EquivPruner: Boosting Efficiency and Quality in LLM-Based Search via Action Pruning

EquivPruner is a novel approach designed to enhance the efficiency and quality of LLM-based search by identifying and pruning semantically equivalent actions during reasoning processes. It introduces the MathEquiv dataset for training a lightweight equivalence detector, demonstrating significant improvements in token consumption and reasoning accuracy—specifically, a 48.1% reduction in token use while enhancing accuracy on the Qwen2.5-Math-7B-Instruct model tested on the GSM8K benchmark. This advancement is crucial for practitioners aiming to optimize LLM performance in domain-specific contexts, particularly in mathematical reasoning.