@

초록

We introduce Reverse-Mutual Reasoning for Mathematical Thinking (ReMi-ReMath), a mutual reasoning framework that enhances mathematical reasoning in small language models without requiring fine-tuning. Building on the mutual reasoning structure of rStar, ReMi-ReMath expands it by incorporating backward reasoning-based discriminative reconstruction and a self-evaluation system that measures logical and semantic consistency, rather than simple answer matching. Across benchmarks, ReMi-ReMath improves performance by 3-5%p over rStar using 3.8B-scale models, demonstrating that combining mutual reasoning and reverse reconstruction establishes a new paradigm for quantitatively verifying reasoning validity beyond answer-centric evaluations.

키워드