Researchers boost AI theorem proving with compiler-guided proof search
New method cuts LLM calls by 21.9% while boosting Lean 4 proof success rates
Researchers Zhuo Liu, Ding Yu, and Hangfeng He have developed a compiler-guided adaptive proof search framework designed to tackle theorem proving in real-world Lean 4 projects. The challenge lies in proofs that depend heavily on project-specific context, where iterative refinement often struggles to balance exploration and exploitation effectively.
Their solution introduces a dual-model generation system that explores diverse starting points through stagnation-triggered resampling, while a compiler-grounded pairwise comparison mechanism exploits promising proof states. In experiments across seven real-world Lean 4 projects from the miniCTX-v2 dataset, this approach achieved a superior effectiveness-efficiency tradeoff compared to traditional pass@k baselines. Within a pass@32 budget, the method improved average pass rates by 12.8 percentage points while reducing LLM calls by 21.9%, demonstrating significant gains in both performance and computational efficiency.
- Framework uses dual-model generation and compiler-guided comparisons to balance exploration and exploitation in Lean 4 theorem proving
- Achieves 12.8 percentage point higher pass rates and 21.9% fewer LLM calls compared to pass@k baselines
- Tested on seven real-world Lean 4 projects from miniCTX-v2 dataset
Why It Matters
This breakthrough accelerates formal verification and automated reasoning for software development and mathematics.