[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-recursive-self-rewrite-terminal-trajectories":3,"topics-all":38,"news-related-d552a77b-3b52-4d59-8962-e897aa06b37c":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"d552a77b-3b52-4d59-8962-e897aa06b37c","腾讯RSR:把借来的轨迹改写成自己的能力","换一套执行 harness,同一个模型能多解出几道难题,但这些增益属于脚手架、不属于模型。腾讯混元联合马里兰大学等提出 RSR 框架,让 Qwen-3.8-27B 自扮三种角色,把 2001 条成功轨迹改写成 11094 条再训练,Terminal-Bench 2 通过率从 57.0% 升到 74.2%,权重已开源。","Terminal-Bench 2 上,Qwen-3.8-27B 的 pass@3 停在 57.0%。有意思的是,模型不动、换一套更强的执行 harness——就是那个管理观察、工具调用、验证、重试、停止的执行脚手架——任务通过数立刻往上走。麻烦在于,多出来的那部分增益记在 harness 账上,不属于模型本身:部署时脚手架一撤,成绩就还回去了。\n\n## 借来的经验要先「翻译」\n\n腾讯 HY LLM Frontier、马里兰大学等机构的 Recursive Self-Rewrite(RSR)框架做的就是这笔「翻译」。他们让同一个 Qwen-3.8-27B 在三套 harness(Terminus 2、Recursive Self-Reflect Terminus、StateM)下跑约 3000 个自建终端任务,三者合计解出 759 个,比最强的单套 harness 多 34.3%。但在专用 harness 下录制的轨迹混着大量控制器介入和工作流约定,直接拿去训练,模型会对外部拐杖产生依赖。\n\nRSR 的解法是把模型自己拆成三个角色:planner 把成功轨迹里的流程蒸馏成 runbook;critic 专门检查验证器和答案有没有泄漏,泄漏就拒绝、带着反馈递归重写;executor 在全新沙箱里按 runbook 走通用 harness 重做一遍。2001 条源轨迹最终扩写成 11094 条「去过拐杖」的轨迹。\n\n## 成绩单与那盆冷水\n\n拿这些轨迹做 SFT,Terminal-Bench 2 pass@3 从 57.0% 提到 74.2%;Terminal-Bench 4 从 1.5% 到 9.1%;自建的 Terminal-Bench Hard 从 39.0% 到 63.0%;Software Terminal-Bench 从 3.0% 到 6.0%;Long-Horizon 的过程奖励从 0.21 升到 0.29,同时超过基座模型和直接轨迹 SFT。\n\n两处冷水:一,最难的 Terminal-Bench 4 仍是个位数,长程过程奖励也只从 0.21 爬到 0.29;二,论文实验显示,不做改写的直接 SFT 虽然在部分基准上提升,却会因为循环行为在另一些基准上退化——反过来印证「改写」才是关键步骤,而不是简单堆数据。另外,RSR-27B 权重已按 Apache-2.0 挂上 Hugging Face,但 model card 的 README 目前还是空的,想复现的人得自己回去啃 16 页论文。\n\n## 和 Hyra 不是一回事\n\n看过腾讯混元 7 月发布的递归自我改进研究智能体 Hyra-1.0 的读者,可能觉得这两个名字像。区别在于:Hyra 是运行时靠递归自我改进的智能体产品,RSR 是训练期的数据工程方法——把改进花在 SFT 数据上,让经验直接沉淀进权重。一个追求边跑边进化,一个把经验折旧进模型。\n\n在评测 harness 越来越强的当下,榜单成绩越来越像「模型 + 脚手架」的合签分。RSR 给了条干净的路径:经验可以借,但改写到不带拐杖也能带走,才算模型自己的。论文见 [arXiv:2610.02826](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.02826),权重见 [IntelligenceLab\u002FRSR-27B](https:\u002F\u002Fhuggingface.co\u002FIntelligenceLab\u002FRSR-27B)。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.02826","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"d4279ea7-c3ac-4f2f-a047-80506234f004","en","Tencent's RSR Rewrites Borrowed Trajectories Into Model Skill","Borrowed harness gains don't stick. RSR rewrites 2,001 trajectories into 11,094 for Qwen-3.8-27B, lifting Terminal-Bench 2 pass@3 from 57.0% to 74.2%.","On Terminal-Bench 2, Qwen-3.8-27B sits at 57.0% pass@3. Swap in a stronger execution harness — the scaffolding that manages observations, tool calls, verification, retries and stopping — and the model suddenly solves more tasks. The catch: those extra gains are booked to the harness, not the model. Remove the scaffolding at deployment and the score goes back down.\n\n## Borrowed Experience Needs Translation First\n\nThat translation is exactly what Recursive Self-Rewrite (RSR), from Tencent HY LLM Frontier, the University of Maryland and other institutions, does. The team ran the same Qwen-3.8-27B under three harnesses (Terminus 2, Recursive Self-Reflect Terminus, StateM) across roughly 3,000 self-curated terminal tasks. Together the three solved 759 tasks, 34.3% more than the strongest single harness. But trajectories recorded under specialized harnesses are laced with controller interventions and workflow conventions; training on them naively teaches the model to lean on external crutches.\n\nRSR splits the model itself into three roles. A planner distills procedures from successful trajectories into runbooks. A critic screens for verifier and solution leakage — leaking candidates are rejected and recursively regenerated with feedback. An executor replays each qualified runbook in a fresh sandbox under a general harness. The result: 2,001 source trajectories expanded into 11,094 rewritten, crutch-free ones.\n\n## The Scorecard, and the Cold Water\n\nFine-tuning on those trajectories lifts Terminal-Bench 2 pass@3 from 57.0% to 74.2%, Terminal-Bench 4 from 1.5% to 9.1%, the self-curated Terminal-Bench Hard from 39.0% to 63.0%, Software Terminal-Bench from 3.0% to 6.0%, and the Long-Horizon process reward from 0.21 to 0.29 — beating both the base model and direct trajectory SFT.\n\nTwo buckets of cold water. First, on the hardest Terminal-Bench 4 the number is still single-digit, and the long-horizon process reward only crawls from 0.21 to 0.29. Second, the paper's experiments show that direct SFT without rewriting improves some benchmarks while degrading others through looping behaviors — which is precisely the counter-evidence that rewriting, not data volume, is the load-bearing step. Also note: RSR-27B weights are on Hugging Face under Apache-2.0, but the model card README is still empty — anyone aiming to reproduce will be reading the 16-page paper themselves.\n\n## Not the Same Thing as Hyra\n\nReaders who remember Hunyuan's July release of Hyra-1.0, the recursively self-improving research agent, may find the naming familiar. The difference: Hyra is an agent product that improves itself recursively at runtime; RSR is a training-time data-engineering method that spends the improvement on SFT data, letting experience settle into weights. One evolves while running; the other amortizes experience into the model.\n\nAs evaluation harnesses keep getting stronger, leaderboard scores increasingly read like a joint signature of model plus scaffolding. RSR offers a clean path: borrowing experience is fine, but it only counts as yours once it is rewritten to travel without crutches. Paper: [arXiv:2610.02826](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.02826); weights: [IntelligenceLab\u002FRSR-27B](https:\u002F\u002Fhuggingface.co\u002FIntelligenceLab\u002FRSR-27B).","recursive-self-rewrite-terminal-trajectories","2026-10-06T21:06:49Z","2026-10-06T21:06:52.818906Z","2026-10-06T21:06:52.818918Z",true,"agent",315,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"fead7184-42cc-4604-9aba-48dcb20ef7ab","阿里2852条长轨迹:喂出12基准全涨的27B智能体","terminalhorizon-long-trajectory-27b-agent","2026-10-13T21:08:01+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"c77381d9-29ba-45ff-89df-855d11f90de2","Terminal-Universe:Qwen 把旧轨迹反向重建为 3.73 万个环境,27B 微调双基准 +11.9\u002F+13.8 分","terminal-universe-trajectory-environments","2026-09-04T17:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"859242dd-0035-41e6-8db9-4583989048b7","AgentSWE基准:让编码智能体造智能体,无人及格","agentswe-coding-agent-benchmark","2026-10-13T19:08:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"feb04a1c-2c75-42d2-9b62-9c4f0585d924","AutoCompact:让编码代理自己学会何时压缩上下文","autocompact-learned-context-compaction","2026-10-04T17:10:39+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"206ea36a-1eea-463f-a241-1e3b32f5ec2d","USTC GraphForge:证据图把任务和 rubric 钉在一起,Qwen3.6-27B 涨三基准","graphforge-ustc-qwen36-27b-evidence-graph","2026-10-04T03:05:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"6f4d1046-ee70-4a06-9261-2cc187c66285","12 万美元 token 把 Copilot 运行时搬进 Rust:AI 智能体包揽 43 万行移植","copilot-runtime-rust-agentic-port","2026-09-20T19:11:22+00:00"]