# 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/2205.11916v4/

## Key points

- Large language models (LLMs) are explored for their zero-shot reasoning capabilities
- The authors focus on the chain of thought (CoT) prompting technique for complex multi-step reasoning
- CoT prompting has achieved state-of-the-art performances in arithmetics and symbolic reasoning tasks
- LLMs can be decent zero-shot reasoners by adding the phrase "Let's think step by step" before each answer
- Zero-shot-CoT significantly outperforms zero-shot LLM performances on various benchmark reasoning tasks
- Improvements are achieved without handcrafted few-shot examples
- Using different large models, such as InstructGPT and PaLM, similar magnitudes of improvements are observed
- The single prompt used suggests untapped and understudied fundamental zero-shot capabilities of LLMs
- High-level cognitive capabilities can be extracted from LLMs through simple prompting without crafting finetuning datasets or few-shot exemplars

## Summary

In their paper titled "Large Language Models are Zero-Shot Reasoners," authors Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa explore the zero-shot reasoning capabilities of pretrained large language models (LLMs). LLMs are widely used in various sub-fields of natural language processing (NLP) and are known for their excellent few-shot learning abilities with task-specific exemplars. The authors specifically focus on the chain of thought (CoT) prompting technique, which elicits complex multi-step reasoning through step-by-step answer examples. CoT prompting has achieved state-of-the-art performances in arithmetics and symbolic reasoning tasks that do not follow the standard scaling laws for LLMs. While these successes are often attributed to LLMs' few-shot learning ability, the authors demonstrate that LLMs can also be decent zero-shot reasoners by simply adding the phrase "Let's think step by step" before each answer. To validate their approach, the authors introduce Zero-shot-CoT, which utilizes a single prompt template across diverse benchmark reasoning tasks. Experimental results show that Zero-shot-CoT significantly outperforms zero-shot LLM performances on tasks such as arithmetics (MultiArith, GSM8K, AQUA–RAT, SVAMP), symbolic reasoning (Last Letter, Coin Flip), and other logical reasoning tasks (Date Understanding, Tracking Shuffled Objects). These improvements are achieved without any handcrafted few–shot examples. For example, using the large InstructGPT model (text–davinci–002), Zero–shot–CoT increases accuracy on MultiArith from 17.7% to 78.7% and on GSM8K from 10.4% to 40.7%. Similar magnitudes of improvements are observed with another off–the–shelf large model called PaLM with 540B parameters. The authors argue that the versatility of this single prompt across diverse reasoning tasks suggests untapped and understudied fundamental zero–shot capabilities of LLMs. They propose that high–level cognitive capabilities can be extracted from LLMs through simple prompting without crafting finetuning datasets or few–shot exemplars. The paper has been accepted to NeurIPS2022 and its code is available on GitHub for further reference.