AI Researchers Propose Scaling Up Automated Theorem Proving

Researchers advocate for a move beyond individual statement autoformalization to theory-level autoformalization, formalizing entire mathematical theories with all interdependencies. Currently, autoformalization primarily focuses on translating single natural language statements into machine-verifiable languages. The authors contend that genuine formalization demands constructing complete theories encompassing axioms, definitions, and lemmas.

They suggest organizing these formalized theories into structured libraries. This approach differs from existing methods and addresses the need for comprehensive, interconnected knowledge bases in formal verification.

The researchers identify open challenges in achieving theory-level autoformalization and propose three potential avenues for future work. They also maintain a survey of autoformalization resources available online.