Try this in your head: "Tom has 5 apples and buys 3 bags with 4 apples each. How many does he have?" Most people do not leap to the answer. They say "three bags of four is twelve, and five plus twelve is seventeen." Writing out the steps is how humans solve multi-step problems.
For years, large language models were asked to answer such problems directly, and they often failed. This 2022 paper from Google researchers asked a simple question: what if the prompt shows the model how to reason, by including worked examples with the steps written out? The answer changed how people talk to language models.
In chain-of-thought prompting, each example in the prompt includes the intermediate reasoning steps before the answer. Large models then write their own steps for a new question, and their accuracy on math, commonsense and symbolic reasoning tasks improves, sometimes dramatically.
What is a chain of thought?
The paper defines a chain of thought as a series of intermediate natural language reasoning steps that lead to the final output. In standard few-shot prompting, you show the model a few question and answer pairs. In chain-of-thought prompting, each answer in your examples also contains the reasoning that gets there.
Nothing is retrained. The only thing that changes is the text in the prompt, which makes the method cheap to try with any model you can already call.
It is like the difference between a teacher who shows only the answer key and a teacher who shows worked solutions. Students given worked solutions learn the method, and then write out their own working.
