Agentic AI Framework Autoformalizes Research Mathematics

Large Language Models (LLMs) have shown impressive mathematical reasoning abilities, but often produce subtle errors that are hard for humans to catch. Formal mathematical languages like Lean 4 offer mechanical proof checking, making autoformalization—translating natural language math into verifiable code—highly desirable. Recent trends indicate that general-purpose LLMs, optimized for coding, now surpass smaller, specialized Lean-tuned models. Leveraging this shift, researchers developed an agentic autoformalization framework powered by these general coding LLMs. The core of the system is an orchestrator managing a multi-agent pipeline specifically designed for research-level mathematics. A key innovation is its ability to handle concepts not yet present in existing formal libraries like Mathlib. The system dynamically extends necessary type definitions and validates them using a novel "Auxiliary Lemma" technique before formalizing primary theorems. This framework was applied to PutnamBench problems, generating machine-checked Lean proofs, and successfully formalized main theorems and proofs from five ACM STOC papers, with human expert validation. Notably, two papers were proved with no axioms beyond Lean's kernel, demonstrating the system's robustness and potential.