通过执行驱动的推理增强提升大型语言模型数学问题解决能力

Large Language Models (LLMs) still face significant challenges in mathematical problem-solving, particularly concerning complex reasoning and numerical precision. This paper introduces a novel paradigm called Execution-Driven Reasoning Augmentation (EDRA), designed to address these limitations by tightly integrating model-generated reasoning steps with an external executor, such as a Python interpreter. The core idea behind EDRA is that as the LLM generates each reasoning step, its correctness is immediately verified by the executor. The execution results are then fed back to the model, creating a closed-loop self-correction process. This approach not only effectively identifies and rectifies errors within the reasoning chain but also significantly enhances the accuracy of numerical computations. Experimental results demonstrate that EDRA achieves substantial performance improvements across various mathematical benchmarks, including GSM8K, MATH, and MMLU-Math, outperforming existing state-of-the-art methods. A key advantage of EDRA lies in its versatility, as it can be combined with diverse LLM architectures and inference strategies. This provides an effective pathway to elevate LLM performance in tasks demanding precise calculations and logical reasoning. The research underscores the critical importance of external tools in augmenting LLM capabilities and opens new avenues for future research in mathematical reasoning for LLMs.