Chain of thought

A model working through intermediate reasoning steps before giving its answer, which improves accuracy on multi-step problems.

Definition

A 2022 Google paper by Wei and colleagues showed that prompting large models to write out their reasoning step by step markedly improved results on arithmetic and logic tasks. Reasoning models now do this by design, often with the reasoning hidden or summarised.

In business use, asking for reasoning helps on complex judgements and makes outputs easier to review. Two cautions apply: the written reasoning is not a guaranteed account of how the model reached its answer, and longer reasoning costs more.

Sources

  1. Wei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (2022)