A Survey on Verifying Reasoning Chains Generated by Large Language Models

Large Languages Models (LLMs) are currently being extensively employed for many Natural Language Processing tasks such as question answering, natural language inference, document summarization etc. Chain-of-Thought (CoT) prompting guides LLMs with the reasoning steps, compelling them to generate reasoning chains. While some of the reasoning chains may follow a correct thought process, they can also suffer from hallucinations, leading to errors in answer generation. Recently, many articles have targeted the problem of verifying these reasoning chains from various aspects. Despite this recent attention, to the best of our knowledge, no comprehensive survey currently summarizes these studies on CoT verification. This work addresses that gap by presenting a detailed overview of the methods for verifying reasoning chains and categorizing them according to their methodology. This paper introduces a novel taxonomy of classification of the methods introduced so far and mainly divides them into approaches that assess entire chains versus those that examine individual steps. This paper also reviews benchmarks for evaluating CoT reasoning and verification methods and further discusses the challenges and future directions associated with these methods. By compiling and analyzing these approaches, our survey aims to advance the understanding and development of robust reasoning techniques in LLMs.