# Chain-of-Thought Reasoning Without Prompting

**Authors:** Xuezhi Wang, Denny Zhou  
**arXiv:** [2402.10200](https://arxiv.org/abs/2402.10200)  
**Date:** 2024-02-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.10200v1/

## Key points

- Study explores effectiveness of large language models (LLMs) in reasoning without specific prompting techniques
- Investigates whether LLMs can effectively reason without prompts by altering decoding process
- CoT reasoning paths can be elicited from pre-trained LLMs by exploring alternative tokens in top-k sequences
- Presence of CoT in decoding path correlates with higher confidence in model's decoded answer
- Proposed CoT-decoding method significantly outperforms standard greedy decoding on various reasoning benchmarks
- Study removes need for CoT prompting and focuses on token-level search during decoding while utilizing confidence scores

## Summary

, , , , 

The study "Chain-of-Thought Reasoning Without Prompting" delves into the effectiveness of large language models (LLMs) in reasoning without the need for specific prompting techniques. Previous research has focused on methods such as few-shot or zero-shot chain-of-thought (CoT) prompting, which require manual prompt engineering. However, this study takes a new approach by investigating whether LLMs can effectively reason without prompts. The findings reveal that CoT reasoning paths can be elicited from pre-trained LLMs by simply altering the decoding process. By exploring alternative tokens in the top-k sequences instead of using conventional greedy decoding, the study uncovers that CoT paths are frequently present in these sequences. This not only eliminates the need for prompting but also allows for an assessment of the LLMs' intrinsic reasoning abilities. Furthermore, it is observed that the presence of a CoT in the decoding path correlates with higher confidence in the model's decoded answer. This confidence metric effectively distinguishes between CoT and non-CoT paths. Extensive empirical studies on various reasoning benchmarks demonstrate that the proposed CoT-decoding method significantly outperforms standard greedy decoding. In contrast to recent works that still rely on CoT prompting to improve generation processes, this study completely removes that need and focuses on searching at the token-level during decoding while utilizing confidence scores. Additionally, other recent works explore how chain-of-thought emerges in language models and highlight how pretraining distribution influences model performance in few-shot reasoning scenarios. Techniques such as instruction-tuning or distillation offer alternative ways to elicit reasoning paths from language models without explicit prompting. Overall, this study provides valuable insights into enhancing LLMs' reasoning capabilities without relying on traditional prompting methods and showcases significant improvements through innovative decoding strategies.