论文开篇的立场是:科学进步发生在科学家找到现象背后的机制、并解释那些挑战现有理解的观察时;好的机制不仅要预测实验数据,还要产出帮助理解更大范围相关现象的洞见。而
AI agent 近年在长期未解的科学问题上取得突破(论文引 OpenAI 的
Navier–Stokes 结果与 Google 的 AI
co-scientist),其搜索与优化能力在"目标明确、结果可验证"的任务上尤其有效。作者接着引用
Terence
Tao(2026)的判断:找到正确解并不必然产出加深理解的科学洞见,甚至可能阻碍洞见的发现;数学与
AI 宣言的签名者把这一落差称为"AI
近年在科学中的用法与传统上驱动科学进步的方法之间的严重错位"。论文的研究问题由此而来:"Can AI agents move beyond solution optimization and discover novel
scientific insights?"
Morrison & Young 2025; Zhao et al. 2024; Chen et al.
2025a; Fieseler et al. 2025; Chen et al. 2026; Dunn et al.
2025; Buhmann et al. 2026; Scheurer & Heilbron 2026; Zhou
et al. 2026; Stoinski et al. 2026
(3) PA–SI 解耦有系统证据支撑。不是只报两个总分:逐问题内的 Kendall τb、反转率、同 PA 组内的 SI
差异(33.6%–63.4%)、+insights
消融、实验中洞见与机制导出洞见的分离,多层证据都指向同一结论。对社区这是定向警告——只优化可验证目标的训练范式,产出的机制在研究性外推上系统性落后。
(6) 跨基准比较的关键是证据单位对齐。EurekaBench 的 SI
依据具体问题、机制和可检验推导评分;因此与
Agent’s Last Exam
等其他评测的分数并列时,只有在任务单位、评分依据和可用资源相容后才适合做数值比较。否则更稳妥的用途是互补诊断,不应拼成排名;此为方法论提醒,不是对该基准结果的复核。
完整图谱 / Figure Gallery
附录 Figure 10–23:agent 原始输出(mechanism.md / experiments.log
摘录)与专家高亮,按 D.3 的十类失败与 D.4
的四类正面行为组织。每图均为论文附录原图的渲染裁剪,caption
为原文统一格式 "Original outputs from AI scientist
agents.";分类标签来自论文 D.3/D.4。
失败模式(D.3)
Figure 10(原文页 23 起):负例 · Missing tests of the target
behavior。agent 原始输出摘录(第一部分)。
Figure 10(续,原文页 24):负例 · Missing tests of the target
behavior(续)。
Figure 11(原文页 25 起):负例 · Explaining only part of the target
phenomenon(第一部分)。
Figure 11(续,原文页 26):负例 · Explaining only part of the
target phenomenon(续)。
Figure 11(续,原文页 27):负例 · Explaining only part of the
target phenomenon(续)。
Figure 11(续,原文页 28):负例 · Explaining only part of the
target phenomenon(续)。
Figure 12(原文页 29 起):负例 · Prediction with limited
explanation(第一部分)。
Figure 12(续,原文页 30):负例 · Prediction with limited
explanation(续)。
Figure 12(续,原文页 31):负例 · Prediction with limited
explanation(续)。
Figure 12(续,原文页 32):负例 · Prediction with limited
explanation(续)。
Figure 13(原文页 33 起):负例 · Unclear explanations(第一部分)。
Figure 13(续,原文页 34):负例 · Unclear explanations(续)。
Figure 14(原文页 35):负例 · Overgeneralization。
Figure 15(原文页 36):负例 · Failing to identify and reduce
uncertainty。
Figure 16(原文页 37 起):负例 · Treating task optimization as
mechanism discovery(第一部分)。
Figure 16(续,原文页 38):负例 · Treating task optimization as
mechanism discovery(续)。
Figure 16(续,原文页 39):负例 · Treating task optimization as
mechanism discovery(续)。
Figure 17(原文页 40):负例 · Interesting observations without
further exploration。
Figure 18(原文页 41):负例 · Conflicting results and conclusions。
Figure 19(原文页 42):负例 · Risk of circular validation。
正面行为(D.4)
Figure 20(原文页 43):正例 · Using controls to rule out
alternative explanations。
Figure 21(原文页 44):正例 · Testing which model better predicts
the observations。
Figure 22(原文页 45):正例 · Revising models after experimental
feedback。
Figure 23(原文页 46):正例 · Informative observations。
Other Discussion Roles / 其他讨论角色
以下七个角色使用 MIT MAS.S60 课程原始
prompt,从不同责任位置审视同一篇论文。
Scientific Peer Reviewer
The paper has not been published yet and is currently submitted to a
top conference where you’ve been assigned as a peer reviewer.
Complete a full review of the paper answering all prompts of the
official review form of the top venue in this research area (e.g.,
NeurIPS). This includes recommending whether to accept or reject the
paper.
Summary. 论文提出 EurekaBench:26
个由领域专家构建的跨 6 学科长程机制发现任务、306 条洞见问题,用
SC(科学约束)/ PA(预测准确率)/
SI(科学洞见)三轴与条件化计分评测 7 个前沿 agent。核心发现:最强
agent 的 PA 逼近人类参考(47.4% vs 48.8%),SI
大幅落后(42.4%/29.4% vs 69.7%),且 PA 与 SI 排名仅弱相关(τb =
0.20,反转率 39.9%)。
This paper was found buried under ground in the desert. You’re an
archeologist who must determine where this paper sits in the context
of previous and subsequent work. Find and report on one older paper
cited within the current paper that substantially influenced the
current paper and one newer paper that cites this current paper.
地层位置。科学发现评测的演化脉络:静态数据假设/方程恢复(DiscoveryBench、LLM-SRBench)→
交互式发现黑箱系统与虚构物理(Geng et al. 2025、DiscoverPhysics)→
本文:真实科研问题上的"洞见可导出性"测量。本文处于"从测机制发现到测研究性外推"的分层位置。
一篇实质影响本文的旧论文:Geng et al. 2025《Are
Large Language Models Reliable AI Scientists? Assessing
Reverse-Engineering of Black-Box Systems》(arXiv
2505.17968,标题经外部核验)。该文确立了"干预—观察—假设修正"的逆向工程循环作为
agent 科研能力的测法;本文的任务循环直接继承这一表述(第 2.1
节原话 "Such an iterative cycle of intervention, observation, and
hypothesis revision constitutes reverse engineering (Geng et al.,
2025a)"),且本文一作 Jiayi Geng 即该文作者,本文 CS 任务的
"Stated belief shift" 也引用了她的 2025b 工作(arXiv
2511.01805)。这是方法谱系上的直接前身。
You’re a researcher who is working on a new project in this area.
Propose an imaginary follow-up project not just based on the current
but only possible due to the existence and success of the current
paper.
可检验的成功标准:(a) judge 校准后 SI
与专家评分的相关 > 0.8;(b) RL 后的 agent 在留出任务 SI
上超过本文最强 baseline 且 SC 失败数不升。风险:judge
循环性若在校准中暴露大偏差,第一部分就要先换评测程序。
Industry Practitioner
You work at a company or organization developing an application or
product of your choice (that has not already been suggested in a
prior session). Bring a convincing pitch for why you should be paid
to implement the method in the paper, and discuss at least one
positive and negative impact of this application.
You’re a hacker who needs a demo of this paper ASAP. Implement a
small part or simplified version of the paper on a small dataset or
toy problem. Prepare to share the core code of the algorithm to the
class and demo your implementation. Do not simply download and run
an existing implementation – though you are welcome to use (and give
credit to) an existing implementation for “backbone” code.
You are a detective who needs to run a background check on one of
the paper’s authors. Where have they worked? What did they study?
What previous projects might have led to working on this one? What
motivated them to work on this project? Feel free to contact the
authors, but remember to be courteous, polite, and on-topic.
对象:一作 Jiayi Geng(脚注:CMU LTI,导师侧 Graham Neubig
组)。证据边界:本节只用论文内公开信息与其公开论文,不做网络深挖,也不联系作者(按约定,联系需用户明确批准)。
Identify how this paper self-assesses its (likely positive) impact
on the world. Have any additional positive social impacts left out?
What are possible negative social impacts that were overlooked or
omitted?
论文的自我评估:"opens up a new path for the
scientific study of AI-driven mechanism discovery across a variety
of scientific fields",并呼吁社区"develop AI scientists that
better support human researchers";结论节列出三个 open
challenges(多路径洞见、洞见评测、洞见导向训练)。资助方为 US DOE
与 Fujitsu(部分)。