作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• •    

大语言模型在数学推理中的研究进展

  • 发布日期:2024-09-18

Research Progress of Large Language Models in Mathematical Reasoning

  • Published:2024-09-18

摘要: 【目的】全面概述大语言模型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。【文献范围】选取与大语言模型在数学推理领域相关的122篇文献。【方法】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,并提出未来可能的研究方向。【局限】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。【结论】基于思维链提示技术、微调、利用编程语言等外部工具和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技术成为了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,解决数学推理问题方法展开深入研究。

Abstract: 【Objective】This review comprehensively outlines the current state, underlying mechanisms, and trends in applications of large language models in mathematical reasoning capabilities, offering references for future research in this area. 【Scope】This analysis incorporates 122 publications related to mathematical reasoning with large language models. 【Method】The paper systematically describes the types of mathematical reasoning issues and their datasets, delving into the principles, utility, and strategies for enhancing model reasoning capabilities and methods of chain-of-thought prompting. Through qualitative analysis, it provides a thorough overview of research progress in the field of mathematical reasoning with large language models and suggests potential future research directions. 【Limitations】Rapid developments in research related to large models may mean this review does not cover all pertinent studies. 【Conclusion】Methods such as chain-of-thought prompting, fine-tuning, the utilization of programming languages and other external tools, and verification mechanisms can effectively enhance the mathematical reasoning capabilities of large language models, with chain-of-thought prompting techniques in particular becoming a major focus of current research in large language models. Future studies could further enhance the reasoning capabilities of large language models and develop methods for solving mathematical problems.