TL;DR
OpenAI has published a preprint showing that GPT-5.2 conjectured a formula for a type of particle interaction that physicists had assumed was impossible. The formula was later proved correct by a scaffolded version of the model and verified by human physicists, marking a notable case of AI contributing original results to theoretical physics.
What GPT-5.2 Found
The preprint, titled “Single-minus gluon tree amplitudes are nonzero,” focuses on gluons — the particles that carry the strong nuclear force. Standard physics textbooks state that when one gluon has negative helicity and the rest have positive helicity, the scattering amplitude at tree level should be zero. The paper shows this is not always the case.
The human authors — from the Institute for Advanced Study, Vanderbilt, Cambridge, and Harvard — worked out the amplitudes for specific cases up to six particles by hand, producing increasingly complex expressions. GPT-5.2 Pro simplified these expressions and identified a pattern, proposing a general formula valid for any number of particles.
An internal scaffolded version of GPT-5.2 then spent roughly 12 hours reasoning through the problem independently, arriving at the same formula and producing a formal proof. The result was verified against established methods including the Berends-Giele recursion relation.
Expert Response
Nima Arkani-Hamed, professor of physics at the Institute for Advanced Study, described the work as “especially well-suited to exploit the power of modern AI tools” and said he looked forward to seeing the approach develop into “a general purpose ‘simple formula pattern recognition’ tool.”
Nathaniel Craig, professor of physics at UC Santa Barbara, called it “journal-level research advancing the frontiers of theoretical physics” and said the preprint “felt like a glimpse into the future of AI-assisted science.”
Looking Forward
The authors report that GPT-5.2 has already helped extend the results from gluons to gravitons, with further generalisations in progress. The preprint has been submitted for publication and is available on arXiv. If the methodology proves reproducible across other areas of physics, it could establish a new model for AI-human collaboration in fundamental research.