When two physicists hit a wall proving a decade-old identity in jamming theory, they turned to Claude for ideas - and got a viable one. The episode is a rare, well-documented case of a large language model contributing to a peer-reviewed physics proof, and it clarifies where LLMs currently add value in formal research: generating tractable directions an expert can verify and refine, not delivering a finished, checked derivation.
What happened
Giorgio Parisi, the 2021 Nobel laureate in physics, and Francesco Zamponi of Sapienza University of Rome had spent years unable to prove why two independently developed theories of the jamming transition - one from Parisi's group, one from Matthieu Wyart's - both predict the same critical exponents, an identity of the form a+b=1. According to Zamponi, the pair asked Claude (Sonnet 4.6, later Opus 4.7) for help: "Quite quickly, Claude came up with an initial idea that was essentially correct," even though the model's first attempt at a formal proof contained errors. Parisi and Zamponi refined that idea into a rigorous proof, published as "A proof of an identity for the critical exponents of jamming" in the Journal of Statistical Mechanics: Theory and Experiment (arXiv:2606.03300).
Technical context
Jamming describes how disordered systems - granular material, foams, dense colloids - abruptly transition from flowing to rigid as density increases without adopting crystalline order. The two competing theoretical routes to the jamming critical exponents were developed independently and were not previously known to reduce to one another; the new proof shows they are mathematically equivalent, connecting two branches of the field's full replica-symmetry-breaking (fullRSB) formalism.
For practitioners
The workflow behind this result relied on tight human supervision: researchers guided Claude to first reproduce known numerical results, then posed the open proof question, then corrected and iterated on its output across multiple rounds before accepting the underlying idea. Teams exploring LLMs for research should expect the same pattern - models are more reliable at surfacing candidate approaches or simplifications than at producing a checked, publishable proof outright, and rigorous human verification remains both the bottleneck and the safeguard.
What to watch
Whether other theory groups report similar model-assisted proof workflows, and whether journals begin asking authors to disclose AI-assisted derivations as routine practice, are the next signals worth tracking as this kind of collaboration becomes less novel.
Key Points
- 1Physicists Giorgio Parisi and Francesco Zamponi used Claude to help prove a jamming-theory identity unsolved for over a decade.
- 2The model's first proof attempt contained errors, but researchers judged its core approach correct and refined it into a rigorous result.
- 3The case shows LLMs currently add the most research value as idea generators, with human experts still required for formal verification.
Scoring Rationale
A verified, peer-reviewed case of an LLM materially contributing to a real physics proof from a Nobel laureate's team, corroborated by multiple outlets and the arXiv preprint. Notable for practitioners as a documented example of AI-assisted formal research, though it is a single result rather than a broad methodological or industry shift.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
