28 January 2022

Chain-of-thought prompting improves step-by-step reasoning in large language models

Jason Wei and colleagues showed that prompting a large language model with worked examples of intermediate reasoning steps dramatically improved its performance on arithmetic, commonsense, and symbolic tasks, reaching state-of-the-art accuracy on a maths benchmark.

Reasoning and planning Language models Natural language processing Transformers and attention Foundational method Demonstrated

Background

Large language models had become remarkably capable at predicting text, but getting them to solve problems that require several steps of logic was harder. A model might know a great deal of maths but still fail at a word problem that requires keeping track of intermediate quantities. The gap was puzzling, because adding more parameters did not obviously close it.

One natural hope was fine-tuning: training the model further on carefully labelled examples. That can help, but it is expensive and requires large datasets for each new task.

What happened

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou published a paper on 28 January 2022 showing that the key was to include worked reasoning in the prompt itself. The method, which they called chain-of-thought prompting, works by providing a few examples where the answer is accompanied by a series of intermediate steps written out in plain language. The model then produces its own chain of reasoning before arriving at an answer.

The technique requires no changes to the model. A handful of examples in the prompt is enough. The paper reports experiments on three large language models across arithmetic, commonsense, and symbolic reasoning tasks, with gains that were large on arithmetic benchmarks and modest or absent on smaller models and some other task types. Prompting a 540-billion-parameter model with just eight chain-of-thought examples achieved state-of-the-art accuracy on GSM8K, a benchmark of maths word problems, surpassing even fine-tuned GPT-3 with a verifier.

The paper shows that chain-of-thought prompting improves performance in sufficiently large language models, and that standard prompting failed to elicit the multi-step reasoning behaviour that chain-of-thought prompts could produce in those models. The method gave practitioners a practical way to draw it out.

Why it mattered

Before this paper, simply scaling language models had not reliably produced complex multi-step reasoning. The work showed that benchmark performance on reasoning tasks improves markedly in sufficiently large models when chain-of-thought prompts are used, without any fine-tuning, and that a 540-billion-parameter model with just eight examples in its prompt could surpass fine-tuned GPT-3 with a verifier on maths word problems.

Sources

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arxiv.org. Primary source
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arxiv.org. Primary source

Cite this page

AI Achievements. (2022). Chain-of-thought prompting improves step-by-step reasoning in large language models. Retrieved 2026-08-29, from https://achievements.ai/milestone/chain-of-thought-prompting-improves-step-by-step-reasoning

@misc{achievements_chain_of_thought_prompting_improves_step_by_step_reasoning,
  title  = {Chain-of-thought prompting improves step-by-step reasoning in large language models},
  author = {{AI Achievements}},
  year   = {2022},
  url    = {https://achievements.ai/milestone/chain-of-thought-prompting-improves-step-by-step-reasoning}
}