WTF is Chain-of-thought?
Getting a model to write out intermediate reasoning steps before giving its final answer.
Also called: cot, chain of thought
TL;DR
Asking a model to "think step by step" makes it noticeably better at maths, logic and multi-step problems. Today's reasoning models do this internally by default, but the idea started as a simple prompting trick.
How it works
The original paper showed few-shot prompts containing worked reasoning examples improved performance on arithmetic and commonsense benchmarks, with gains emerging at larger model scales. Later work found zero-shot cues ("Let's think step by step") also help. Modern reasoning models are trained with reinforcement learning to produce long internal reasoning before answering, often called extended or test-time thinking.
Origin
- Coined / introduced by
- Wei et al. (Google Research)
- When
- 28 Jan 2022
- Story
- Introduced in 'Chain-of-Thought Prompting Elicits Reasoning in Large Language Models'.