Axiom Math verifies proof of the 246 theorem on prime gaps
Axiom Math, an AI-for-mathematics startup, has used its AI system AxiomProver to automatically verify the formal proof of the "246 theorem," a landmark result in number theory. The 246 theorem states that there are infinitely many pairs of primes that differ by exactly 246, making it the closest mathematicians have gotten to proving the still-unsolved twin prime conjecture, which holds that there are infinitely many pairs of primes separated by just two.
The twin prime conjecture was first precisely formulated in the 19th century by French mathematician Alphonse de Polignac. The first real progress toward it came in 2013, when Yitang Zhang, now a professor at Sun Yat-sen University in Guangzhou, China, proved that infinitely many pairs of primes are separated by a gap of 70 million. Months later, University of Oxford professor James Maynard used a different technique to cut that gap dramatically, from 70 million down to 600, a result that helped earn him the 2022 Fields Medal. Maynard and fellow Fields Medalist Terence Tao, a professor at the University of California, Los Angeles, then worked with the Polymath8b collaboration to bring the gap down further, to 246. It is this result, the 246 theorem, that AxiomProver has now verified.
Ken Ono, Axiom Math's founding mathematician, called the 246 theorem "the threshold of human knowledge about prime numbers." Axiom Math has already used AxiomProver, an autonomous multi-agent system that converts mathematical statements into machine-checkable proofs, to crack several unsolved problems and formalize other proofs this year, but the company frames the 246 theorem as its most significant result yet. Earlier this year, rival Math, Inc. used its own Gauss agent to formalize Maryna Viazovska's 2022 Fields Medal-winning proof of the sphere-packing problem in 8 and 24 dimensions. Sidharth Hariharan, a Carnegie Mellon University Ph.D. student who led the human effort behind that earlier formalization and is now an intern at Axiom Math, says formalizing the 246 theorem is a more comprehensive and useful achievement: rather than a one-shot effort tied to a single problem, Axiom Math built the work so its components can be reused for other formalization tasks, assembling a library of results about prime gaps with the 246 theorem as its flagship entry.
Formal verification means a computer checks a machine-readable version of a proof line by line. The article notes this is not a full guarantee of correctness: a recent demonstration showed that a bug in the verification method could be exploited to get a false, AI-generated proof accepted, even though the process is close to a rubber stamp. Ono argues the bigger significance lies beyond number theory. Because these techniques underpin cryptography and cybersecurity, they could help verify how digital data is kept secure. More broadly, Ono sees formal proof verification as a stepping stone toward verifying AI-generated computer code, checking properties like whether an algorithm terminates or whether a program's output is correct for any input, by translating them into precise mathematical statements that AxiomProver-derived tools could then prove. "The world is about to run on computer code that nobody has read," Ono said.
Key facts
- Axiom Math's AI system AxiomProver has, for the first time, automatically verified the formal proof of the 246 theorem, which shows there are infinitely many pairs of primes differing by 246.
- The 246 theorem is the closest progress yet toward the unproven twin prime conjecture (infinitely many primes differing by two), first formulated in the 19th century by Alphonse de Polignac.
- The result traces through Yitang Zhang's 2013 proof of a 70 million prime gap, James Maynard's reduction of that gap to 600 (part of his 2022 Fields Medal-winning work), and the Polymath8b collaboration with Terence Tao that brought it down to 246.
- Sidharth Hariharan, a Carnegie Mellon Ph.D. student who led the human formalization of Maryna Viazovska's Fields Medal-winning sphere-packing proof (formalized this year by rival Math, Inc.'s Gauss agent) and is now an Axiom Math intern, calls the 246 theorem formalization more comprehensive and reusable.
- Founding mathematician Ken Ono frames the achievement as a stepping stone toward using AI to verify the correctness of AI-generated code, though the article notes formal verification is not a 100 percent guarantee against bugs.
Why it matters
The 246 theorem is, in Ken Ono's words, "the threshold of human knowledge about prime numbers," the closest anyone has come to the still-unproven twin prime conjecture. Getting an AI system to formally verify a proof at this level shows automated formalization has moved from toy problems to genuinely hard, current mathematics. Axiom Math frames the achievement as more than a one-off: it built the formalization so its components form a reusable library of results about prime gaps, with the 246 theorem as the flagship entry.
Who it affects
The result matters most to mathematicians working on formal verification and number theory, and to the small set of companies, including Axiom Math and rival Math, Inc., competing to automate mathematical formalization. Ono also points to a wider audience: number theory underlies present-day cryptography and cybersecurity, and he argues the same verification techniques could eventually be applied to AI-generated software running infrastructure, finance, and data-protection systems.
How to use it
This is a research result, not a shipped product. Axiom Math's practical offer is the library itself: components of the 246 theorem formalization are built to be reused in other formalization tasks rather than solving one problem and stopping. Ono's longer-term pitch is a method, not a tool release: translate properties of code, such as whether an algorithm terminates or produces correct output for any input, into precise mathematical statements that AxiomProver-derived technology could then prove, as a way to check AI-generated code before it is trusted.
How solid is it
The verification is a formal proof check, a computer confirming a machine-readable version of the proof line by line, which the article calls close to a rubber stamp. But it is not an absolute guarantee: the article notes a recent demonstration exposed a bug in the verification method that could be exploited to get a false, AI-generated proof accepted. The 246 theorem result itself builds on proofs already established by human mathematicians, Zhang, Maynard, and the Polymath8b collaboration with Tao; AxiomProver's contribution is the automated formal verification of that existing proof, not a new mathematical discovery.
Risks and caveats
The twin prime conjecture that motivates this work, that primes differing by just two occur infinitely often, remains unproven; the 246 theorem only narrows the gap to 246, not two. The article gives no specific date for when AxiomProver completed the verification. And the formal-verification method itself has a demonstrated weak point: a bug that could let a false AI-generated proof pass as verified, which bears directly on Ono's own pitch of using the same approach to certify AI-generated code.
“The world is about to run on computer code that nobody has read.”
— Ken Ono, Axiom Math's founding mathematician