人工智能正在显著加速理论物理研究。2025 年一组研究胶子散射振幅的理论物理学家在计算中遇到瓶颈，与 OpenAI 合作后使用 AI 模型在 数周内完成原本需要 数月的工作。研究成果以 两篇预印本形式在 2026 年初发布。问题涉及量子力学下的粒子碰撞概率计算，其中散射振幅通常包含 数百个复杂数学项。研究人员怀疑某些被认为为零的“single-minus tree-level”振幅实际上并非完全消失。

AI 在推导过程中发挥关键作用。研究人员向 GPT-5.2 Pro 提供公式后，模型不仅发现了新的数学简化，还提出一个适用于 任意数量胶子的通用表达式。随后他们使用更强的内部模型进行验证，该模型经过 12 小时计算给出了完整证明。论文于 2 月 12 日上传到 arXiv。研究人员随后进一步扩展工作，探索与引力子相关的散射振幅，这类计算比胶子问题更复杂。

在 3 月 4 日发布的第二篇论文中，GPT-5.2 Pro 在研究人员简单提示下利用胶子结果推导出引力子的对应振幅公式，物理学家只需验证数学推导。研究者认为，这种协作改变了研究流程：难点从推导公式转变为检查结果并撰写论文。随着 AI 能提出猜想、推广公式并生成证明，它在理论物理中的角色正从计算工具逐渐转向研究合作者。

Artificial intelligence is accelerating research in theoretical physics. In 2025 a group of physicists studying gluon scattering amplitudes encountered a computational barrier and collaborated with OpenAI, using AI models to finish work in weeks that would normally take months. Their findings appeared in two preprints released in early 2026. The problem concerns probabilistic particle collisions in quantum physics, where scattering amplitudes often involve hundreds of complicated mathematical terms. Researchers suspected that certain amplitudes known as “single-minus tree-level,” previously thought to vanish, might not actually be zero.

AI played a central role in deriving new results. After being given the researchers’ formulas, GPT-5.2 Pro identified additional mathematical simplifications and proposed a generalized expression valid for any number of gluons. A more powerful internal model was then asked to verify the conjecture and produced a complete proof after 12 hours of computation. The paper describing these results was posted on arXiv on February 12. The team then attempted to extend the findings to gravitons, hypothetical particles that would mediate gravity, whose scattering calculations are even more complex.

A second paper released on March 4 showed that GPT-5.2 Pro could derive analogous single-minus amplitudes for gravitons using the gluon results with minimal prompting. The physicists’ main task became checking the mathematics rather than discovering the formulas themselves. Researchers suggest this shift alters the research workflow: deriving results may become easier while verification and interpretation become the main challenges. As AI systems generate conjectures, generalizations and proofs, their role in theoretical physics is beginning to blur the line between computational tool and research collaborator.