Abstract
Large language models (LLMs) have demonstrated impressive capabilities in natural language understanding, yet their logical reasoning performance remains limited. In this work, we propose LFMC (Learning from Logical Mistake Correction), a fine-tuning framework designed to enhance the logical reasoning abilities of LLMs. We construct a dedicated dataset, LOCD (Logical Error Correction Dataset), which provides both erroneous reasoning paths and corrected logical chains to guide model learning. Experimental results verify that fine-tuning with LOCD consistently enhances logical reasoning performance on large-scale LLMs, with accuracy gains ranging from approximately 2\% to 13\% across different benchmarks. Overall, our study demonstrates that logical error correction learning is a general and effective paradigm for improving LLMs’ reasoning robustness and consistency. These findings provide empirical guidance for designing diverse logical fine-tuning datasets and pave the way for extending LFMC to larger, multilingual, and multimodal logical reasoning scenarios. All corresponding code and data is publicly available at
https://github.com/Zhuxingxing45/LFMC.