Aha.
正在载入中英对照阅读…

arXiv:2602.10177 · 中英对照阅读

迈向自主数学研究

Towards Autonomous Mathematics Research

Tony Feng、Trieu H. Trinh、Garrett Bingham、Dawsen Hwang、Yuri Chervonyi、Junehyuk Jung、Joonkyung Lee、Carlo Pagano、Sang-hyun Kim、Federico Pasqualotto、Sergei Gukov、Jonathan N. Lee、Junsu Kim、Kaiying Hou、Golnaz Ghiasi、Yi Tay、YaGuang Li、Chenkai Kuang、Yuan Liu、Hanzhao Lin、Evan Zheran Liu、Nigamaa Nayakanti、Xiaomeng Yang、Heng-Tze Cheng、Demis Hassabis、Koray Kavukcuoglu、Quoc V. Le*、Thang Luong*、*Project leads. Mathematicians. Work conducted under Google DeepMind

中文速览

把竞赛数学做得像金牌,不等于能在前沿研究里发现新定理;Aletheia要解决的正是AI能否把研究问题做出来、并减少编造和漏错。它让模型反复生成解答、由独立的验证环节检查,再根据反馈修改,同时借助联网检索查找数学文献。研究中,Aletheia参与产出了多篇达到发表水准的论文,还在700个开放问题的评估中协助解决四道题,并在研究级数学测试FirstProof上表现领先;不过这些结果的重要性不一,仍需人类专家判断。它的意义不仅在于展示AI开始能参与真正的数学研究,也在于提醒大家清楚记录AI贡献了多少、结果有多新颖,才能让人机合作更透明、可信。

摘要

基础模型的最新进展催生了推理系统,使其能够达到国际数学奥林匹克竞赛(IMO)金牌水平。我们致力于从竞赛级解题迈向专业研究,这一转变带来了重大的新挑战。为此,我们推出了 Aletheia,这是一种数学研究智能体,基于 Gemini Deep Think 所采用的一项新颖的推理时扩展定律,以自然语言端到端地迭代生成、验证和修订解答。我们通过自主数学研究的若干里程碑展示了 Aletheia 的能力:多篇可发表级论文,其中一篇完全没有人工干预(17);Feng 等人(2026b)针对 Bloom 的 Erdős 猜想数据库中的 700 个开放问题所开展的大规模半自主评估,其中包括自主解决四个开放问题;以及在 FirstProof 上取得领先表现。FirstProof 是一组由数学家提出、用于评估 AI 数学研究能力的研究级问题集合。提示词和模型输出的完整记录见 https://github.com/google-deepmind/superhuman/tree/main/aletheia。为帮助公众更好地理解 AI 与数学相关的发展,我们建议对 AI 辅助成果的自主程度和新颖性制定标准化量化等级,并提出“人机交互卡”这一概念,以实现透明记录。最后,我们就数学领域的人机协作进行了探讨。

Recent advances in foundational models have yielded reasoning systems capable of achieving a gold-medal standard at the International Mathematical Olympiad. We undertake the transition from competition-level problem solving to professional research, which presents significant new challenges. To this end we introduce Aletheia, a math research agent that iteratively generates, verifies, and revises solutions end-to-end in natural language, leveraging a novel inference-time scaling law based upon Gemini Deep Think. We demonstrate the capability of Aletheia through several milestones in autonomous mathematics research: multiple publication-grade papers, including one with no human intervention (17); an extensive semi-autonomous evaluation Feng et al. (2026b) on 700 open problems from Bloom’s Erdős Conjectures database, including autonomous solutions to four open questions; and a leading performance on FirstProof, a collection of research-level problems proposed by mathematicians to assess AI capabilities for mathematical research. Full transcripts of prompts and model outputs are shared at https://github.com/google-deepmind/superhuman/tree/main/aletheia. In order to help the public better understand the developments pertaining to AI and mathematics, we suggest quantifying standard levels of autonomy and novelty of AI-assisted results, and propose the concept of “human-AI interaction cards” for transparent documentation. We conclude with reflections on human-AI collaboration in mathematics.

术语表

foundational models
基础模型
reasoning systems
推理系统
International Mathematical Olympiad (IMO)
国际数学奥林匹克竞赛(IMO)
Aletheia
Aletheia
math research agent
数学研究智能体
inference-time scaling law
推理时扩展定律
Gemini Deep Think
Gemini Deep Think
autonomous mathematics research
自主数学研究
publication-grade papers
可发表级论文
Erdős Conjectures database
Erdős 猜想数据库
FirstProof
FirstProof
human-AI interaction cards
人机交互卡
AI-driven scientific discovery
AI 驱动的科学发现
hallucination
幻觉
eigenweights
特征权
arithmetic geometry
算术几何
independent sets
独立集
Autonomous Mathematics Research Levels
自主数学研究等级
SAE Levels of Vehicle Autonomy
SAE 车辆自动化等级
Generator
生成器
Verifier
验证器
Reviser
修订器
subagent orchestration
子智能体编排
formal language
形式化语言
AlphaGeometry
AlphaGeometry
AlphaProof
AlphaProof
parallel thinking
并行思考
inference-time scaling
推理时扩展