[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-baai-arex-2-self-improving-agent":3,"topics-all":38,"news-related-20db9794-7f34-4a8c-8271-ec5d646d311e":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},"20db9794-7f34-4a8c-8271-ec5d646d311e","智源开源AREX-2:27B把自我改进循环装进模型","智源研究院开源 27B 模型 AREX-2:微调自 Qwen3.8-27B,把自我改进循环从外部脚手架移进模型本体,自己提出、度量、反思、修正,权重宽松开源。模型卡评测里,MLE-Lite 拿 81.8 分,压过表内最强闭源模型的 72.7;深度研究零新增数据上台阶,改进只发生在测试时,本周在海外社区走红。","一个 27B 参数的开源模型,在机器学习工程基准 MLE-bench Lite 上拿下 81.8 分,把评测表里的全部闭源旗舰——GPT-5.6 Sol、Claude Opus 4.8、GPT-5.5——压在身后;而它的参数量,大约只有 2.8T 的 Kimi-K3 的百分之一。这就是智源研究院(BAAI)的 AREX-2:Apache 2.0 许可,权重挂在 [Hugging Face](https:\u002F\u002Fhuggingface.co\u002FBAAI\u002FAREX-2),论文编号 2609.38288 九月就已挂出,本周却在海外开发者社区集中走红——daily.dev 的解读帖拿下 538 个 upvote,知名 AI 博主 Sam Witteveen 专门出了视频。\n\n## 循环本来不在模型里\n\n论文开篇就把行业现状挑明了:在大多数 agent 系统里,重试循环主要实现在脚手架(scaffold)上——harness 决定什么时候重试、保留什么,模型只负责一次产出一个尝试。AREX-2 把问题反了过来:怎么把循环搬进模型本体。模型自己跑「提出—度量—反思—修正」的多轮循环,而且轮数预算和墙钟时间会直接告诉它,量级是数小时、每个任务数百次工具调用。论文里的细节很有意思:知道预算有多大,模型会先建立 baseline,用早期轮次做度量,把修正留到后期——而不是一口气交出最好的猜测。这就是它对自我改进的定义:在单个任务内,把更多轮次变成更好的解,本质是任务内的 test-time scaling。\n\n## 训练数据:保留失败的轨迹\n\n方法层最有味道的是数据构造。教师模型把 GitHub 仓库和在线判题题库编译成可执行环境(任务加打分函数);agent 在环境里多轮迭代;轨迹按整体取舍——最终分数高、过程合规就整条保留,不要求每一轮都成功。失败的运行、倒退、被放弃的路线全部留在训练数据里,论文的理由很直接:留着它们,模型才学得会受挫之后继续迭代。训练本身,是在这些轨迹与 AREX 既有深度研究数据的混合上微调 Qwen3.8-27B——论文写的词是 fine-tune,方法部分没有强化学习的影子;daily.dev 的解读也点破了这层窗户纸:发布文案里写着 reinforcing,实际做的是监督微调。要强调的是,这种自我改进只发生在测试时——权重在任务之间不变,离递归自我改进还差把轨迹喂回训练那一步,论文结论里把这一步明明白白列为下一步。\n\n## 跑分要连着口径一起读\n\n按模型卡:MLE-bench Lite(三种子平均的 Any Medal)81.8,是对比表内最高——GPT-5.6 Sol 72.7、Kimi-K3(2.8T)72.7、Naive-N0.5-Flash(309B)73.7、Claude Opus 4.8 63.6;Frontier-CS(188 任务 Agent Track)70.7 是开源阵营最高,但低于 Sol 的 76.4、Opus 的 74.5、GPT-5.5 的 72.1。深度研究侧没加任何新训练数据:BrowseComp 84.0、HLE 52.6(text-only 子集口径)、GAIA 92.2、DeepSearchQA 93.8——GAIA 这项超过表内全部闭源模型(最高 GPT-5.5 的 87.4),但同表 35B 的 XYZ-Aquila-mini 还有 97.1,谈碾压为时过早。消融更能说明机制:基座裸跑 28.8,加通用技能到 41.2,任务特定技能加更多轮次(仍未训练)到 68.2,换成训练后的 AREX-2 到 81.8——训练净贡献 13.6 分;训练后的模型用更少轮次反超未训练多轮版 7.6 分;轮次继续加,还能再涨 6.0 分,Frontier-CS 上一路涨到第五个小时。\n\n## 三个判断\n\n第一,长程任务上参数量不是瓶颈,轨迹数据的形态才是——2.8T 的 Kimi-K3 与 1.6T 的 DeepSeek-V4-Pro(54.5)都被 27B 甩开,差的是整条改进轨迹这种数据,不是算力堆出来的参数。第二,这是对外挂 harness 路线的一次正面叫板:平台九月刚报道过 DeepSeek 把 Agent 拆成插件栈的外挂式 Harness,智源这波把循环直接焊进权重——循环到底住在 harness 里还是模型里,两条路线现在都有人重注押注。第三,别把 test-time 自我改进传话成 RSI:权重不变,任务之间不积累;论文自己说,让模型自己的轨迹成为下一份训练数据,才是递归的起点。\n\n复现口径也放得很开:[VectorSpaceLab\u002FAREX-2](https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-2) 仓库(49 star、80 commits)给了六个赛道的评测 runner 和数据集定义,深度研究四项基准共享同一套默认设置。对做 agent 的人来说,该问的问题正在换:从「我的 harness 写得好不好」,变成「我的模型会不会自己变好」。","https:\u002F\u002Fhuggingface.co\u002FBAAI\u002FAREX-2","31482a3a-8cde-4b8a-b0ae-1f0c0358eb06",[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},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"42311b64-8765-4335-a954-aaa838d45ba8","en","BAAI’s AREX-2: 27B Model Runs Its Own Improvement Loop","BAAI open-sources AREX-2, a 27B fine-tune of Qwen3.8-27B that runs its own improvement loop. Card: MLE-bench Lite 81.8 over GPT-5.6 Sol 72.7.","A 27B-parameter open-weight model has taken the top spot on a machine-learning engineering benchmark, outscoring every closed flagship in its comparison table — GPT-5.6 Sol, Claude Opus 4.8, GPT-5.5 — while holding roughly one hundredth of the parameters of the 2.8T Kimi-K3. The model is AREX-2 from the Beijing Academy of Artificial Intelligence (BAAI): Apache 2.0, weights on [Hugging Face](https:\u002F\u002Fhuggingface.co\u002FBAAI\u002FAREX-2), paper 2609.38288 first posted in September, and this week it went viral across overseas developer communities — a daily.dev explainer picked up 538 upvotes and Sam Witteveen published a dedicated video.\n\n## The loop was never in the model\n\nThe paper opens by naming the industry default: in most agent systems the retry loop lives largely in the scaffold. The harness decides when to retry and what to keep; the model mostly produces one attempt at a time. AREX-2 flips the question: how do you move the loop inside the model? The model runs its own propose–measure–reflect–revise loop over multiple test-time rounds, and it is told its budget of rounds and wall-clock time up front — on the order of hours and hundreds of tool calls per task. The detail matters: knowing the budget changes how the agent works. It establishes a baseline first, spends early rounds measuring, and saves revisions for later, instead of submitting its best guess at once. That is the paper’s definition of self-improvement: turning more rounds on one task into a better solution — test-time scaling within a single task.\n\n## Training data that keeps the failures\n\nThe data construction is the tasteful part. A teacher model compiles GitHub repositories and online-judge problems into executable environments — a task plus a scoring function. An agent then works each task over many rounds, and trajectories are selected whole: kept when the final score is high and the process meets requirements, with no requirement that any individual round succeed. Failed runs, regressions, and abandoned approaches all stay in the training data, on the explicit logic that keeping them is what teaches the model to keep iterating after a setback. Training itself fine-tunes Qwen3.8-27B on a mixture of these trajectories and the existing AREX deep-research data — the paper says fine-tune, and its method section contains no reinforcement learning; daily.dev’s write-up calls out the same gap between the word “reinforcing” in launch language and the supervised fine-tuning actually used. And note the boundary: this self-improvement is test-time only. Weights do not change between tasks; recursive self-improvement would require feeding trajectories back into training, which the paper itself lists as the next step.\n\n## Read the scores together with their caveats\n\nPer the model card: MLE-bench Lite (Any Medal averaged over three seeds) at 81.8 is the highest in its comparison table — GPT-5.6 Sol 72.7, Kimi-K3 (2.8T) 72.7, Naive-N0.5-Flash (309B) 73.7, Claude Opus 4.8 63.6. Frontier-CS (the 188-task Agent Track) at 70.7 is the best among open weights but sits below Sol’s 76.4, Opus 4.8’s 74.5, and GPT-5.5’s 72.1. On deep research, with no new training data added: BrowseComp 84.0, HLE 52.6 (text-only subset), GAIA 92.2, DeepSearchQA 93.8 — the GAIA number beats every closed model in the table (GPT-5.5’s 87.4 is the frontier best), though the 35B XYZ-Aquila-mini still posts 97.1 in the same table, so talk of crushing would be premature. The ablation explains the mechanism best: the base model alone scores 28.8; general skills lift it to 41.2; task-specific skills plus more rounds — still untrained — reach 68.2; swapping in the trained AREX-2 reaches 81.8, a net training contribution of 13.6 points. The trained model with fewer rounds beats the untrained one with more rounds by 7.6 points, and more rounds still buy another 6.0 — on Frontier-CS the curve keeps rising through the fifth hour.\n\n## Three takeaways\n\nFirst, on long-horizon tasks parameter count is not the bottleneck — trajectory data shape is. The 2.8T Kimi-K3 and the 1.6T DeepSeek-V4-Pro (54.5) are both left behind by a 27B model; what differs is whole-improvement-trajectory data, not parameters bought with compute. Second, this is a direct challenge to the external-harness route: this platform covered DeepSeek’s pluggable external Harness in September, and BAAI now welds the loop into the weights — whether the loop lives in the harness or in the model is now a two-sided bet. Third, do not relay test-time self-improvement as RSI: weights are fixed, nothing accumulates across tasks, and the paper itself says the recursive version starts when the model’s own trajectories become its next training data.\n\nReproduction is unusually open: the [VectorSpaceLab\u002FAREX-2](https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-2) repository (49 stars, 80 commits) ships evaluation runners and dataset definitions for six tracks, with the four deep-research benchmarks sharing one default profile. For anyone building agents, the question is changing — from “is my harness good enough?” to “does my model improve itself?”","baai-arex-2-self-improving-agent","2026-11-07T19:20:00Z","2026-11-07T19:14:07.215361Z","2026-11-07T19:14:07.215376Z",true,"agent",29,[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},"26337a90-e049-44cb-9738-16d10a1fe84b","微软4B混合智能体:OSWorld 66%,点击还是敲代码","microsoft-mage-vl-cowork-hybrid-computer-use","2026-11-03T07:33:53+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"3719a369-508f-4fbe-96df-15f6fbb49f3b","Unlimited-OCR 2:把无限长文档塞进有限上下文","baidu-unlimited-ocr-2-sr-swa","2026-10-27T15:11:27+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"a9912456-9f4f-4a4c-b564-83a339402ced","达摩院开源RynnUnify:16B具身大脑先想象再动手","rynnunify-16b-unified-embodied-brain","2026-10-26T13:16:55+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"9597f154-5515-457e-9590-8b32347ea750","Nanbeige4.3-33B:33B教师模型自测80.9分","nanbeige-4-3-33b-open-agent-teacher-model","2026-10-26T01:15:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"7d735b5c-c7a8-4a8d-afac-abd3034170e7","Sage-A1.5开源:35B智能体GAIA 95.1","sage-a1-5-agentic-scientific-discovery","2026-10-24T23:10:02+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"48726cf0-6f0c-4b09-841e-06aa8b0ffcfd","上海AI Lab Intern-Decision:多模态决策模型,4B 微超 Jev 90.02%","intern-decision-qwen3-5-multimodal-decision-model","2026-10-18T11:13:06+00:00"]