From Sorry to QED: Refining LLM Math Proofs via Angelic Execution
Summary
- Researchers introduced MathEye, a framework that converts imperfect LLM-generated Lean proof scripts into fully verified mathematical proofs.
- MathEye uses angelic proof checking to replace failing proof fragments with placeholders, isolating exact unresolved obligations.
- By performing residual analysis on a proof dependence graph and formulating obligation selection as a MaxSAT problem, MathEye minimizes human interaction.
- Experiments on 30 real-world analysis and algebra benchmarks showed MathEye solved 93.3% of problems while reducing user interactions by 55%.
Why it matters
MathEye improves the reliability of LLMs in formal mathematics by bridging the gap between flawed automated proof scripts and verified proofs through targeted human-AI collaboration.