ArXiv: Munkres' Entire Topology Textbook Formalized in Isabelle/HOL with LLM Assistance

A record formalization of mathematics

Formalizing mathematics — writing proofs in a language a computer can mechanically verify — has historically been one of the slowest and most arduous tasks in science. A team led by Bryant published a paper on ArXiv on April 10 showing that this is changing drastically thanks to LLMs.

What they did

The team ran Munkres’ “General Topology” textbook in full through an automatic LLM-assisted pipeline that generates Isabelle/HOL code for every definition, lemma, and theorem. The final result:

• 85,000+ lines of verified Isabelle code
• All 806 formal results from the textbook covered
• Mechanical verification of every statement — the Isabelle prover explicitly accepts every proof

Why this matters

Munkres’ “General Topology” is a standard graduate-level textbook containing nontrivial mathematics — Tychonoff’s theorem, compactness, separation axioms. Manual formalization of such a volume would take years even for a team of specialized formalist mathematicians.

The LLM-assisted approach changes the equation in several ways:

• Scale — formalization becomes accessible for hundreds of textbooks, not just elite projects
• Proof certainty — every result is mechanically verified, not just “obvious” to the reader
• Training data — all the generated code becomes a new training corpus for the next generation of models

Specific tools

The pipeline combines an LLM (likely Claude or GPT-4 class) for generating proof hypotheses, the Isabelle prover for verification, and automatic strategies for filling in routine steps. When the LLM fails to produce a valid proof on the first try, the system uses feedback from the prover for iterative refinement.

The paper fits into the broader “AI for math” trend in which formal verification is finally beginning to scale. The next goal is likely the entire Springer Graduate Texts in Mathematics series.