(2306.03872) Deductive verification of chain-of-thought reasoning

View a PDF of the article titled Deductive Verification of Chain-of-Thought Reasoning, by Zhan Ling and 5 other authors

View PDF

Abstract:Large Language Models (LLMs) benefit significantly from Chain-of-Thought (CoT) prompts when performing various reasoning tasks. While CoT allows models to produce more elaborate reasoning processes, the emphasis on intermediate reasoning steps may unintentionally introduce hallucinations and accumulated errors, limiting the ability of models to solve complex reasoning tasks. Inspired by the way humans perform careful and precise deductive logical reasoning processes to solve tasks, we aim to enable language models to perform explicit and rigorous deductive reasoning, and also ensure the reliability of their reasoning process through self-verification. However, directly verifying the validity of a complete deductive reasoning process is challenging, even with advanced models such as ChatGPT. In light of this, we propose to decompose a reasoning verification process into a series of stepwise subprocesses, each of which receives only its necessary context and premises. To facilitate this procedure, we propose Natural Program, a natural language-based deductive reasoning format. Our approach enables models to generate precise reasoning steps where subsequent steps are more rigorously grounded in previous steps. It also enables language models to perform self-verification of reasoning in a stepwise manner. By integrating this verification process into each deductive reasoning stage, we significantly improve the accuracy and reliability of generated reasoning steps. In the process, we also improve the correctness of answers to complex reasoning tasks. The code is released at this https URL.