GPT-5.2 and AI find new formula in theoretical physics | Keryc
GPT-5.2 proposed a new formula for a gluon amplitude that was later checked by an internal OpenAI model and verified by the authors in a preprint that's already on arXiv. What does this mean for physics and for AI applied to science? Let’s take it step by step.
What the preprint announced
A group of authors, including researchers from the Institute for Advanced Study, Vanderbilt, Cambridge, Harvard and OpenAI, published the preprint titled "Single-minus gluon tree amplitudes are nonzero." The main result: a configuration of gluon interactions that many assumed to be zero can actually be nonzero in a very specific region of momentum space.
The striking contribution is both methodological and substantive. GPT-5.2 Pro suggested a closed-form formula (it appears as equation 39 in the preprint) after analyzing complicated base cases. Then, an internal model reasoned for several hours and produced a formal proof that the authors verified analytically using standard discipline procedures like the Berends–Giele recursion relation and checks against the soft theorem.
What is a scattering amplitude and why does it matter?
A scattering amplitude is the quantity physicists use to calculate the probability that particles interact in a certain way. In field theory, many tree-level amplitudes (that is, without quantum loops) turn out to be surprisingly simple, and those simplifications often reveal deep structure in the theoretical framework.
In the case studied here, one gluon has negative helicity and the others positive. By standard arguments, that configuration was treated as if its amplitude were zero. The preprint shows that argument is too strong: there is a well-defined slice of momentum space, called the mid-collinear regime, where the amplitude does not vanish.
Think of it like expecting a mirror to reflect nothing in a certain angle, and then discovering that under a very specific tilt it does reflect. Why does that matter to you? Because these kinds of corrections change how we predict and interpret particle interactions, even if only in narrow circumstances.
The technical finding in plain words
Using a geometric condition on the directions and energies of the gluons, the team shows that the usual cancellation fails and the amplitude takes a simple, computable form. The authors worked the full integer cases separately and obtained very complicated expressions that grow superexponentially, and it was GPT-5.2 that proposed a far more compact form that generalizes across all cases.
This not only corrects an established idea about the vanishing of a certain amplitude, but also opens the door to extensions — for example to gravitons, where quantum gravity might show interesting analogues.
The role of AI in this process
GPT-5.2 Pro simplified very complex expressions and suggested a general conjecture based on concrete cases.
An internal model, using a guided reasoning process, spent around 12 hours producing a formal proof of the formula.
The authors checked the validity with standard analytical methods and consistency tests.
In other words, the AI acted as a discovery and proof-assistant, but final validation remained in human hands and followed the field’s formal checks.
For some experts, finding simple formulas is both delicate work and something that could be automated. This example shows that automation is becoming realistic and useful for concrete problems.
Why should this interest you even if you're not a physicist?
Because it’s a clear case of AI’s practical promise in science: not just finishing text or generating code, but helping spot patterns in enormous expressions and turning intuitions into testable conjectures. And—critically—doing so within processes that are then subjected to the scientific community’s rules.
It also raises legitimate questions. How exactly should we document model participation in scientific results? What review standards are needed to accept AI-assisted findings? These are discussions that will mature as more cases appear.
Looking ahead
The work has already extended gluon results to gravitons and announces more generalizations. The community is invited to review, reproduce, and comment on the preprint while the authors submit the work to journals. Broadly, this is an example of collaboration between human experts and AI tools: machines suggest patterns, humans verify and place findings in the wider conceptual framework.
Science doesn't change overnight, but it does change when tools accelerate the search for hidden simplicity in complicated expressions. Can you imagine what other fields might benefit from this AI-assisted recognition and proof strategy?
Summary: GPT-5.2 proposed and helped prove a new formula for a gluon amplitude previously considered zero in theoretical physics. A preprint on arXiv describes the finding, its analytical verification, and the implications of using AI as a scientific discovery tool.
Stay up to date!
Get AI news, tool launches, and innovative products straight to your inbox. Everything clear and useful.
GPT-5.2 and AI find new formula in theoretical physics
