Live·Open questions in longevity research
All news
Science ResearchScientific Computing

OpenAI’s code proves that a quantum field theory satisfying the Wightman axioms can be constructed as a limit of lattice Yang-Mills theory, while uniqueness of the limit and a mass gap remain unproved, David Gross’s committee concludes

11 November 2026· 261111003

OpenAI’s code proves that a quantum field theory satisfying the Wightman axioms can be constructed as a limit of lattice Yang-Mills theory, while uniqueness of the limit and a mass gap remain unproved, David Gross’s committee concludes

On 10 November, the committee published its assessment and the Lean code submitted by OpenAI on 28 October. Lean checks every step of the proof, which OpenAI says its models discovered. Yang-Mills theory describes gluons, the particles that hold quarks together inside protons and neutrons. The Clay Mathematics Institute’s Millennium Prize Problem requires a proof that this theory rigorously exists in four dimensions and has a mass gap: the smallest nonzero energy that can be added to a state with no particles.

In their statement of the problem, Arthur Jaffe and Edward Witten wrote:

“At present, we know of no nontrivial relativistic field theory that satisfies the Wightman axioms (or any other reasonable axioms) in four dimensions.”

The Wightman axioms specify the conditions for a rigorous quantum theory consistent with relativity. The code formalizes the existence of a pure gluon theory only. This is one part of the claimed proof of the existence of quantum chromodynamics, a theory that includes quarks.

The Wilson model approximates the gluon theory by replacing spacetime within a box with a four-dimensional lattice. The lattice spacing is taken to zero while the coupling constant is adjusted according to asymptotic freedom: the shorter the distance, the weaker the interaction. The box is then expanded to infinite size. Inequalities keep correlations between points from diverging, allowing a convergent subsequence to be selected as the boxes grow. Whether different subsequences converge to the same limit remains unknown.

The Wightman axioms require a unique vacuum, a state with no particles. Uniqueness follows from clustering, which means that distant regions are independent, but clustering has not been proved for the limiting theory as a whole. The code therefore decomposes the limit into “pure phases” and proves clustering within each phase. A reconstruction theorem, which recovers a theory from the limiting correlations, then yields a Wightman theory with a unique vacuum for each phase.

As evidence that the resulting theory is indeed Yang-Mills theory, the committee cites the weakening of the coupling at short distances and the behavior of the trace of the conserved energy-momentum tensor. This tensor describes the flow of energy and momentum; its trace is the sum of its diagonal components. The trace is zero in the classical theory and nonzero in the constructed theory, where its correlation at two nearby points follows a law containing a logarithm. The coefficient in this law agrees with the result from perturbation theory, which calculates small corrections, and finite collections of free fields do not produce such a logarithm.

All 9 470 modules were rebuilt from scratch without errors or placeholders standing in for proofs. The Comparator program checks the theorem against a reference statement to prevent its conditions from being changed. It accepted two statements: OpenAI’s public version, which the company later said omitted some connections to the lattice construction, and the committee’s stronger version.

Lean checks deductions from the definitions as written. The committee examined their correspondence to the physics extensively, though not exhaustively. It used OpenAI’s Astra model to analyze the code and writes: “We do not yet claim to understand the proof fully.”

The committee considers the absence of a uniqueness proof consistent with the general expectation that the limit and the true vacuum of Yang-Mills theory are unique. The committee’s repository lists a second submission from OpenAI, dated 4 November, with its assessment marked “in progress.”

Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#yang-mills#quantum-field-theory#lean#wightman-axioms#mass-gap#formal-verification