# Enhancing Zero-Shot Chain-of-Thought Reasoning in Large Language Models through Logic

**Authors:** Xufeng Zhao, Mengdi Li, Wenhao Lu, Cornelius Weber, Jae Hee Lee, Kun Chu, Stefan Wermter  
**arXiv:** [2309.13339](https://arxiv.org/abs/2309.13339)  
**Date:** 2023-09-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.13339v1/

## Key points

- Recent advancements in large language models have demonstrated impressive ability to generalize across domains
- Large language models struggle with multi-step reasoning tasks, indicating a need for improvement in their reasoning abilities
- Large language models often fail to effectively utilize extensive knowledge to establish coherent thinking patterns, particularly in reasoning
- Generative language models sometimes exhibit hallucinations due to unconstrained reasoning procedures lacking adherence to logical principles
- The authors propose a neurosymbolic framework called LogiCoT to enhance the zero-shot chain-of-thought reasoning ability of large language models
- LogiCoT leverages principles from symbolic logic to verify and revise reasoning processes accordingly
- Incorporating logical principles improves coherence and accuracy of output in various language tasks such as arithmetic, commonsense understanding, symbolic reasoning, causal inference, and social problems
- Experimental assessments validate the effectiveness of LogiCoT's enhanced reasoning paradigm in improving overall reasoning capabilities of large language models.

## Summary

Recent advancements in large language models have demonstrated their impressive ability to generalize across various domains. However, these models still struggle with multi-step reasoning tasks, indicating a need for improvement in their reasoning abilities. Despite possessing extensive knowledge, large language models often fail to effectively utilize this knowledge to establish coherent thinking patterns, particularly when it comes to reasoning. In fact, generative language models sometimes exhibit hallucinations due to their unconstrained reasoning procedures that lack adherence to logical principles. To address this limitation and enhance the zero-shot chain-of-thought reasoning ability of large language models, the authors propose a neurosymbolic framework called Logical Chain-of-Thought (LogiCoT). This framework leverages principles from symbolic logic to verify and revise the reasoning processes accordingly. By incorporating logical principles into the model's reasoning procedures, LogiCoT aims to improve the coherence and accuracy of its output. The efficacy of LogiCoT's enhanced reasoning paradigm is evaluated through experimental assessments on various language tasks spanning different domains such as arithmetic, commonsense understanding, symbolic reasoning, causal inference and social problems. The results demonstrate that incorporating logic into the model's reasoning process significantly improves its performance in these tasks. In conclusion, this study introduces LogiCoT as a neurosymbolic framework that enhances the zero-shot chain-of-thought reasoning ability of large language models by leveraging principles from symbolic logic. The experimental evaluations conducted on diverse language tasks validate the effectiveness of this approach in improving the model's overall reasoning capabilities.