Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

"Chain-of-Thought Prompting Elicits Reasoning in Large Language Models" (Jason Wei, Xuezhi Wang, Dale Schuurmans, and colleagues at Google Brain, 2022) is the paper behind the most widely used prompting trick in practice: asking a model to show its work.

What problem it solved

By 2022, the standard way to use a large model on a new task was few-shot prompting: show a handful of input-output examples and let the model continue the pattern. That worked well for classification and lookup-style questions and badly for anything needing several steps, like grade-school word problems or multi-hop logic. Scaling the model up helped on many tasks but barely moved these, which led some to argue that reasoning wasn't something language models would pick up from scale alone.

The key idea

Change what the examples in the prompt contain. Instead of question → answer, each example is question → reasoning steps → answer, with the steps written out in natural language, such as working through the arithmetic of a word problem one line at a time. Given those examples, the model imitates the format on a new question: it produces its own intermediate steps first, then an answer.

This works for the reason covered in Prompt Engineering: a model generates each token conditioned on everything before it, so reasoning steps that are part of the output become context the final answer can use, rather than something that has to happen invisibly inside one forward pass.

Why it mattered

On math word problems, commonsense, and symbolic tasks, the gains were large, and a sufficiently big model with eight worked examples beat earlier approaches that fine-tuned for the task. Two details are worth holding onto. First, the effect appeared mostly at scale: smaller models produced fluent but illogical chains and often did worse than with plain prompting, so reasoning-by-prompt looked like something that emerged with model size. Second, a follow-up (Kojima et al., 2022) found that for large models you often don't need examples at all; appending "let's think step by step" is enough, which is the form most people now use.

The idea also outlived the prompt. Training models to produce long reasoning traces before answering, rather than relying on a prompt to coax them, builds directly on this one, and benchmarks like GSM8K in Evaluation & Benchmarks are the multi-step arithmetic tests this technique was first measured on. It carries a standing caution too: the written chain is what the model produced, not a verified account of how it reached the answer.

Authors: Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, Denny Zhou (Google Brain)

Read the paper — arXiv:2201.11903

Learn more: Prompt Engineering · Prompt Engineering lesson · Evaluation & Benchmarks