工调查和会话评审介绍 AI 如何参与构建后续模型；本期收录 9 月 18 日更新。", "refs": [ "main" ] }, "conclusion": { "value": "更新给出按任务难度分组的会话成功率；由 Claude judge 判断是否完成且无需纠正。文章明确尚未实现自主构建后继模型。", "refs": [ "main" ] } }, "dimensions": { "scope": { "value": "模型参与 AI 研发的产业证据；涵盖编码、实验执行与研究下一步选择。", "refs": [ "main" ] }, "reading": { "value": "先读内部证据与图注，再读限制：代码量不等于生产率，选择过的会话样本不能代表一般人机能力。", "refs": [ "main" ] }, "feedback": { "value": "会话成功由 Claude judge 判断；训练代码优化另有正确性检查和计时，不能把两种指标合成一个成功率。", "refs": [ "main" ] }, "limits": { "value": "本条日期是页面明确标注的 9 月 18 日更新日，不是首次发布日期；公司自报与观察性结果不等于独立因果验证。", "refs": [ "main" ] } }, "visibleKeys": [ "scope", "reading" ], "tags": [ "IndustryReport", "ResearchAutomation", "Update20260921" ], "links": [], "reviewedAt": "2026-09-21", "attributions": [ { "tag": "org:anthropic", "label": "Anthropic", "kind": "institution", "sources": [ { "label": "Evidence from within Anthropic；Update 9/18/2026；图注与脚注", "url": "https://www.anthropic.com/institute/recursive-self-improvement" } ] } ], "priorityBasis": "本期阅读推荐；依据主题相关性、作者来源和公开关注信号，不作为质量排名。" }, { "id": "ibm-rsi-serious-question", "title": "Why recursive self-improvement suddenly became a serious question", "date": "2026-09-16", "dateLabel": "博客发布", "category": "overview", "methodType": "", "contentType": "blog", "priority": "K", "url": "https://www.ibm.com/think/news/why-recursive-self-improvement-ai-serious-question", "references": { "main": { "label": "What changed；Testing pieces of the idea；The safety gap；Still an open question", "url": "https://www.ibm.com/think/news/why-recursive-self-improvement-ai-serious-question" } }, "tldr": { "gap": { "value": "“AI 帮忙研发”与“每代自主制造更强后继”常被混为一谈，容易误读当前进展。", "refs": [ "main" ] }, "position": { "value": "IBM Think 采访研究者，对照 coding agents、AIDE² 与前沿实验室的公开说法。", "refs": [ "main" ] }, "conclusion": { "value": "强调已有局部改进循环，但完整自主递归仍缺验证；讨论错误奖励被放大、人类审核和计算成本的约束。", "refs": [ "main" ] } }, "dimensions": { "scope": { "value": "RSI 现状与证据边界的采访型入门文章，作者 Sascha Brodsky。", "refs": [ "main" ] }, "reading": { "value": "适合与 AIDE² 原始报告、Anthropic 内部证据对读，分清受访者判断和实验结果。", "refs": [ "main" ] }, "evidence": { "value": "IBM／Brown University／Weco 受访者的一手观点；论文与实验细节仍应回到各原始报告。", "refs": [ "main" ] }, "limits": { "value": "这是新闻采访和观点汇总，没有提出新的统一 benchmark 或自主训练方法。", "refs": [ "main" ] } }, "visibleKeys": [ "scope", "reading" ], "tags": [ "IndustryReport", "RSIRoadmap", "Update20260921" ], "links": [], "reviewedAt": "2026-09-21", "attributions": [ { "tag": "org:ibm", "label": "IBM", "kind": "institution", "sources": [ { "label": "What changed；Testing pieces of the idea；The safety gap；Still an open question", "url": "https://www.ibm.com/think/news/why-recursive-self-improvement-ai-serious-question" } ] } ], "priorityBasis": "本期阅读推荐；依据主题相关性、作者来源和公开关注信号，不作为质量排名。" }, { "id": "sol-pi-author-blog", "title": "SoL-Pi: Scaling Auto-Research Loops for Efficient Agent Harnesses", "date": "2026-09-21", "dateLabel": "收录日期（配套博客未标首发）", "category": "overview", "methodType": "", "contentType": "blog", "priority": "K", "url": "https://nvlabs.github.io/SoL-Pi/", "references": { "main": { "label": "Method；Discoveries；Results；Humans set priors；Recursive Efficient Improvement", "url": "https://nvlabs.github.io/SoL-Pi/" } }, "tldr": { "gap": { "value": "长时间运行的 agent 会反复付出上下文与工具输出成本，单看完成率看不出这种浪费。", "refs": [ "main" ] }, "position": { "value": "SoL-Pi 作者的图解与工程说明，解释四种机制如何省成本及研究中人类实际做了什么。", "refs": [ "main" ] }, "conclusion": { "value": "人类提供早期方向、筛选搜索空间并整理成功代码；用更省钱的 harness 改进下一代仍是未来计划，而非已经验证的复利。", "refs": [ "main" ] } }, "dimensions": { "scope": { "value": "论文的配套解释，重点是机制设计、人工参与和成本／得分取舍。", "refs": [ "main" ] }, "reading": { "value": "先看四种机制图解，再读 Results 和 Humans set priors；严格的数据隔离以论文 §2.5 为准。", "refs": [ "main" ] }, "limits": { "value": "页面未明确给出首发日期，因此按 9 月 21 日收录日记录；与同名论文属于同一工作，不算第二项独立研究。", "refs": [ "main" ] } }, "visibleKeys": [ "scope", "reading" ], "tags": [ "ResearchAutomation", "Update20260921" ], "links": [], "reviewedAt": "2026-09-21", "attributions": [ { "tag": "org:nvidia", "label": "NVIDIA", "kind": "institution", "sources": [ { "label": "Method；Discoveries；Results；Humans set priors；Recursive Efficient Improvement", "url": "https://nvlabs.github.io/SoL-Pi/" } ] } ], "priorityBasis": "本期阅读推荐；依据主题相关性、作者来源和公开关注信号，不作为质量排名。" } ]