# Zero-Shot Verification-guided Chain of Thoughts

**Authors:** Jishnu Ray Chowdhury, Cornelia Caragea  
**arXiv:** [2501.13122](https://arxiv.org/abs/2501.13122)  
**Date:** 2025-01-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2501.13122v1/

## Key points

- Jishnu Ray Chowdhury and Cornelia Caragea from the University of Illinois at Chicago explore the use of Large Language Models (LLMs) in guiding reasoning processes through Chain-of-Thought (COT) prompts and verifiers.
- The study focuses on LLM-based self-verification of self-generated reasoning steps in a zero-shot regime, unlike previous studies that relied on fine-tuned verifiers or manually crafted few-shot examples.
- Researchers introduce a new zero-shot prompt called COT STEP to facilitate the decomposition of reasoning steps and design two new zero-shot prompts for LLM-based verifiers.
- Evaluation centers around assessing the accuracy of these verifiers in classifying the correctness of reasoning chains and exploring different ways to utilize verifier scores in guiding mathematical and commonsense tasks with various LLMs.
- The introduction highlights how LLMs have revolutionized NLP by achieving state-of-the-art performance through intelligent prompting strategies.
- Authors emphasize the importance of COT prompting for multi-step reasoning tasks, paving the way for advancements in efficiently guiding complex chains of thoughts.
- Research has potential implications for enhancing real-time reasoning processes and improving overall performance in diverse cognitive tasks.

## Summary

In their paper titled "Zero-Shot Verification-guided Chain of Thoughts," Jishnu Ray Chowdhury and Cornelia Caragea from the University of Illinois at Chicago explore the use of Large Language Models (LLMs) in guiding reasoning processes through Chain-of-Thought (COT) prompts and verifiers. Previous studies have demonstrated the effectiveness of these tools, but they often rely on fine-tuned verifiers or manually crafted few-shot examples. In contrast, this study focuses on LLM-based self-verification of self-generated reasoning steps in a zero-shot regime. To investigate this novel approach, the researchers introduce a new zero-shot prompt called COT STEP to facilitate the decomposition of reasoning steps. They also design two new zero-shot prompts for LLM-based verifiers. The evaluation centers around assessing the accuracy of these verifiers in classifying the correctness of reasoning chains and exploring different ways to utilize verifier scores in guiding mathematical and commonsense tasks with various LLMs. The introduction highlights how LLMs have revolutionized NLP by achieving state-of-the-art performance through intelligent prompting strategies. The authors emphasize the importance of COT prompting for multi-step reasoning tasks, paving the way for advancements in efficiently guiding complex chains of thoughts. Furthermore, they discuss how their research has potential implications for enhancing real-time reasoning processes and improving overall performance in diverse cognitive tasks.