Axiom Math Claims AI Agents Formalized Math's 15,000-Page Enormous Theorem in 24 Days
Axiom Math announced that its AxiomProver system converted the Classification of Finite Simple Groups into 43.1 million lines of computer-checked code, uncovering 34 literature errors along the way.
In short
- Axiom Math announced that dozens of AI agents organized under its AxiomProver system converted the 15,000-page Classification of Finite Simple Groups into machine-checked Lean code in 24 days.
- The resulting formalization spans 43.1 million lines of code and 2.05 million declarations, building on Mathlib and 1.2 million lines reused from the Qiuzhen-CFSG project.
- The agents identified 34 noteworthy deviations from the mathematical literature, including a flaw in an intermediate 1986 proof step that required producing an alternative argument.
- The formal proof currently represents a self-assessed achievement by Axiom Math; independent researchers have not yet executed a third-party rebuild of the complete repository.
On October 26, 2026, Palo Alto startup Axiom Math announced that swarms of AI agents had successfully translated the Classification of Finite Simple Groups—often called "the enormous theorem"—into computer-verifiable code in 24 days. The human mathematical proof for the theorem, established across decades by roughly 100 mathematicians, spans approximately 15,000 pages across more than 500 published papers. According to Axiom Math's project website and news coverage, the automated process produced roughly 43.1 million lines of code in the Lean proof assistant language.
The project marks a major shift in the scale of automated proof checking. In a mathematical proof assistant like Lean, an argument is validated line by line through an algorithmic kernel that accepts deductions only if they strictly satisfy core logical axioms without placeholders. Rather than discovering a fundamentally new classification theorem from scratch, Axiom Math targeted the existing human literature, systematically encoding definitions, intermediate lemmas, and overarching deductions into a single executable repository.
Cataloging the Atoms of Symmetry
The Classification of Finite Simple Groups serves as a definitive catalog of symmetry in mathematics. In group theory, a group captures the ways to manipulate or transform an object while leaving its overall structure intact. A standard cube, for example, possesses 24 rotational symmetries that return it to an identical appearance. Just as chemical molecules break down into indivisible chemical elements, finite groups break down into indivisible components known as finite simple groups.
The classification theorem proves that all finite simple groups belong to a set of known families or a finite collection of exceptions, as reported by DailyAI. These include cyclic groups of prime order, alternating groups on five or more symbols, 16 infinite families of groups of Lie type, and exactly 26 sporadic groups that fall outside those regular patterns. The largest of these exceptions, the Monster group, has an order of approximately 8 × 10⁵³ elements.
Mathematician Daniel Gorenstein announced the classification completed in 1983, but subsequent analysis revealed substantial gaps. A major omission involving the quasithin case was not fully resolved until Michael Aschbacher and Stephen Smith published a 1,221-page work in 2004, followed by further corrections in 2008. As group theorist Michael Aschbacher remarked in 2005, writing a massive and complex mathematical argument without errors is practically impossible for humans.
Inside the 24-Day AxiomProver Workflow
Axiom Math launched the formalization initiative on October 1, 2026, setting up an initial operational outline with 108 milestones and roughly 380 smaller subtasks. The system deployed on October 2 under an architecture named AxiomProver, as reported by DailyAI and Fast Company. The platform orchestrated dozens of AI agents running across a dynamically shifting selection of underlying AI models, with human mathematicians resolving theoretical obstructions and software engineers managing build infrastructure.
Lead agents decomposed complex mathematical arguments into subproblems and verified formal statements, while subordinate agents retrieved known results and generated Lean code. Coordination took place across GitHub issues, with operational updates directed to Slack. The team also incorporated prior formalization efforts, integrating Lean's community-driven Mathlib library and approximately 1.2 million lines of code from the Qiuzhen-CFSG project, while borrowing the high-level formulation of the classification statement from the TauCeti repository.
The team encountered severe friction during the project's early phases. Some early milestones hid unproved assumptions that agents took for granted, forcing engineers to delete tens of thousands of lines of dependent code and restart entire sections. On October 13, progress slowed to a crawl, completing only a single milestone in 24 hours. Axiom reorganized the pipeline into smaller, strictly defined tasks. As agent drafting accelerated, Lean's compilation engine struggled to keep pace, prompting engineers to implement a distributed build system to verify proofs across machines.
Uncovering Flaws in the Mathematical Literature
While encoding the literature into formal code, the system uncovered 34 noteworthy deviations where published claims did not align with mathematical reality. Most discrepancies involved typographical errors, missed boundary preconditions, or misstated table values, but several were substantial.
In Aschbacher and Smith's 2004 quasithin volumes, agents located an error in a lemma asserting that two quantities were always equal. The system identified a counterexample using the group SL₂(5), where one quantity evaluated to 1 and the other evaluated to 2. To preserve the proof chain, mathematicians adjusted the claim to an inequality and re-derived the downstream arguments, as detailed by MACCUS.
A larger issue emerged within the Unbalanced Group Theorem (UGT), a core pillar of the classification. When formalizing John Walter's 1986 memoir, the agents found a counterexample to a "closure assertion" that assumed a specific operation kept simple components within an intended family. Using the group PSL₂(49), the system showed that an order-two field automorphism produced the centralizer PGL₂(7) with simple component PSL₂(7), a group explicitly excluded from the family because 7 is a Mersenne prime. While this did not invalidate the overarching theorem, it broke Walter's published inference step.
To repair the gap, Axiom Math consulted original classification contributors Richard Lyons and Ronald Solomon. Axiom stated that, drawing on Solomon's 1976 theory of maximal 2-components, AxiomProver produced and formalized an alternative path, establishing Theorem 2.3 ("Preservation of simple tops"), though the accompanying paper is labeled "primarily human-written". Axiom Math published this repair in a preprint on Hexagon, authored by Colin Defant and Ken Ono. The paper also documents a subsequent, shorter deduction discovered by Axiom mathematicians that bypasses Theorem 2.3 entirely.
Independent Audits and Open Questions
Axiom Math has released the formal proof code across a public GitHub repository, containing roughly 2.05 million declarations across 43.1 million lines of code. However, as noted in analyses by Kingy AI and Xenospectrum, the achievement remains self-assessed.
No third-party research group has completed an independent rebuild or mathematical audit of the full repository. Furthermore, Axiom Math noted in its repository documentation that compiling the entire proof in under 24 hours requires unreleased modifications to Lean's Lake build system that have yet to be upstreamed. Independent experts also caution that while Lean's kernel guarantees internal deductive consistency, human specialists must inspect the top-level definitions to confirm that the formal code faithfully reflects the intended mathematical concepts rather than a weakened assertion.
The project demonstrates how AI agents combined with algorithmic verification engines can process and debug interconnected scientific literature. As independent group theorists evaluate the preprint and attempt third-party builds, the release provides a concrete foundation for testing automated reasoning on the largest structures in pure mathematics.
Questions people ask
What is the Classification of Finite Simple Groups?
It is a central theorem in abstract algebra that provides a complete catalog of finite simple groups, which serve as the fundamental building blocks of all finite symmetry. The original human proof spans roughly 15,000 pages across more than 500 published papers.
Did the AI discover a new mathematical classification?
No. The AI system verified and formalized the existing human proof into the Lean proof assistant rather than discovering a new classification. However, Axiom claims that while repairing a flawed step in an intermediate 1986 paper, AxiomProver proved a new result known as Theorem 2.3, though Axiom also stated the system verified proofs rather than discovering new ones, and the accompanying preprint is labeled "primarily human-written".
Has the formal proof been independently verified by outside mathematicians?
Not yet. While Axiom Math published its codebase on GitHub, as of late October 2026 no external party had reproduced the build or completed an independent audit of the entire repository.
Sources
- Formalizing the Enormous TheoremAxiom
- Anthropic's IPO is selling investors a story. The numbers tell another onefastcompany.com, Oct 29, 2026
- Agenti di intelligenza artificiale verificano la dimostrazione matematica più lunga in 24 giornidailyai.com, Oct 28, 2026
- AI Formalizes the Classification of Finite Simple Groups in 24 Days, Exposing a Flaw in an Intermediate Stepxenospectrum.com, Oct 28, 2026
- Axiom Says AI Formalized a 15,000-Page Proof in 24 DaysKingy AI, Oct 27, 2026
- Axiom used AI agents to turn math's 15,000-page proof into computer-checked code in 24 daysMACCUS - SEO, Influence & Organic Marketing Agency, Oct 29, 2026
How this story was made: the newsroom picked it up from Reddit, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (49 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 30, 2026 at 02:29 UTC
