OpenAI and collaborators extend single-minus amplitudes to gravitons

Research announcement

OpenAI published a new research update on March 4, 2026 describing joint work with researchers from the Max Planck Institute for Physics and the University of Chicago on extending single-minus scattering amplitudes from gluons to gravitons. In amplitude research, compact all-multiplicity structure is highly valuable because direct calculations become difficult very quickly as particle counts increase. The announcement frames this result as a continuation of recent progress in finding analytic structure that scales beyond low-point examples.

What this builds on

According to the post, a 2025 preprint by Bern and collaborators found nonzero all-multiplicity tree-level single-minus amplitudes in Yang-Mills theory. At that stage, no graviton counterpart was known. The new work reports conjectural all-multiplicity formulas for both mixed graviton-gluon amplitudes and pure graviton amplitudes, extending the scope from gauge-theory settings toward gravity-inclusive cases.

How GPT-5.2 Pro was used

OpenAI states that a team member and a collaborator used GPT-5.2 Pro while searching for and refining candidate formula structure. Importantly, the post also says that independent collaborators then produced rigorous mathematical proofs for the conjectured formulas. In addition to proof-level validation, the team reports independent numerical checks: up to six gravitons and fifteen gluons for the mixed case, and up to seven gravitons for the pure-graviton case.

Why the result is notable

The practical significance is less about replacing formal methods and more about a hybrid research workflow. AI-assisted hypothesis generation can reduce search time over large symbolic spaces, while human-led derivation and proof keep standards of correctness intact. In that sense, the work is presented as a method demonstration as much as a physics result: a concrete example of combining model-guided pattern discovery with strict independent verification.

For the broader AI-and-science landscape, the key next question is reproducibility across other hard symbolic domains. If similar workflows continue to produce verifiable outputs, this style of collaboration could become a repeatable pattern for theoretical research where candidate generation is expensive but final validation must remain rigorous.