AI Yang-Mills proof goes public, and a Nobel laureate's panel gives a cautious nod

What's going on

OpenAI says one of its internal AI models has found a mathematically rigorous proof that quantum Yang-Mills theory exists. This theory is the mathematical foundation of the Standard Model of particle physics. The company announced the claim back in late October 2026, together with two other AI results in mathematics and computer science. Unlike with the other two results, OpenAI didn't publish the material itself at first. Instead, it handed the proof to an outside panel of physicists and mathematicians.

This Yang-Mills working group is led by physics Nobel laureate David J. Gross. On November 9 it announced it would release the material, and on November 10 two GitHub repositories followed. The first contains OpenAI's complete proof in Lean, a so-called proof assistant. In Lean, a mathematical argument is written down so precisely that a computer can check every single step. The second repository contains the panel's assessment. The code is available at github.com/yang-mills-group/OpenAI-YM-20261028. The proof folder alone holds dozens of subfolders with names like "Lattice", "RG", "Wilson", "OS" and "Spectral".

What Yang-Mills is and why the problem is so famous

The easiest way into Yang-Mills is through electromagnetism. Physicists describe it as a gauge theory, a mathematical framework built on a particular symmetry that determines how particles interact. In electromagnetism, that symmetry is very simple, basically a rotation around a circle. In 1954, Chen-Ning Yang and Robert Mills extended the idea to more complicated symmetries where the order of operations matters. That generalization is exactly what describes the strong nuclear force, which holds quarks together inside protons and neutrons, and the weak force, which drives certain radioactive decays. The theory of the strong force is called quantum chromodynamics (QCD).

In practice, these theories work brilliantly. Their predictions have been confirmed at particle accelerators like the LHC for decades. But nobody has ever rigorously proven that the theory is even consistently defined at all energy scales. Officechai puts it this way: physicists are running a program that works beautifully, without being able to prove it will never crash.

In 2000, the Clay Mathematics Institute made this one of its seven Millennium Prize Problems, each carrying a $1 million prize. Arthur Jaffe and Edward Witten wrote the problem statement. It asks for two things. For every compact simple gauge group, meaning every suitable symmetry structure, a non-trivial quantum Yang-Mills theory must exist in four dimensions. And that theory must have a mass gap. This means there is a real, positive gap between the vacuum at zero energy and the first excited state. Among other things, that would explain why quarks and gluons are never observed on their own in nature.

What OpenAI claims and how the proof is built

OpenAI explicitly claims only existence, not the mass gap. According to Officechai, the claim actually goes beyond pure Yang-Mills. The model reportedly delivered existence proofs for a whole family of four-dimensional quantum field theories with gauge fields and fermions (matter particles like quarks and electrons). These are said to include QCD and a "unified completion" of the Standard Model. According to OpenAI, the part covering pure Yang-Mills, meaning gluons that only interact with each other, is "part" of the Millennium Problem.

The timeline according to OpenAI and the panel:

According to Crypto Briefing, the submitted Lean proof runs to 1,147 pages. It builds on established tools. One is Wilson's lattice regularization, where spacetime is first replaced by a grid of points, much like a photo made of pixels. The other is the renormalization group methods developed by mathematician Tadeusz Balaban in the 1980s, which move step by step from coarse to fine scales. Balaban's program was never fully completed for theories this complex.

Kingy AI explains the construction in more detail. On a finite lattice, group elements are assigned to the links between neighboring points. A probability distribution weights the possible configurations based on the smallest squares of the lattice, called plaquettes. The real work is showing that meaningful limits emerge as the lattice spacing goes to zero and the volume grows. From the resulting "Euclidean" correlations, a relativistic quantum field theory is then reconstructed, complete with a state space, fields, a vacuum, spacetime symmetry, positive energy and locality. On top of that comes a short-distance test meant to show the theory is non-trivial, meaning it differs from a free theory with no interactions.

What the panel checked and where it adds caveats

Besides Gross, the panel includes Michael Douglas, who has tracked the state of the problem for the Clay Institute for two decades, and Arthur Jaffe, co-author of the original problem statement and co-founder of the Clay Institute. Also on board are Igor Klebanov, an expert on confinement (the trapping of quarks), and Nikita Nekrasov, winner of the 2023 Dannie Heineman Prize. OpenAI first approached Gross, who picked the other members himself.

According to Crypto Briefing, the reviewers focused mainly on one question: do the definitions in the proof match the ones the field actually uses? A formal proof can be perfectly correct and still prove the wrong thing. If the definitions are even slightly off, the computer will happily confirm the answer to a question nobody asked.

In its 35-page assessment, the panel reaches a verdict that Kingy AI summarizes like this: the construction is Yang-Mills in the sense of an existence criterion that doesn't require a precise characterization of the vacuum. A closer look reveals several important caveats:

• No mass gap: The construction provides "no evidence for or against" a mass gap.
• No uniqueness: When passing to infinite volume, the proof allows a subsequence to be selected. So it shows that there is one path to the desired object, but not that every path leads to the same theory.
• A revised target: According to the assessment, the original Lean target "jointExistence" left out links to the lattice construction that appear elsewhere in the proof. The verdict therefore rests on the stricter target "combined_acceptance". It ensures that the lattice family, the limits and the reconstructed theory really describe the same object.
• Limited non-triviality test: The test rules out certain free alternatives, but not every conceivable one.

Kingy AI stresses that it did not rerun the Lean checks itself.

Background: OpenAI more cautious after the Navier-Stokes debate

According to Officechai, OpenAI's restraint with this result is a lesson from the dispute over its claimed solution to another Millennium Problem, the Navier-Stokes equations. Critics questioned whether the version it solved actually matched the Clay Institute's definition. OpenAI published its math results directly and brought in an outside group for the computer science result. For the physics result, the company let the panel decide whether and when to publish. OpenAI isn't the only one working on this, either. According to Crypto Briefing, Ethan Sussman and Levent Alpöge are pursuing approaches from constructive quantum field theory in their own repository.

The bigger picture

If the proof holds up under further scrutiny, it would be a genuine milestone for mathematical physics. A rigorous construction of the Standard Model's gauge theories has been considered a central open problem since the 1960s and 70s. It won't produce new physical predictions. Officechai compares it to calculus, which engineers used for centuries before mathematicians put it on solid footing. It's also remarkable that an AI model is said to have produced such a long, deep argument and formalized it in machine-checkable form.

Still, the "solved and verified" headline making the rounds on Reddit and elsewhere is wrong. What's verified so far is that the Lean code meets its stated target, and a top-tier panel considers that target a valid version of the existence question. The mass gap, which is the harder half of the Millennium Problem, remains open. The prize hasn't been awarded, and neither confinement nor a full solution of QCD has been proven. The broader claims about QCD and the Standard Model haven't been publicly reviewed at all. Now it's up to the wider research community. Only once independent mathematicians have worked through the definitions and the roughly 1,100 pages of Lean code will we know how solid the result really is.