AI math firm Axiom used dozens of AI agents to formalize and verify the 15,000-page Classification of Finite Simple Groups theorem in just 24 days using Lean. The process uncovered a logical flaw in an intermediate lemma, which the AI resolved by generating a new proof strategy that passed rigorous computer verification.
Autoformalization
自动形式化
Converting natural language math statements into machine-verifiable formal languages
Curated Intelligence
AI safety startup Resolution demonstrated the use of agents to automatically formalize complex mathematical statements from statistical learning theory, such as resolution of singularities. This work aims to establish a rigorous science of research automation through the 'industrialization of mathematics,' laying groundwork for understanding the relationship between loss landscape geometry and generalization.
A new study argues that autoformalising natural language mathematical proofs, such as OpenAI's claimed Navier-Stokes blow-up proof, into formal languages like Lean risks semantic disconnection due to unresolvable ambiguities. The translation process is shown to sit at the highest level of the Solvability Complexity Index (SCI), theoretically harder than the Halting problem, implying current AI verification lacks reliability for complex mathematics.