How LLMs Are Made / Glossary

Chain of thought

Chain of thought is the step-by-step working a language model writes out before its final answer. Reasoning models are trained to produce long chains of thought to solve harder problems.

Researchers showed in 2022 that simply prompting a large model to write out its working step by step improved its accuracy on maths word problems. Writing the steps gives the model more text to build on before it commits to an answer.

Reasoning models take this further with training. They practise on problems with checkable answers, such as maths and code, and reinforcement learning rewards the chains that reach the right result. Habits like checking their own work get reinforced, and thinking longer often helps.

A chain of thought is useful but not a perfect window into the model. Researchers have found that the written steps don't always reflect what really drove the answer, so the final result still needs checking.