大语言模型在数学推理中的研究进展

Research Progress of Large Language Models in Mathematical Reasoning

全面概述大语言模型(LLM)在数学推理中的研究进展、机制原理以及应用趋势,为后续开展相关研究提供参考借鉴。选取与大语言模型在数学推理领域相关的122篇文献。系统描述了数学推理问题的类型及其数据集,分别从增强模型推理能力的策略和思维链提示方法这两方面深入解析各技术的原理、应用价值和存在问题。通过定性分析,提出未来可能的研究方向。大语言模型相关研究发展迅速,相关调研工作可能未覆盖完整。基于思维链提示技术、微调、利用编程语言等外部工具、验证机制等方法可以有效提升大语言模型的数学推理能力,特别是基于思维链提示的方法成为当前大语言模型的主要研究热点。未来研究工作可在进一步提升大语言模型的推理能力、提出解决数学推理问题的新方法等方面展开深入研究。

This review comprehensively outlines the current state,underlying mechanisms,and trends in the applications of Large Language Model(LLM)in terms of mathematical reasoning capabilities.Moreover,it provides references for future research in this area.This review incorporates data from 122 publications related to mathematical reasoning with LLMs and systematically describes different types of mathematical reasoning problems and their datasets.It explores the principles,application values,and issues of various techniques from two perspectives-strategies to enhance model reasoning capabilities and methods of Chain-of-Thought(CoT)prompting.Through qualitative analysis,the review provides a thorough overview of the progress of research in the field of mathematical reasoning with LLMs and suggests potential directions for future research.However,rapid developments in research related to large models may mean that this review does not cover all pertinent studies.Methods such as CoT prompting,fine-tuning,the utilization of programming languages and other external tools,and verification mechanisms can effectively enhance the mathematical reasoning capabilities of LLMs.In particular,CoT prompting techniques are becoming a major focus of current research in LLMs.Future studies could further enhance the reasoning capabilities of LLMs and develop new methods for solving mathematical problems.