AI-Assisted Physics Breakthrough: GPT-5.2 Derives Novel Gluon Amplitude Formula

For decades, theoretical physicists believed certain gluon interactions were mathematically impossible. Standard quantum field theory arguments suggested that tree-level scattering amplitudes must vanish when one gluon has negative helicity while all others have positive helicity. This assumption became textbook orthodoxy, leading researchers to largely ignore this configuration. Now, in a surprising twist, a new preprint co-authored by OpenAI researchers reveals this long-held conclusion was incomplete - and the discovery was facilitated by GPT-5.2.

The breakthrough centers on gluons, the force-carrying particles responsible for the strong nuclear interaction that binds quarks. Researchers identified a precisely defined kinematic regime called the "half-collinear limit" where the traditional arguments break down. In this specific momentum configuration - where gluon momenta follow a special alignment that's mathematically consistent but non-generic - the amplitude doesn't vanish. GPT-5.2 derived a compact formula (Eq. 39 in the preprint) that describes these previously overlooked interactions.

What makes this discovery remarkable is the process behind it. Human physicists first computed amplitudes manually for small particle counts (n=3 to n=6), resulting in increasingly complex expressions through Feynman diagram expansions. When presented with these unwieldy results, GPT-5.2 Pro simplified the expressions dramatically. More significantly, it detected a pattern across these cases and conjectured a general formula valid for any number of gluons. An internal scaffolded version of GPT-5.2 then independently spent approximately 12 hours constructing a formal proof of this formula's validity.

The formula was rigorously verified through multiple methods: it solves the Berends-Giele recursion relation (a fundamental technique for building multi-particle amplitudes) and satisfies the soft theorem governing particle behavior at low energy. According to co-author Alex Lupsasca of Vanderbilt University and OpenAI, "This opens the door to many new questions that will be the subject of subsequent investigations," including extensions to graviton interactions.

Reactions from prominent physicists highlight the significance:

• Nima Arkani-Hamed (Institute for Advanced Study): "To me, 'finding a simple formula' has always been fiddly, and something I've long felt might be automatable by computers. This example seems especially well-suited to exploit modern AI tools. I look forward to seeing this trend continue."
• Nathaniel Craig (UC Santa Barbara): "This preprint felt like a glimpse into the future of AI-assisted science. By coupling GPT-5.2 with human domain experts, the paper provides a template for validating LLM-driven insights."

The research demonstrates AI's emerging capability not just as a computational tool but as a collaborator in fundamental discovery. While traditional methods produced super-exponentially complex expressions, the AI identified elegant simplicity where humans saw only zero. This collaboration has already yielded extensions to graviton physics, suggesting we may be witnessing the early stages of a paradigm shift in theoretical physics methodology.

Important questions remain: How general is this approach beyond specific kinematic regimes? Can AI similarly simplify other intractable problems across physics? What safeguards ensure the validity of AI-generated conjectures? As noted in the preprint, these AI-assisted results "satisfy what we expect from rigorous scientific inquiry" through multi-layered verification - suggesting a promising path forward for human-AI research partnerships in theoretical domains.