# Large Language Models are Zero-Shot Reasoners

**Authors:** Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, Yusuke Iwasawa  
**arXiv:** [2205.11916](https://arxiv.org/abs/2205.11916)  
**Date:** 2022-05-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2205.11916v1/

## Key points

- Pretrained large language models (LLMs) have excellent few-shot learning abilities
- Chain of thought (CoT) prompting elicits complex multi-step reasoning
- CoT prompting achieves state-of-the-art performances in difficult system-2 tasks
- LLMs can perform well as zero-shot reasoners by adding the phrase "Let's think step by step" before each answer
- Zero-shot-CoT outperforms zero-shot LLM performances on diverse benchmark reasoning tasks
- Zero-shot CoT increases accuracy on benchmark tasks significantly
- The single prompt used in Zero-shot CoT is versatile across various reasoning tasks
- LLMs have untapped and understudied fundamental zero shot capabilities
- Simple prompting techniques can extract high level cognitive capabilities from LLMs
- This paper serves as a strong zero shot baseline for challenging reasoning benchmarks

## Summary

The paper titled "Large Language Models are Zero-Shot Reasoners" discusses the capabilities of pretrained large language models (LLMs) in natural language processing (NLP). LLMs are known for their excellent few-shot learning abilities with task-specific exemplars. The authors highlight the success of a recent technique called chain of thought (CoT) prompting, which elicits complex multi-step reasoning through step-by-step answer examples. CoT prompting has achieved state-of-the-art performances in difficult system-2 tasks such as arithmetics and symbolic reasoning, which do not follow the standard scaling laws for LLMs. While these successes are often attributed to LLMs' ability for few-shot learning, the authors demonstrate that LLMs can also perform well as zero-shot reasoners by simply adding the phrase "Let's think step by step" before each answer. They introduce a new method called Zero-shot-CoT, which significantly outperforms zero-shot LLM performances on diverse benchmark reasoning tasks including arithmetics, symbolic reasoning, and other logical reasoning tasks. Notably, this improvement is achieved without any handcrafted few-shot examples. Experimental results show that Zero-shot CoT increases accuracy on benchmark tasks such as MultiArith from 17.7% to 78.7% and GSM8K from 10.4% to 40.7% using an off the shelf 175B parameter model. The authors emphasize the versatility of this single prompt across various reasoning tasks, suggesting untapped and understudied fundamental zero shot capabilities of LLMs. The findings of this study suggest that high level multi task broad cognitive capabilities can be extracted from LLMs through simple prompting techniques. The authors hope that their work serves as a strong zero shot baseline for challenging reasoning benchmarks and highlights the importance of exploring and analyzing the vast amount of hidden zero shot knowledge inside LLMs before crafting finetuning datasets or few shot exemplars. Overall, this paper sheds light on the zero shot reasoning abilities of LLMs and emphasizes the potential for extracting high level cognitive capabilities through simple prompting techniques.