Lean Refactor: Multi-Objective Controllable Proof Optimization via Agentic Strategy Search

Researchers introduced Lean Refactor, a plug-and-play framework for refactoring Lean proofs using a frozen LLM and a curated strategy database.

The system optimizes for shorter proofs, lower compilation cost, and better robustness across Lean and Mathlib versions.

In experiments, Lean Refactor delivered strong compression gains, faster compilation, and improved transfer across versions.

The approach tackles a practical bottleneck in formal verification without requiring repeated retraining as models and Lean evolve.