Caroline Bishop
Mar 03, 2026 19:59
GPT-5.2 Professional proposed a novel components for gluon amplitudes that human physicists later verified, marking a primary for AI-generated scientific discoveries.
OpenAI’s GPT-5.2 Professional has accomplished one thing no AI system has publicly achieved earlier than: it proposed an authentic components in theoretical physics that human researchers subsequently verified as right. The preprint, printed March 3, 2026, demonstrates that sure gluon scattering amplitudes beforehand assumed to be zero can really happen beneath particular circumstances.
The invention challenges textbook assumptions about particle physics. For many years, physicists believed that when one gluon has detrimental helicity whereas the remaining particles have constructive helicity, the tree-level amplitude should equal zero. GPT-5.2 recognized a mathematical exception—the “half-collinear regime”—the place this reasoning breaks down and the amplitude turns into nonzero.
How the AI Bought There
Human researchers first calculated amplitudes for small particle numbers as much as n=6 by hand, producing what the paper describes as “very sophisticated expressions” whose complexity grows superexponentially. GPT-5.2 Professional simplified these unwieldy formulation dramatically, noticed a sample throughout the circumstances, and proposed a normal components legitimate for all n.
Here is the place it will get attention-grabbing: a scaffolded model of GPT-5.2 then spent roughly 12 hours reasoning by the issue independently, arriving on the identical components and producing a proper proof. The equation was verified towards the Berends-Giele recursion relation and gentle theorem constraints—normal validation strategies within the subject.
The preprint lists authors from the Institute for Superior Research, Vanderbilt College, Cambridge, Harvard, and OpenAI. Kevin Weil seems on behalf of OpenAI.
What Physicists Are Saying
Nima Arkani-Hamed, Professor of Physics on the Institute for Superior Research, referred to as the outcomes “strikingly easy expressions” and famous that discovering easy formulation has “at all times been fiddly” work he lengthy believed could possibly be automated. “I’m wanting ahead to seeing this development proceed in direction of a normal function ‘easy components sample recognition’ software within the close to future,” he mentioned.
Nathaniel Craig at UC Santa Barbara was extra direct about implications: “That is clearly journal-level analysis advancing the frontiers of theoretical physics.” He described the work as “a glimpse into the way forward for AI-assisted science” and a template for validating LLM-driven insights.
The Larger Image
GPT-5.2 launched December 11, 2025, with OpenAI touting large enhancements in reasoning and agentic workflows. The Professional variant particularly takes extra compute time to work by advanced issues—precisely the potential that apparently enabled this physics breakthrough.
The group has already prolonged these amplitude calculations from gluons to gravitons utilizing GPT-5.2 help, with extra outcomes forthcoming. The preprint is accessible on arXiv and beneath evaluate for journal publication.
For AI corporations racing to reveal real-world worth past chatbots and code completion, a verified contribution to elementary physics represents a unique class of feat solely. Whether or not this turns into a repeatable methodology or stays a one-off demonstration will seemingly form how analysis establishments method AI collaboration within the coming years.
Picture supply: Shutterstock
