[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-decision-model-actor-judge-teacher":3,"topics-all":38,"news-related-6c8ae96a-53e0-4ef7-8609-40b345cdee7c":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},"6c8ae96a-53e0-4ef7-8609-40b345cdee7c","决策模型实测:自己演砸42.9%,教学生82.9%","同一批任务上，决策模型JEV自己当智能体只完成42.9%，退到幕后当老师却把Qwen2.5-1.5B教到82.9%。浙大与百度的论文把决策模型拆成演员、裁判、教师三角色实测：演员被顺手的子目标带偏，教师看完整轨迹、按每步贡献打分后，GRPO训练在两个基准全面上涨，一轮调用费仅22到30美元。","同一个模型，同一批任务，换个角色，成绩单完全变样：决策模型 JEV 自己下场当智能体，在 ALFWorld 家务环境里只完成 42.9% 的任务；退到幕后给 Qwen2.5-1.5B 当老师，学生把成功率刷到 82.9%。这篇 11 月 12 日提交到 arXiv 的论文（2611.17167）来自浙江大学 ZJU-REAL 团队与百度，把最近很热的「决策模型」拆成演员、裁判、老师三个角色挨个实测，结论相当反直觉：演不好的模型，照样能教好。\n\n## 三重角色怎么定义\n\n决策模型是最近冒出来的新品类：输入问题和上下文，输出不是生成的长文本，而是带概率的选项，快且便宜。论文选了三个专用决策模型——TypeSafe AI 的 JEV、Perplexity 开源的 Decider v1.2、OpenAI 基于 GPT-6 Luna 的 GPT Decisions，再加上三个通用模型当对照：Qwen3.8-27B 的提示用法与受限读出用法，以及 GPT-6 Luna 直接提示。测试环境 ALFWorld（家务指令）与 WebShop（网购下单）的共同点是任务里有必须按顺序完成的子目标，比如「把加热过的杯子放到桌上」，杯子得先加热再放。\n\n三种角色的分法：演员在每个状态选下一个动作；裁判给已执行的动作打分，看多少上下文（动作之前的前缀，还是完整轨迹）与问什么（任务会不会成功，还是这一步有没有帮到目标）两两组合；老师把裁判打分加进 GRPO 训练的优势项，决策模型全程冻结，测试时不再调用。\n\n## 演员为什么演砸\n\n六个模型全栽在同一类任务上。JEV 在两个子目标可以乱序完成的任务上拿到 70.8%，到了必须先加热或清洁再放置的任务，每类最多 18.8%；Decider v1.2，总分 63.6%，照样在 Heat 任务上只拿 18.8%；GPT Luna 总分 34.3%。\n\n模型并不是「不会」。目标遮蔽实验里，在 116 个关键状态上，JEV 把 63.8% 的概率押在「直接放杯子」这个顺手的目标上，只给该先做的步骤 14.6%；把简单目标从指令里删掉，先决步骤的概率立刻涨到 77.6%。知识在，是「先做容易的」这股引力把概率拉走了。往提示词里加依赖提醒基本没用，六个模型的成绩变化落在 -2.8% 到 +5.2% 之间；作者更建议直接把目标重构成有序子任务。\n\n## 换个问法，裁判就灵了\n\n同一批模型当裁判，胜负手是两个设计：看什么、问什么。看完整轨迹时，JEV 给里程碑动作的打分比无效动作高 0.36；改问「任务最后会不会成功」，差距只剩 0.05——整条轨迹里每个动作的答案都一样，等于把结果奖励复读了一遍。三档动作的均分也从后一种问法的 0.632、0.604、0.577（几乎糊在一起）拉开到前一种的 0.659、0.440、0.298。论文还测得，两个基座公开的专用决策模型当裁判，比自家基座通用模型直接用更准，每次打分成本低 28% 到 43%。\n\n## 当老师：小模型反超大模型\n\n「看全程 + 问贡献」组合搬进 GRPO：Qwen2.5-1.5B 在 ALFWorld 从 72.8% 涨到 82.9%，WebShop 从 56.8% 涨到 71.4%；换成问成功率的问法只有 69.3% 与 68.3%。1.5B 学生在两个环境都超过了只用结果奖励训练的 7B 模型。增益恰好集中在演员自己也做不好的那类任务上：有序子目标任务策略做到 86.8%，而 JEV 自己当演员只完成 14.7%。\n\n账单也便宜：JEV 走 OpenRouter，输入每百万 token 0.042 美元、输出计零，一轮训练的教师调用费在 22 到 30 美元之间。不用人工标注每一步，不用额外训一个评论家，冻结的决策模型每集结束打一轮分就够。\n\n## 为什么值得记住\n\n这篇论文真正把「模型会不会做事」和「模型能不能监督做事」拆成了两件事：挑监督者该看裁判成绩单，而裁判可以在训练前用极低成本筛出来。这个「总分高不等于靠得住」的判断，另一个独立团队在同一周给了佐证——新加坡国立大学的 SO-Arena 基准评了 21 个 System One 模型——TypeSafe AI 当初发布 JEV 用的正是这个词——覆盖 11 个领域，发现按它的总分榜排头名的模型，在决策可靠性维度只排第七。两篇论文互不相识，指向同一件事：这类模型的「能力」和「可靠」是两个维度。\n\n论文自己划了边界：只有两个文本环境、每条轨迹最多 30 步；在线场景只有前缀可看，事后裁判帮不上忙。代码以 Apache 2.0 许可证开源在 [GitHub](https:\u002F\u002Fgithub.com\u002FZJU-REAL\u002Fdecision-agents)，论文见 [arXiv 2611.17167](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.17167)。留一个问题：你手头的 agent 管线里，负责打分的那个模型，问的问题问对了吗？\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.17167","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},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":19,"name":20,"slug":20,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"997dcc43-5f98-4924-962e-bdc456fa72b3","en","Acts at 42.9%, teaches to 82.9%: decision models' roles","JEV completes 42.9% of tasks as an actor yet teaches a Qwen2.5-1.5B policy to 82.9%: decision models tested as actor, judge and teacher.","Same model, same tasks, a completely different report card. The decision model JEV, acting as an agent in the ALFWorld household environment, completes only 42.9% of tasks; moved behind the curtain as a teacher for Qwen2.5-1.5B, it drives its student to 82.9%. The paper, submitted to arXiv on November 12 (2611.17167) by the ZJU-REAL group at Zhejiang University together with Baidu, takes the much-discussed decision-model category and tests it in three roles - actor, judge, and teacher. The conclusion is counterintuitive: a model that acts poorly can still teach well.\n\n## How the three roles are defined\n\nDecision models are a recent category: given a question and a context, they return a typed output - a choice among options with a probability for each - instead of generated text, which makes them fast and cheap. The paper evaluates three dedicated decision models: JEV from TypeSafe AI, the open-weight Decider v1.2 from Perplexity, and GPT Decisions from OpenAI built on GPT-6 Luna. Three general LLMs serve as controls: Qwen3.8-27B used via prompting, the same model via constrained readout, and GPT-6 Luna via prompting. The two test environments, ALFWorld (household instructions) and WebShop (online shopping), share a property: tasks contain subgoals that must be completed in order - in 'put the heated mug on the table', the mug must be heated before it is placed.\n\nThe three roles split cleanly. The actor picks the next action at every state. The judge scores already-executed actions, along two axes that combine: what it sees (the trajectory prefix up to the action, or the full trajectory) and what it is asked (whether the task will succeed, or whether this action helped the goal). The teacher feeds those scores into the advantage term of GRPO training; the decision model stays frozen throughout and is never queried at test time.\n\n## Why actors fail\n\nAll six models fall over on the same kind of task. JEV scores 70.8% on tasks whose two subgoals can be done in either order, but at most 18.8% on tasks that require heating or cleaning before placement; Decider v1.2, overall actor score 63.6%, still gets only 18.8% on Heat tasks; GPT Luna manages 34.3%.\n\nThe models do know the required step. In a goal-masking experiment over 116 key states, JEV puts 63.8% of its probability on the easy subgoal - placing the mug directly - and only 14.6% on the step that must come first; remove the easy subgoal from the instruction, and the probability on the required step jumps to 77.6%. The knowledge is there; the gravity of 'do the easy thing first' pulls the probability away. Adding a dependency reminder to the prompt barely helps - success changes by -2.8% to +5.2% across the six models - and the authors suggest restructuring the goal into ordered subtasks instead.\n\n## Change the question, and the judge works\n\nFor the same models as judges, two design choices decide everything: what they see and what they are asked. Seeing the full trajectory, JEV scores milestone actions 0.36 higher than no-op actions; asked whether the task will eventually succeed, the gap collapses to 0.05 - every action in a trajectory then gets the same answer, effectively replaying the outcome reward. Mean scores for milestone, other, and no-op actions spread from 0.632, 0.604, and 0.577 under the success question to 0.659, 0.440, and 0.298 under the contribution question. The paper also finds that the two dedicated decision models with disclosed base LLMs judge more accurately than those base LLMs given the same question, at 28% to 43% lower cost per score.\n\n## As a teacher: the small model overtakes the big one\n\nThe full-trajectory-plus-contribution-question combination goes into GRPO: Qwen2.5-1.5B rises from 72.8% to 82.9% on ALFWorld and from 56.8% to 71.4% on WebShop; the success-question variant yields only 69.3% and 68.3%. The 1.5B student beats a 7B model trained with outcome rewards alone on both environments. The gains land exactly on the tasks the actor itself rarely completes: on tasks with ordered subgoals the policy reaches 86.8%, while JEV itself completes only 14.7% of them as an actor.\n\nThe bill is modest: JEV is queried through OpenRouter at 0.042 USD per million input tokens with zero output cost, and teacher queries for a training run cost between 22 and 30 USD. No per-step labels, no separately trained critic - a frozen decision model scoring each episode once it ends is enough.\n\n## Why it matters\n\nThe paper genuinely separates 'can the model do the task' from 'can the model supervise the task': pick supervisors by their judging record, and judges can be screened cheaply before any RL run. The broader point - a strong scoreboard does not mean reliable decisions - got independent confirmation the same week: SO-Arena, a benchmark from the National University of Singapore covering 21 System One models (the very term TypeSafe AI used when launching JEV) across 11 domains, finds the model at the top of its scoreboard ranks only seventh on decision reliability. Two unrelated papers point at the same thing: capability and reliability are two different axes for these models.\n\nThe authors state their own limits: two text environments only, at most 30 actions per trajectory, and online settings where only the prefix is visible are out of reach for hindsight judging. Code is open-sourced under Apache 2.0 on [GitHub](https:\u002F\u002Fgithub.com\u002FZJU-REAL\u002Fdecision-agents); the paper is at [arXiv 2611.17167](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.17167). One question to leave you with: in your agent pipeline, is the model doing the scoring being asked the right question?\n","decision-model-actor-judge-teacher","2026-11-15T21:45:00Z","2026-11-15T21:12:55.325282Z","2026-11-15T21:12:55.325295Z",true,"agent",7,[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},"bd350599-7253-4f32-a874-9224bab89a49","多智能体搜索实测：32组对照里委派多数时候反而更差","multi-agent-search-delegation-accuracy","2026-11-15T09:10:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"a2038f05-8ad2-4b57-bb29-47ff522e09a0","失败轨迹偷走权重:ALFWorld成功率涨7到9.4点","outcome-mass-alignment-multi-turn-agent-rl","2026-11-14T11:20:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"9531b4f7-3f4d-474c-b233-c715925a3e76","8.4B轨迹压过25.3B技能语料:形式化证明中训配方","davinci-prover-agent-native-mid-training","2026-11-14T07:20:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"343f6cae-ae14-4617-a317-7316748e1d6e","删掉推理链再转发:同档32B模型差51分","deletion-sensitivity-reasoning-trace-routing","2026-11-13T07:05:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"365a0af2-d564-4eff-bf03-a138cdea0c5f","研究智能体大考:只查交付过69%,查全要求剩28%","are-bench-research-agent-context-requirements","2026-11-13T01:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"f2723003-43ac-47fe-bb59-49030446919a","28个决策模型大考:多选最高才36.81%","mmdm-decision-model-benchmark","2026-11-12T13:30:00+00:00"]