Coding
Formalize Once, Edit the Rest: Efficient Lean-Based Answer Selection for Math Reasoning
The article presents BASE, a novel pipeline for efficient answer selection in mathematical reasoning using Lean, which formalizes a single base candidate and edits it to generate additional candidates, significantly reducing computational costs. The approach incorporates a rewriter model, LEANSCRIBE, which localizes answers and creates reusable edit functions, achieving a fivefold reduction in autoformalization calls at K=8 across multiple benchmarks and solvers. This method enhances selection accuracy while lowering the formalization burden, making it a valuable advancement for practitioners working with LLMs in mathematical contexts.
mathreasoningllmselection