AI推动理论物理新突破
双语精读
AI is helping expand the frontier of theoretical physics
人工智能助力拓展理论物理学前沿
It is blurring the line between tool and collaborator
它正在模糊工具与合作者的界限
March 12th 2026
2026 年 3 月 12 日
IN 2025 A GROUP of theoretical physicists studying the behaviour of fundamental particles called gluons hit a brick wall in their calculations. In search of a fresh perspective, the physicists teamed up with OpenAI, an artificial-intelligence lab, to see whether an AI assistant might be able to help. Two preprints, published in early 2026, report the results of this collaboration. The AI’s role was central, say the researchers, enabling them to complete in weeks what would have typically required months. The long-touted idea that AI could help with work at the frontiers of theoretical physics is now a reality.
2025 年，一群研究名为胶子的基本粒子的行为的理论物理学家在计算中遭遇了瓶颈。为寻找新思路，这些物理学家与人工智能实验室 OpenAI 合作，试探 AI 助手能否提供帮助。2026 年初发表的两篇预印本论文报告了此次合作的成果。研究人员表示，人工智能起到了核心作用，让他们在数周内完成了通常需要数月才能完成的工作。长期以来被广为宣扬的 “人工智能可助力理论物理前沿研究” 的设想，如今已成现实。
What makes the interactions of subatomic particles so difficult to model is the fact that they obey the probabilistic laws of quantum physics. That means that when two particles enter a collision, it is impossible to definitively predict how many particles will leave. All physicists can ever do is determine the probability of various outcomes, which is done with the help of mathematical quantities called scattering amplitudes.
亚原子粒子的相互作用之所以难以建模，是因为它们遵循量子物理学的概率规律。这意味着，当两个粒子发生碰撞时，无法确切预测会产生多少个粒子。物理学家所能做的，只是计算各种结果出现的概率，而这要借助一种名为散射振幅的数学量来完成。
These quantities are challenging to compute, often involving many hundreds of intricate mathematical terms. In certain cases, however, mathematical patterns emerge that collapse these mammoth equations into simple, elegant forms. This simplicity is particularly striking for gluons—fundamental particles that transmit the strong nuclear force. A subset of their scattering amplitudes, known as single-minus tree-level, appear to vanish completely, implying that the associated processes could never occur. The authors of the new studies, however, suspected this conclusion was too strong.
这些物理量计算难度极大，往往包含数百个复杂的数学项。然而在某些情况下，会出现某种数学规律，能将这些庞大的方程简化为简洁优美的形式。这种简洁性在传递强核力的基本粒子 —— 胶子身上表现得尤为突出。胶子的一类散射振幅（被称为单负树图阶）似乎完全消失，意味着对应的物理过程不可能发生。但新研究的作者们怀疑，这一结论过于绝对。
The researchers had noticed that if the momenta of the particles entering and leaving a collision are made to take certain values, the amplitudes become non-zero. Calculating the simplest examples, involving only a few gluons, was straightforward. But as the number of particles increased, so did the complexity of the maths.
研究人员注意到，如果让参与碰撞及碰撞后产生的粒子的动量取特定数值，振幅就会变为非零。只涉及少数胶子的最简单情形计算起来并不困难，但随着粒子数量增加，数学复杂度也随之飙升。
When Alexandru Lupsasca, a physicist at Vanderbilt University and OpenAI, invited the researchers to test the physics capabilities of OpenAI’s latest models, the single-minus gluon scattering amplitudes seemed like the perfect problem. Given the physicists’ formulae, GPT-5.2 Pro both spotted simplifications they had missed and conjectured a generalisation—an expression valid for any number of gluons. The researchers then asked a more capable OpenAI model—one not publicly available—to confirm it. After 12 hours of thinking, the AI handed them a proof. The physicists checked through the mathematics; the AI’s working was correct.
The researchers posted their findings, which have not yet been peer reviewed, on arXiv on February 12th. But that was not the end of the story. They immediately wondered if the results could be extended to gravitons—hypothetical particles thought to carry the gravitational force. Gravitons have not been observed, but calculating their theoretical scattering amplitudes allows physicists to investigate how gravity might behave at the smallest scales.
研究人员于 2 月 12 日将尚未经过同行评审的成果发表在 arXiv 预印本平台。但故事并未就此结束。他们立刻想知道，这一结论能否推广到引力子—— 一种被认为传递引力的假想粒子。引力子尚未被观测到，但计算其理论散射振幅，能帮助物理学家研究引力在微观尺度下的行为。
Graviton calculations are even more complex than those for gluons. Yet on March 4th the researchers released a second paper. Using only the gluon results and some gentle prompting from the physicists, GPT-5.2 Pro was able to construct the analogous single-minus scattering amplitudes for gravitons. All that was left for the physicists to do was check its working.
引力子的计算比胶子更为复杂。但 3 月 4 日，研究团队发表了第二篇论文。仅借助胶子的研究结果和物理学家的适度提示，GPT-5.2 Pro 就成功推导出引力子对应的单负散射振幅。物理学家只需负责验证其推导过程即可。
“The physics problem now is not the hard part. The hard part is verifying the results and writing it up,” said Dr Lupsasca. “This feels surreal to me.”
“如今物理问题本身已不再是难点，难点变成了验证结果和撰写论文。” 卢普斯卡博士说，“这让我感觉很不真实。”
The physicists are now working with the models to investigate what these results mean for their theories. But the true significance of these two preprints may lie in their means, rather than the ends. For the researchers, the AI model has begun to blur the line between tool and collaborator. “It came back to me and said ‘Well, the obvious generalisation is…’ and wrote down the whole formula,” said Andrew Strominger, a physicist at Harvard University and co-author of the studies. “Which is just the kind of thing some of my more obnoxious colleagues would say.”
物理学家目前正与这些模型合作，探究这些结果对他们的理论有何意义。但这两篇预印本的真正意义，或许在于研究方式，而非研究成果。对研究人员而言，人工智能模型已开始模糊工具与合作者的界限。“它回过头跟我说‘显然，更一般的结论是……’，然后直接写出了完整公式。” 哈佛大学物理学家、该研究合著者安德鲁・斯特罗明格说，“这口吻跟我某些比较讨人嫌的同行一模一样。”
重点单词/短语
theoretical physics 理论物理学
frontier n. 前沿，尖端领域
gluon n. 胶子
hit a brick wall 遇到瓶颈，碰壁
team up with 与…… 合作
preprint n. 预印本
central adj. 核心的，关键的
long-touted adj. 长期被宣扬的
subatomic particles 亚原子粒子
probabilistic laws 概率规律
quantum physics 量子物理学
collision n. 碰撞
definitively adv. 确定地，确切地
scattering amplitudes 散射振幅
intricate adj. 复杂精细的，错综复杂的
mammoth / ˈmæməθ /adj. 庞大的，巨大的
elegant adj. 简洁优美的
strong nuclear force 强核力
single-minus tree-level 单负树图阶
vanish v. 消失
momenta n. 动量（momentum 的复数）
straightforward adj. 简单的，易懂的
formula n. 公式
simplification n. 简化
conjecture v. 推测，猜想
generalisation n. 推广
peer reviewed 同行评审的
graviton n. 引力子
gravitational force 引力
analogous adj. 类似的，类比的
verify v. 验证，核实
surreal adj. 超现实的，不真实的（助记：sur-表示“超越，超过”，sur(超过)+real(现实的) -> 超现实的）
obnoxious adj. 讨厌的，令人不快的