11-19 更新 实时
首页/ AI 大模型/ 正文

OpenAI GPT-6 Astra 攻克标准模型格点数学难题

AI 大模型 来源:The Decoder 2026-11-17 03:19 0 阅读 机器翻译
OpenAI GPT-6 Astra 攻克标准模型格点数学难题

Nano Banana Pro,提示词由 THE DECODER 提供

OpenAI 的 GPT-6 Astra 协助完成了一项证明,有可能解决粒子物理学中一个存在了 50 年的难题。

加州大学圣塔芭芭拉分校(UC Santa Barbara)物理学家 Nathaniel Craig 目前是 OpenAI 的访问科学家,他在 OpenAI 博客上以及一篇预印本论文中介绍了这一方法。Craig 领导 UCSB 的粒子理论小组,同时也是该校卡弗里理论物理研究所(Kavli Institute for Theoretical Physics)的永久成员。

物理学家可以在被称为格点(lattice)的计算机网格上模拟标准模型的部分内容,标准模型是描述所有已知粒子和基本作用力的理论。这已经带来了诸如从第一性原理计算质子质量这样的突破。但把完整理论放到格点上一直无法实现。弱力支配着某些类型的放射性衰变,它对左手粒子和右手粒子的处理方式不同。在简单的格点上,这种不对称会产生不需要的幽灵粒子。被称为重叠费米子(overlap fermions)的特殊格点粒子可以将其滤除,但 25 年多来,没有人能正确设定它们的相位。

该图展示了比较越来越精细的格点如何有助于修正原始格点上的相移。|图片:Nathaniel Craig / GPT-6 Astra

Craig 有一套策略,通过比较逐步变精细的格点来修正相位,但他无法证明这套策略在它需要覆盖的所有情形下都成立。GPT-6 Astra 将这个想法转化为可行的构造,并在详细证明中承担了主导工作。格点场论学者 Maarten Golterman、Andreas Kronfeld 和 Yigal Shamir 审阅了该证明,并提出了一些问题;Craig 说这些问题有助于增强对结果的信心。他将这项工作称为“与其说是终点,不如说更像起点”,并表示在能够进行逼真的模拟之前还有大量工作要做。

这一成果出现在 2026 年 1 月开启的一轮 AI 数学进展之后。5 月,OpenAI 对 Erdős 单位距离猜想的证伪成为首个足以登上数学顶刊的 AI 证明。几天后,Google DeepMind 的 AlphaProof Nexus 仅用几百美元就解决了 353 个未解 Erdős 问题中的 9 个,到 10 月,其结果已通过同行评审。

最引人瞩目的说法出现在 9 月,当时 OpenAI 提交了一份 166 页的证明,针对纳维-斯托克斯千禧年大奖难题,克莱数学研究所表示该问题“似乎已被解决”。该证明仍在评审之中,数学家们质疑 OpenAI 是否完全理解自己的证明。在整个领域,随着 AI 不断攻克未解难题,数学家们的心情颇为复杂。

没有炒作的 AI 新闻——由人类精选

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(OpenAI's GPT-6 Astra tackles the Standard Model's toughest math problem on a lattice)

Nano Banana Pro prompted by THE DECODER

OpenAI's GPT-6 Astra helped produce a proof that could solve a 50-year-old problem in particle physics.

Physicist Nathaniel Craig of UC Santa Barbara, currently a visiting scientist at OpenAI, presented the approach on the OpenAI blog and as a preprint paper. Craig leads the particle theory group at UCSB and is a permanent member of its Kavli Institute for Theoretical Physics.

Physicists can simulate parts of the Standard Model, the theory describing all known particles and fundamental forces, on a computer grid called a lattice. That has already led to breakthroughs like calculating the proton's mass from first principles. But putting the full theory on a lattice has been impossible. The weak force, which governs certain types of radioactive decay, treats left- and right-handed particles differently. On simple lattices, this asymmetry creates unwanted ghost particles. Special lattice particles called overlap fermions can filter them out, but for over 25 years, nobody could set their phase correctly.

The diagram shows how comparing increasingly finer lattices helps correct the phase shift on the original lattice. | Image: Nathaniel Craig / GPT-6 Astra

Craig had a strategy to fix the phase by comparing progressively finer lattices, but couldn't prove it worked in all the cases it needed to cover. GPT-6 Astraturned the idea into a working construction and took the lead on the detailed proof. Lattice field theorists Maarten Golterman, Andreas Kronfeld, and Yigal Shamir reviewed the proof and raised questions that Craig says helped build confidence in the result. He calls the work "much more a beginning than an end" and says a lot remains before realistic simulations are possible.

The result follows a run of AI advances in mathematics that started in January 2026. In May, OpenAI's disproof of Erdős's unit distance conjecture became the first AI proof worthy of math's top journal. Days later, Google Deepmind's AlphaProof Nexus solved nine of 353 open Erdős problems for a few hundred dollars, and by October its results had passed peer review.

The most far-reaching claim came in September, when OpenAI submitted a 166-page proof for the Navier-Stokes Millennium Prize Problem and the Clay Mathematics Institute said it had "apparently been settled". The solution remains under review, and mathematicians have questioned whether OpenAI fully understands its own proof. Across the field, mathematicians have mixed feelings as AI keeps cracking unsolved problems.

AI News Without the Hype – Curated by Humans

原文为英文,正文由自动抓取并保留原文。

本文由采集器自动同步

原文发布于 The Decoder(OpenAI's GPT-6 Astra tackles the Standard Model's toughest math problem on a lattice),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗