An Analysis of Large Language Model Reasoning Path Based on Dynamic Search and Self-Correction

Abstract

Although large language models have been developing rapidly in recent years, there are still shortcomings in reasoning capacity and path in dealing with complex problems. The reasoning capacity of large language models has been studied in the paper based on the dynamic search and self-correction. By reviewing relevant essays, three kinds of methods, namely thought exploration, self-correction, and reinforcement learning and guided search, are summarized in this article. In addition, this paper summarizes the commonly used datasets and evaluation parameters in this field to verify the effect of advancing the reasoning ability of large language models. Finally, the current challenges and optimization methods in this field are provided at the end of the research. It is found that multi-path collaborative search is conducive to improving the comprehensiveness of dynamic exploration, and self-correction can optimize the reasoning path, but there are also failure scenarios in intrinsic self-correction. This study aims to clarify the importance of synergy between the two ways and provide a reference for subsequent related research.