# Faithful Logical Reasoning via Symbolic Chain-of-Thought

**Authors:** Jundong Xu, Hao Fei, Liangming Pan, Qian Liu, Mong-Li Lee, Wynne Hsu  
**arXiv:** [2405.18357](https://arxiv.org/abs/2405.18357)  
**Date:** 2024-05-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2405.18357v1/

## Key points

- The paper introduces the concept of Symbolic Chain-of-Thought (SymbCoT) as a framework to enhance logical reasoning in large language models (LLMs).
- SymbCoT integrates symbolic expressions and logic rules with CoT prompting to strengthen LLMs' logical reasoning capability.
- The framework translates natural language context into symbolic format within an LLM, derives step-by-step plans using symbolic logical rules, and includes a verifier for checking translation and reasoning chains.
- Evaluation on five standard datasets shows that SymbCoT outperforms the traditional CoT method consistently in First-Order Logic and Constraint Optimization tasks.
- SymbCoT is praised for being more faithful, flexible, and explainable in logical reasoning processes compared to existing methods.
- This research marks the first attempt to combine symbolic expressions and rules into CoT for logical reasoning with LLMs.
- The paper has been accepted by ACL 2024 for inclusion in its main proceedings, and the code for SymbCoT is available on GitHub.

## Summary

The paper "Faithful Logical Reasoning via Symbolic Chain-of-Thought" by Jundong Xu, Hao Fei, Liangming Pan, Qian Liu, Mong-Li Lee, and Wynne Hsu addresses the limitations of the Chain-of-Thought (CoT) technique in enhancing the reasoning ability of large language models (LLMs) when it comes to handling logical reasoning that heavily relies on symbolic expressions and rigid deducing rules. To overcome this challenge, the authors propose a novel framework called Symbolic Chain-of-Thought (SymbCoT), which integrates symbolic expressions and logic rules with CoT prompting to strengthen the logical reasoning capability of LLMs. The SymbCoT framework operates by first translating natural language context into a symbolic format within an LLM. Subsequently, it derives a step-by-step plan to solve problems using symbolic logical rules and includes a verifier to check the translation and reasoning chain. Through comprehensive evaluations on five standard datasets encompassing both First-Order Logic and Constraint Optimization symbolic expressions, SymbCoT demonstrates significant improvements over the CoT method consistently. This refreshes current state-of-the-art performances in logical reasoning tasks. Furthermore, the authors showcase that their system advances in terms of being more faithful, flexible, and explainable in logical reasoning processes. Notably, this work represents the first attempt to combine symbolic expressions and rules into CoT for logical reasoning with LLMs. The research has been accepted by ACL 2024 for inclusion in its main proceedings. Interested individuals can access the code for SymbCoT on GitHub at https://github.com/Aiden0526/SymbCoT.