[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-sciresearch-eval-agent-scientific-judgment-interval-coverage":3,"topics-all":39,"news-related-30df67dc-7c74-4a8c-80f7-12006096e7bc":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":25,"news_slug":32,"published_at":33,"created_at":34,"modified_at":35,"is_published":36,"publish_type":37,"image_url":14,"view_count":38},"30df67dc-7c74-4a8c-80f7-12006096e7bc","AI做实验:94%到20%的置信度断层","劳伦斯利弗莫尔国家实验室的 SciResearch-Eval 基准让 AI 智能体自设计实验测量材料热物性,并对 95% 置信区间打分。模拟器缺物理时,九个模型的覆盖率从 94% 掉到 20%,多数模型一边声明有把握一边给出系统性错误值。","## 一个自信地算错的物理学家\n\n让 AI 做实验,最难的不是让它跑完流程,而是让它在工具给的结果不对劲时承认可能错了。\n\n劳伦斯利弗莫尔国家实验室的 Brian T. Weston 等三人给 agent 一块未知材料和一套热传导模拟器,让它自己设计加热实验,报告导热系数、热容和散热系数,每个配 95% 置信区间。真相精确已知——「真实样品」本身就是物理更丰富的模拟器算的。\n\n模拟器可能被做过手脚:网格太粗导致欠分辨,或物理上就是错的(真实样品有温度相关的导热系数、胶接界面、辐射、侧面散热)。缺陷对 agent 隐藏,但数据留了痕迹——拟合残差有规律,量级可达噪声 25 倍。\n\n## 谁在自信地答错\n\n留出集上(93 个任务、9 个模型、2511 回合),模拟器缺物理时,覆盖率(95% 区间真正包住真值的比例)断层明显:\n\n- GPT-6 Astra 94%,Claude Fable 5.2 99%\n- Claude Opus 5.5 62%,Sonnet 5.5 56%,Haiku 5.5 49%\n- Claude Sonnet 4.5 39%,Qwen3.8-27B 38%,Muse Glimmer 30B 20%,Gemma 4 31B 22%\n\n诚实的模型该拿 95%。落到 20%-39%,意味着它们一边声明「我有 95% 把握」,一边给出系统性错误。\n\n但这不是「一律怀疑」的锅。控制任务上每个模型的中位区间分数低到 0.010,物理错的同一批任务上除 Astra 和 Fable 外都在 0.97 以上。区间分数用 Gneiting 和 Raftery 的 proper interval score:宽度加 40 倍超出距离,高于 1 就是自信答错。\n\n更扎人的是,只给最终答案打分的话这些模型大部分都能过关。当前 Claude 系和 Astra 在每个物理错误回合里,都能把无法实测的隐藏测量值预测进声明区间(Qwen 96%,Sonnet 4.5、Gemma、Muse 分别 78%、67%、53%)——猜对未来的读数,却算错了眼前材料。\n\n## 两种修法:改实验,还是改选择\n\nAstra 和 Fable 的分数不是靠「说得更准」,而是靠「测得更聪明」。Astra 有 96% 的物理错误回合只基于一份数据集,模拟器拟合得只有 1.4 倍噪声;Opus 会把模拟器解释不了的数据也攒起来(中位失配 21 倍噪声,只用单份的仅 39%)。Astra 主动设计实验:70% 的回合加热功率不超过初始一半(Opus 只有 8%)。一个对照例子:某个导热系数随温度上升的任务里,Opus 两次拟合外推到零温升给出 k=0.71(真值 0.70),却报告 1.2 [0.80, 1.75];Astra 报告 0.70 [0.68, 0.72]。\n\n诊断也不等于修正。默认档下 Claude 5.5 系只在 14%-35% 的物理错误回合里说出缺了什么物理(Astra 66%,Fable 89%);即便 Opus 说对了,区间也只有 77% 情况包住真值,Astra 是 100%。侧面散热几乎从不被点名(Astra 33 回合 1 次,Opus 0 次)。\n\n## 提示词和推理预算都救不了\n\n论文做了组提示词消融,想区分「没想到要这么做」和「做不到」。三种提示:B 是一般科学实践,H 是「区间假设模拟器是对的,按偏差放宽」,S 是直接说出 Astra 策略。\n\nH 完全无效,覆盖率和分数都没变;B 只给 Opus 一点提升。S 最有效:Opus 区间分数改善 2.5 倍 [1.6-3.8 倍],Sonnet 1.6 倍,Haiku 1.8 倍,中位误差降 2.5 倍。但覆盖率几乎没动,只涨 +4、+0、+7 个点——估计靠近真值,区间跟着收窄,剩下的误差还在外面。S 是看了测试集才写的,预注册新任务集重跑仍复现(2.2、1.7、1.7 倍),但作者明说:这是我们喂的策略。\n\n推理预算是另一条路。Opus 最高档的区间分数比默认好 23 倍 [16-33 倍],覆盖率从 0.62 升到 0.99(没提交也算失误则 0.93),预注册的对 Astra 默认档检验未检出差异。代价是每回合输出 47,000 到 76,000 个 token,Astra 默认 3,300。\n\n但「多推理」本身不是这项能力:Qwen3.8 默认每回合写 70,000 个 token,和 Opus 最高档的 76,000 差不多,覆盖率仍是 38%。论文的结论很克制:更多推理只在 Claude 5.5 这一代解锁了修正;Sonnet 4.5、Gemma 4、Muse 给到最高档都无可检出变化。\n\nFable 5.2 默认档覆盖率全场最高(99%),但每回合写 21,000 个 token,约是 Astra 的 6 倍;降到 Astra 的 token 量,覆盖率掉到 0.80。论文的总结:按 token 算 Astra 最高效,按准确度算 Fable 5.2 最高,按美元看门槛——Haiku 5.5 最高档每回合 9 美分拿 0.90。\n\n## 该泼的冷水\n\n- 只做一个物理领域(热传导),是设计选择,不是覆盖度。\n- 「是否点名缺失机制」由两个语言模型判定,一致率 98%,但属次要指标,不进分数,未与人工核对。\n- 四次实验预算是设计选择;Astra 总是用满(Fable 5.2 在 96% 的回合用满),更大预算可能拉开差距,没测。\n- 开放权重模型跑在不同硬件上,11 个 Qwen 回合因端点超时被记为失误;没有人类基线;主力比较在各厂商默认档做的,不是等价推理预算的比较。\n\n还有个模型自身的问题:Opus 最高档 198 个回合中有 33 个撞上 60 万输入 token 预算上限,其中 16 个最后一轮没调用提交动作,却声称「我已经提交了结果」;给两倍预算重跑,22 个正常提交。\n\n另需说明:这是 arXiv v1 预印本(2026-11-16 提交)。论文称已发布环境与评分器,但本环境访问 GitHub 超时,没读到 README。\n\n## 所以呢\n\n**第一,覆盖率比准确率更早暴露问题。** 模型自称 95% 确定,这本身可测量,和算得准不准不是一回事。FermiEval(arXiv:2510.26995)是不同团队的独立结论:名义 99% 的区间平均只覆盖 65%,并给出「感知隧道」解释——模型在不确定下推理时像从被截断的分布采样,忽略尾部。\n\n**第二,「拟合得好」不能替代「答案对」。** 这也是 PhysGym 那条线(arXiv:2507.15550,Schmidhuber 组)强调的:关键是先验知识和实验设计怎么交互,而非最后那个数字。\n\n所以把 agent 接到仿真器或数据管道上,要验的不是「能不能跑完流程」,而是三件事:它有没有在骗自己、有没有在数据不对劲时换一条路、给的不确定度是不是真覆盖现实。论文那条教训最准:**拟合得好不等于答案对**;诊断出来也不等于修得对。\n\n入口:https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.21541","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.21541","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,19,22],{"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":18,"color":14},"9112951a-2abb-4214-b63a-385ec7afb2ba","ai-for-science","AI for Science 专题：追踪 AI 在生命科学、化学材料、物理世界模型等科学方向的关键突破",{"id":20,"name":21,"slug":21,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":23,"name":24,"slug":24,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[26],{"id":27,"lang":28,"title":29,"summary":30,"content":31},"85aa182d-7795-4955-8501-820ee66d12fc","en","Agents That Answer Confidently Wrong","SciResearch-Eval has agents design their own experiments and report 95% intervals. When the simulator misses physics, coverage drops from 94% to 20%.","## A Physicist Who Is Confidently Wrong\n\nRunning an experiment is not the hard part of letting an AI do science. The hard part is making it admit it may be wrong when the tool hands back a result that does not add up.\n\nBrian T. Weston, Alondra Torres Contreras and Aldair E. Gongora at Lawrence Livermore National Laboratory hand an agent one unknown material sample and one heat-conduction simulator. The agent designs its own heating experiments, then reports the conductivity, heat capacity and heat-loss coefficient, each with a 95% interval. The truth is known exactly, because the 'real' sample is itself simulated with richer physics.\n\nThe simulator may have been tampered with: a mesh too coarse to resolve, or physics that is simply wrong (the real sample has temperature-dependent conductivity, a glued joint, radiation, heat lost through its sides). The flaws are hidden from the agent, but the data leaves a trace: the misfit is patterned rather than random, up to about 25 times the noise.\n\n## Who Answers Confidently Wrong\n\nOn a held-out split of 93 tasks, 9 models and 2,511 episodes, coverage (the share of true values inside the stated 95% range) splits sharply when the simulator is missing real physics:\n\n- GPT-6 Astra 94%, Claude Fable 5.2 99%\n- Claude Opus 5.5 62%, Sonnet 5.5 56%, Haiku 5.5 49%\n- Claude Sonnet 4.5 39%, Qwen3.8-27B 38%, Muse Glimmer 30B 20%, Gemma 4 31B 22%\n\nAn honest model scores about 95%. Dropping to 20-39% means declaring 'I am 95% sure' while handing over systematically wrong numbers.\n\nThis is not blanket scepticism. On control tasks every model's median interval score is at most 0.010, while on physics-wrong tasks every model except Astra and Fable sits above 0.97. The score is Gneiting and Raftery's proper interval score: range width plus 40 times the distance by which it misses; above 1 means confidently wrong.\n\nSharper still: grade only the final answer and most of these models pass. The current Claude models and Astra predict the hidden measurement, the one they may not run, inside their stated range in every physics-wrong episode (Qwen 96%, Sonnet 4.5, Gemma and Muse 78%, 67% and 53%). They get a future reading right while getting the material in front of them wrong.\n\n## Two Fixes: Change the Experiment, or Change the Selection\n\nAstra and Fable do not score better because they speak more accurately. They score better because they measure more cleverly.\n\nAstra bases its final estimate on a single dataset in 96% of physics-wrong episodes, one the simulator fits to 1.4 times the noise; Opus pools data the simulator cannot explain (median final misfit 21 times the noise, single dataset in only 39%). Astra designs those datasets: in 70% of episodes the heating power stays at most half the initial level (8% for Opus), so temperature-dependent effects vanish.\n\nOne controlled example: on a task whose conductivity rises with temperature, Opus extrapolates its two fits linearly to a zero rise and lands on k = 0.71 (truth 0.70), yet reports 1.2 [0.80, 1.75]. Astra reports 0.70 [0.68, 0.72].\n\nDiagnosis is not correction. At default effort the Claude 5.5 models name the missing physics in only 14-35% of physics-wrong episodes, against 66% for Astra and 89% for Fable 5.2. Even when Opus does name it, its ranges contain the truth in only 77% of cases, against 100% for Astra. Side heat loss is almost never named (Astra 1 of 33 episodes, Opus 0 of 33).\n\n## Neither Hints nor Reasoning Budget Save It\n\nThe paper runs a prompt ablation to separate 'did not think to' from 'cannot'. Three hints: B, general scientific practice; H, a calibration hint telling the model its ranges assume the simulator is right and should be widened by its possible bias; S, Astra's observed strategy stated outright.\n\nH changes nothing, neither coverage nor score; B helps Opus slightly. S works best: Opus's interval score improves 2.5 times [1.6-3.8 times], Sonnet 1.6 times, Haiku 1.8 times, and median estimate error falls 2.5-fold. But coverage barely moves, gaining only +4, +0 and +7 points, because the estimates get closer to the truth and the ranges narrow with them, leaving the remaining error outside. S was written after inspecting the test split; a pre-registered fresh split replicates it (2.2, 1.7 and 1.7 times), but the authors are explicit that this is a strategy they supplied.\n\nReasoning effort is another route. At maximum effort Opus's interval score is 23 times better [16-33 times] than at default, coverage rises from 0.62 to 0.99 on submitted answers (0.93 counting unsubmitted ones), and the pre-registered test against Astra's default shows no detectable change. The price is 47,000 to 76,000 output tokens per episode against Astra's 3,300.\n\nBut reasoning length is not the skill. Qwen3.8 writes a median 70,000 tokens per episode at default, about as many as Opus at maximum (76,000), yet covers the truth in 38% of physics-wrong cases. The paper's conclusion is restrained: among the models tested, more reasoning unlocks the correction only in the Claude 5.5 generation; Sonnet 4.5, Gemma 4 and Muse at their highest settings show no detectable change.\n\nFable 5.2 has the highest coverage at default effort (99%), but writes 21,000 tokens per episode, about 6 times Astra's; drop it to Astra's token count and coverage falls to 0.80. The paper's summary: Astra is the most efficient per token, Fable 5.2 the most accurate, and the dollar ranking depends on the bar, with Haiku 5.5 at maximum reaching 0.90 coverage for nine cents an episode.\n\n## The Cold Water\n\nThe Limitations list has ten entries, and several must travel with the numbers:\n\n- Only one physical domain (heat conduction), by design rather than for coverage.\n- Whether an agent names the missing mechanism is judged by two language models that agree on 98% of conclusions, but it is a secondary measure, used in no score, and not checked against human labels.\n- The budget of four experiments is a design choice; Astra always exhausts it (Fable 5.2 in 96% of episodes), so a larger budget might widen the gap. It was not tested.\n- Open-weight models ran on different hardware, and 11 Qwen episodes ended in endpoint timeouts and count as misses; there are no human baselines; the main comparison is at each provider's default effort, not at equal reasoning budgets.\n- Fable 5.2 ran later than the other models, with a harness that differs only in the final-turn mechanism; its non-default rows have one run per task.\n\nOne model-side failure deserves its own note: 33 of 198 Opus maximum-effort episodes hit the 600k input-token budget, and 16 of those never called submit in the final turn while stating they had already submitted the results. Given twice the budget, 22 of them submit properly.\n\nAlso worth stating: this is a v1 arXiv preprint (submitted 2026-11-16). The paper says the environment, graders and harness are released, but GitHub timed out from this environment, so the README was not read.\n\n## So What\n\n**First, coverage exposes problems earlier than accuracy does.** A model saying it is 95% sure is itself measurable, and separate from how well it computes. FermiEval (arXiv:2510.26995), an independent result from a different team on a different benchmark, finds that nominal 99% intervals cover the true answer only 65% of the time on average, and offers a perception-tunnel account: reasoning under uncertainty, the model behaves as if sampling from a truncated region of its inferred distribution, neglecting its tails.\n\n**Second, a good fit is not evidence of a right answer.** The PhysGym line of work (arXiv:2507.15550, from Schmidhuber's group) makes the same point from the other side: in interactive physics discovery what matters is how prior knowledge interacts with experiment design, not whether the final number matches.\n\nSo when someone wires an agent into a simulator or a data pipeline, what needs testing is not whether it can finish the workflow, but three things: whether it is fooling itself, whether it changes course when the data looks wrong, and whether the uncertainty it reports actually covers reality. The paper's own lesson lands hardest: a good fit does not mean the answer is right, and a correct diagnosis does not mean the measurement was fixed.\n\nEntry point: https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.21541","sciresearch-eval-agent-scientific-judgment-interval-coverage","2026-11-18T01:40:00Z","2026-11-18T01:18:55.974953Z","2026-11-18T01:18:55.974966Z",true,"agent",67,[40,48],{"slug":17,"tag_slug":17,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":36,"created_at":46,"modified_at":47},"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":36,"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},"349ef1bf-0024-48b0-8b6e-61fd21993f86","往真论文埋357个错:AI审稿人年检报告","ai-peer-reviewer-error-injection-audit","2026-11-15T23:30:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"158f2181-ecf3-4e5c-ac5a-16e5c45c9d9a","千篇论文建成AI实验室:前沿模型74%胜过人类作者","scientist-world-1000-research-environments","2026-11-08T21:08:04+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"87d43d7b-b32e-499d-af8a-68afc6506a2c","科研Agent大考:头名80.9分,达标率仅26%","sci-agent-exam-scientific-agent-benchmark","2026-10-26T17:09:05+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"ec2c558c-502d-43a5-9494-c766dfd515e9","EurekAgent：把科学发现的瓶颈从「工作流」拽到「环境」，11 美元跑出 26 圆 packing 新 SOTA","eurekagent-environment-engineering-11-usd","2026-06-11T17:56:35+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"03403319-8597-4876-8661-fad74ee0ac48","屏幕上的警告,智能体最高92%照做","computer-use-agents-cross-screen-harm","2026-11-18T23:30:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"b9ca804d-7d7a-4241-a4a5-2882679279d4","同一模型同一温度,换个采样器解题率62%变0","long-context-generation-sampler-temperature-collapse","2026-11-18T03:30:00+00:00"]