Company Updates

OpenAI's AI proof of Yang-Mills existence passes Lean verification

Share
OpenAI's AI proof of Yang-Mills existence passes Lean verification

OpenAI's internal model produced a proof for the existence half of the Yang-Mills Millennium Prize Problem, which was then formally verified in the Lean proof language by a committee led by physicist David Gross. The mass gap, the harder half of the problem, remains unresolved.

TL;DR

  • OpenAI's AI-generated proof for the Yang-Mills existence problem was formally verified in Lean, marking a milestone in AI-assisted mathematics.
  • The mass gap, the second half of the Millennium Prize Problem, remains unaddressed and unverified.
  • The verification process sparked debates about transparency, peer review, and the role of AI in formal mathematics.

What happened

On 26 October 2026, OpenAI announced that an internal model had generated a proof for the existence half of the Yang-Mills Millennium Prize Problem, one of the Clay Mathematics Institute's seven unsolved problems worth $1 million. The company stated the result was achieved on 9 September 2026 but was kept private until encoded in Lean, a formal proof language. OpenAI did not release the proof publicly but instead shared it with a committee of experts for review.

The committee, led by physicist David Gross and including Michael R. Douglas of Harvard, issued a statement on 9 November confirming the proof's validity in Lean. On 10 November, they published a GitHub document explaining the proof's structure and affirming its formal verification. Scientific American reported on these developments on 11 November, with updates on 12 November.

The committee's statement emphasized that while the proof was verified, they did not fully understand it. The explanatory document was described as cursory, and the committee used OpenAI's Astra model to parse the Lean code. The mass gap, the second and harder half of the Yang-Mills problem, was explicitly left untouched.

The verification process drew criticism for its lack of transparency. Some mathematicians argued that the review process conflicted with transparency recommendations from an OpenAI-affiliated advisory group. Others, like NYU's Scott Armstrong, suggested the review could be conducted more efficiently in the open. Meanwhile, a parallel effort led by Levent Alpöge of Anthropic and Northwestern's Ethan Sussman claimed progress on a three-dimensional version of the existence problem, though this does not meet the Clay problem's four-dimensional criteria.

Why it matters

This milestone highlights the growing role of AI in formal mathematics and proof verification, offering a template for how AI-generated proofs might be vetted in the future. For researchers and startups, it underscores the importance of transparency and reproducibility in AI-assisted discoveries, particularly when formal verification tools like Lean are involved.

The debate over the verification process reflects broader tensions in AI-driven research: who controls the release of AI-generated insights, how peer review adapts to AI outputs, and whether formal verification alone is sufficient to validate a result. For investors and founders, this raises questions about the reliability of AI-generated proofs and the need for robust validation frameworks.

The unresolved mass gap and the committee's admission of incomplete understanding of the proof serve as a reminder that AI's role in mathematics is still evolving. While AI can assist in generating and verifying proofs, human expertise remains critical for interpreting results and ensuring they align with the intended problem statements.

The process also highlights the competitive landscape among AI labs. OpenAI's approach—controlling the release and shepherding the proof through a private committee—contrasts with open alternatives, such as the parallel efforts using Anthropic's Claude-based agents. This divergence could shape future norms for AI-assisted research and publication.

Key facts

  • OpenAI's internal model generated a proof for the existence half of the Yang-Mills Millennium Prize Problem on 9 September 2026, according to the company.
  • The proof was formally verified in Lean by a committee led by David Gross on 9–10 November 2026.
  • The committee included Michael R. Douglas of Harvard and was formed to shepherd the proof into the mathematical literature.
  • The verification process did not address the mass gap, the second and harder half of the Yang-Mills problem, which remains unresolved.
  • The committee stated it did not fully understand the proof, describing the explanatory document as cursory.
  • A parallel effort led by Levent Alpöge of Anthropic and Northwestern's Ethan Sussman claimed progress on a three-dimensional version of the existence problem, though it does not meet the Clay problem's four-dimensional criteria.
  • The verification process sparked debates about transparency, with critics arguing the review could be conducted more efficiently in the open.
  • The Clay Mathematics Institute's Millennium Prize Problem offers a $1 million reward for solving the Yang-Mills problem in full.

Context

The Yang-Mills Millennium Prize Problem is one of seven unsolved problems in mathematics, with a $1 million prize for a complete solution. The problem has two parts: proving the existence of a rigorous four-dimensional theory and demonstrating the emergence of a mass gap in its equations.

Formal proof languages like Lean are increasingly used to verify mathematical proofs, offering a way to check the correctness of complex arguments with computational precision. This approach is gaining traction in both pure mathematics and AI-assisted research.

The role of AI in mathematics has grown rapidly, with models like OpenAI's Astra and Anthropic's Claude being used to generate, parse, and verify proofs. This development is reshaping how researchers approach problem-solving and peer review, though it also raises questions about transparency and accountability.

Topics

Related coverage

Join the discussion

Have a take on this story? Weigh in with our community on Facebook.

💬 Discuss on Facebook →