VibeMathedMath problems solved with AI

A new bound for small gaps between primes: H1212H_1 \le 212

Write H1=lim infn(pn+1pn)H_1 = \liminf_{n\to\infty}(p_{n+1}-p_n). Stadlmann had recently proved H1240H_1 \le 240, improving the bound 246246 of Polymath8b. Building on her work, this paper proves H1212H_1 \le 212: infinitely many pairs of consecutive primes are at most 212212 apart.

Result
Proved(see note)
Status
Partial result
AI contribution
AI-assisted
Method
Argument
Field
Analytic number theory
Posed by
Alphonse de Polignac (the twin prime conjecture); the bounded form since Goldston, Pintz and Yıldırım
Year posed
1849
Years open
177y
Solved
2026-09-03
Model
AxiomProver
Vendor
Axiom Math
Collaborators
François Charton, Letong Hong, Kenny Lau, Ken Ono, Guillaume Remy, Ho Chung Siu, Ashvin A. Swaminathan, Jesse Thorner, Yunzhou Xie
Verification
Lean-checked, statement unaudited
Publication
Preprint
Significance
62 / 100
Disclosed cost
Wikipedia
No dedicated article

What was actually shown

H1212H_1 \le 212, improving Stadlmann's 240240 of three days earlier and the 246246 of Polymath8b that had stood since 2014. The twin prime conjecture, H1=2H_1 = 2, is untouched. Held the record for hours at most: OpenAI's paper claiming 186186 is dated 30 August, four days before this one, though its Lean development appeared on 2 September.

What the AI did

The mathematics is the authors'. The AI contribution is the formal certificate, and the paper is precise about it in Appendix A: "AxiomProver, an AI system under development by AxiomMath, autonomously generated from natural-language specifications a Lean certificate of the deduction of Theorem 1.1." The certificate takes as hypotheses the five Type I, Type II and Type III equidistribution estimates of Section 5, the bilinear Bombieri-Vinogradov theorem below the half-level, the Harman decomposition, and the variational certificate. Nothing in the paper claims the model found the argument, and the abstract does not mention AI at all.

The same group's AxiomProver had formalised the twelve-year-old 246 bound in Lean a few weeks earlier, which is the work this builds its tooling on.

Verification

A preprint one day old, not peer reviewed. Its Lean certificate was produced by AxiomProver and is conditional on the equidistribution estimates and the Bombieri-Vinogradov theorem stated in the paper, so it certifies the deduction rather than the analytic inputs. Nine authors, several of whom work on exactly this, take responsibility for the mathematics. No independent expert has read it on the record, and this site has not rebuilt the certificate.

Sources

Changelog1 change

Discussion