Chain-of-thought prompting

Large language models in 2021 could write fluently and failed at arithmetic word problems a child could do, and the standard explanation was that they did not reason. Jason Wei's paper, posted on 28 January 2022, suggested the failure was in the asking. Instead of showing the model a few problems with answers, show it a few problems with worked solutions, the intermediate steps written in words. The model then wrote steps of its own before answering, and on a benchmark of grade-school maths problems the 540-billion-parameter PaLM went from 18 percent correct to 57. Smaller models did not benefit; the ability appeared only above a certain scale.

The finding was simple enough to be adopted everywhere within months. "Let's think step by step", a five-word prompt that a Tokyo group showed worked on its own, became the most reproduced result of the year. Chain-of-thought was also an early example of an emergent ability, something a large model could do that a smaller one of the same design could not, and it fed the argument about what scale was buying.

It is the ancestor of the reasoning models of 2024, o1 and its successors, which are trained to generate long chains of thought and to search over them, rather than merely prompted to.