[KongchangAI]
· 2 min read· 1,397 words

AI Automatically Verifies the Longest Proof in Mathematical History, a Key Step Toward Formalizing All of Mathematics

AI Automatically Verifies the Longest Proof in Mathematical History, a Key Step Toward Formalizing All of Mathematics

Axiom Math uses an AI cluster to automatically formalize CFSG and Stacks, proving that formalizing all of mathematics can be engineered.

Axiom Math's AI cluster completed two formalization feats: the 15,000-page Classification of Finite Simple Groups (CFSG) and the 7,500-page Stacks project, producing about 43 million and 16 million lines of code. CFSG tackled the engineering challenge of an ultra-long proof by breaking it into about 100 independently verifiable milestones; Stacks relied on a hierarchical AI agent system to ensure consistency across thousands of definitions. The formalization process also discovered and cleaned up several errors in the original proofs, but it has also raised concerns among mathematicians about "missing the opportunity to gain knowledge through hands-on work." The two achievements are seen as a key step toward formalizing all of mathematics.

Axiom Math's AI cluster completes mathematical formalization verification

Mathematical formalization has long been seen as a computational black hole: translating human-written proofs into machine-checkable code often takes longer than the proof itself. Axiom Math recently used an AI cluster to tackle two hard problems at once—the 15,000-page Classification of Finite Simple Groups (CFSG) and the 7,500-page Stacks project—giving the goal of "formalizing all of mathematics," which sounds like science fiction, an engineering path for the first time.

Two Hard Problems: One Extremely Long, One Extremely Wide

According to IEEE Spectrum, Axiom Math's AI cluster completed the final automated verification of CFSG on Sunday (October 25). CFSG, colloquially known as the "Enormous Theorem," is widely recognized as the longest mathematical proof ever written by humans. Ken Ono, the company's founding mathematician, remarked while watching the team count down to completing the verification: "I can't tell you how excited I am."

Colva Roney-Dougal, a group theorist at the University of St Andrews who was not involved in the work, was surprised by the result. She had regarded CFSG as a typical example of something "extremely difficult to formalize," believing the required compute was simply unattainable.

Just nine days later, Axiom Math announced the automated formalization of the Stacks project. This open-source collaborative mathematics textbook, initiated in 2005 by Columbia University mathematician Aise Johan de Jong, focuses on highly abstract algebraic geometry and algebraic stacks, and is a collection of thousands of interlocking definitions, constructions, and lemmas.

Axiom Math's Ken Ono summed up their complementarity in one sentence: "They show the two directions in which formalization must scale: one is a very long proof, and the other is a very wide web of interconnected theory."

The Classification of Finite Simple Groups (CFSG) is a foundational result in group theory and indeed in all of algebra. It asserts that every finite simple group—that is, a finite group with no nontrivial normal subgroups, analogous to primes among the integers—can be classified into about 18 infinite families or 26 sporadic simple groups. The proof of the theorem spans the second half of the 20th century, was completed接力 by over a hundred mathematicians, exceeds ten thousand pages in total, and has long been subject to disputes over whether certain subproofs are complete. Precisely because of its size and degree of interlocking, CFSG is regarded as the "Mount Everest" for testing formalization engineering capability: any local change can affect global consistency, making automated verification far more difficult than for ordinary theorems.

The Engineering Challenge of Long Proofs: Map First, Then Break into Milestones

The proof of CFSG is scattered across more than 500 journal papers, completed by about 100 authors over half a century. No individual mathematician can fully digest this interlocking web of knowledge.

Axiom Math's Niels Voss led the CFSG formalization. He said the hardest challenge was the sheer length of the proof: over 15,000 pages of dense mathematical content, containing a large number of theorems, lemmas, and definitions, and until all of it becomes machine-readable, no part of it can be checked.

The team's approach was to first process the entire CFSG literature, map the network structure of the informal proof, and then split the statement of CFSG into about 100 milestones—landmark results that can be independently verified along the way—using them to guide the formalization and maintain progress.

The resulting CFSG formalization code reached 43 million lines, roughly comparable in scale to the Linux kernel. What Roney-Dougal particularly appreciated was that the formal proof discovered and cleaned up several errors in the original human proof. She considers this crucial for the formalization of any modern research in group theory.

However, she also pointed out a concern: because the verification was done automatically, few mathematicians can currently extract information from the proof and share it with the community, and a new generation of mathematicians may "miss out on the opportunity to gain knowledge from the hands-on process."

Formal verification is fundamentally about having a computer check, line by line, whether each step of a proof conforms to axioms and inference rules; typical tools include proof assistants such as Lean, Coq, and Isabelle. Unlike natural-language proofs, formal code has no room for "obviously" or "left to the reader": every definition, lemma, and inference step must be explicitly given and accepted by the machine. This also explains why formal code is often tens of times longer than the original proof—43 million lines corresponding to 15,000 pages is a manifestation of this inflation. The value of formalization lies in providing absolute reliability and exposing hidden assumptions or typos in human proofs.

The Coordination Challenge of Wide Theories: Keeping Thousands of Definitions Aligned

The formalization of the Stacks project was led by Vasily Ilin. The final code was "only" 16 million lines, but the challenge was entirely different: maintaining faithfulness and consistency across thousands of interlocking definitions and theorems.

To this end, the team designed a hierarchical AI agent system, with agents having specialized responsibilities working in parallel to keep every definition in the library aligned, while checking faithfulness through multi-agent auditing, cross-checking against the source, and hard tests. This ensured that every element and task remained consistent throughout the project.

The impact of the Stacks formalization may be even more profound: it is now an efficient, general-purpose, reusable, and maintainable formal foundation library that anyone and any robotic program can access and build new formalized mathematics on top of.

The Stacks project, initiated by Aise Johan de Jong, aims to build an open, continuously updated modern textbook for algebraic geometry and algebraic stacks, and has been called "the Wikipedia of algebraic geometry." Algebraic stacks are a generalization of schemes, used to handle delicate issues such as quotient structures in moduli spaces, and their definitions are deeply nested and interdependent. Formalizing such a "wide" system is difficult not because of the length of any single proof, but because the interfaces between thousands of definitions must fit together perfectly: a slight adjustment to one definition may affect hundreds of downstream conclusions. This is precisely the consistency problem that the hierarchical multi-agent system needs to solve.

A hierarchical AI agent system refers to decomposing the formalization task into a pipeline shared by different specialized agents: some agents translate definitions, some verify lemma dependencies, and some cross-check against the original literature. Multiple agents working in parallel can increase throughput, but shared state and auditing mechanisms must ensure that their outputs are consistent with one another. This architecture draws on the idea of continuous integration in software engineering—every commit must pass hard tests, and any "hallucination" that deviates from the source is intercepted, thereby maintaining faithfulness while scaling automation.

After Formalization, What Do Human Mathematicians Do?

Ono and Voss see multiple advantages of formalized mathematics becoming central: rubber-stamp confirmation of theorems and proofs, preserving knowledge as expertise is lost over time, and building shared libraries as a foundation for creating new mathematics.

They also remain optimistic about the human role, viewing formalization as a launchpad for humans to use and understand mathematics in new ways. "Shouldn't we think more broadly about what the future of mathematics can be?" Ono said. "I'm excited about it, really very excited."

From an engineering perspective, the significance of these two achievements by Axiom Math is not in replacing mathematicians, but in demonstrating that AI clusters can handle the two extreme forms of formalization—ultra-long single proofs and ultra-wide theory libraries. When these two capabilities are combined, the formalization of all mathematics turns from a slogan into a problem that can be scheduled.

One-line Summary

Axiom Math uses an AI cluster to automatically formalize CFSG and Stacks, proving that formalizing all of mathematics can be engineered.

Paragraph Summary

Axiom Math's AI cluster has completed two formalization feats: the 15,000-page Classification of Finite Simple Groups (CFSG) and the 7,500-page Stacks project, producing about 43 million and 16 million lines of code, respectively. CFSG tackled the engineering challenge of an ultra-long proof by breaking it into about 100 independently verifiable milestones; Stacks relied on a hierarchical AI agent system to ensure consistency across thousands of definitions. The formalization process also discovered and cleaned up several errors in the original proofs, but it has also raised concerns among mathematicians about "missing the opportunity to gain knowledge through hands-on work." The two achievements are seen as a key step toward formalizing all of mathematics.

#Axiom Math#formal verification#Classification of Finite Simple Groups#Stacks project#AI cluster#mathematical formalization#Lean#proof assistant
Share:

Related articles