Automated theorem provers often rely on sampling language models to generate proof steps. But proof assistants like Lean reject each failed step with a precise error message. According to a new arXiv preprint, pure sampling discards that information, missing an opportunity to correct course.
The preprint introduces a bounded compiler-guided repair approach for Lean 4. Instead of simply resampling after a failure, the method uses the compiler's error messages to guide fixes. The title emphasizes that repair, not resampling, is the key contribution, and the bounded nature suggests a controlled search for corrections.
This matters because proof assistants already provide rich feedback that most sampling-based systems ignore. By leveraging that feedback, the approach could make language-model-driven proof generation more efficient and reliable, though the abstract does not yet report experimental results.