Compiler-Guided Adaptive Proof Search with Cross-Model Synergy on Context-Dependent Theorem Proving
Source
arxiv.org
Author
Zhuo Liu, Ding Yu, Hangfeng He
Date
Context
An AI proof search can waste attempts by repeatedly revising an unpromising proof or damaging a partially correct one. A Lean 4 study addresses this by generating starting points with two models, trying fresh candidates when progress stalls and using compiler feedback to compare which proof state deserves further refinement.
Across seven projects in miniCTX-v2, the authors report a 12.8-percentage-point increase in average proof success and 21.9% fewer model calls within a comparison budget based on 32 attempts. The result supports using the proof checker's feedback to allocate search effort more selectively. Model-call savings alone do not establish equivalent savings in time or money.
Terms in this piece · Glossary
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.