Curiosity · Concept
Chain-of-Thought Prompting
Chain-of-thought prompting, introduced by Wei et al. (Google) in 2022, elicits intermediate reasoning steps from an LLM instead of a direct answer. Adding 'Let's think step by step' — or providing few-shot examples of reasoning — can turn a failure into a success on arithmetic and logic problems. It's the idea that led directly to today's reasoning models (o1, o3, R1, Claude thinking).
Quick reference
- Proficiency
- Beginner
- Also known as
- CoT, step-by-step prompting, reasoning prompting
- Prerequisites
- Basic prompt engineering
Frequently asked questions
What is chain-of-thought prompting?
A prompting technique that asks the model to produce intermediate reasoning steps before the final answer. The simplest version just appends 'Let's think step by step.' to the prompt; richer versions provide few-shot examples showing the reasoning pattern you want.
Why does chain-of-thought improve accuracy?
By forcing the model to use its output space as a scratchpad, it effectively gets more compute per answer and can decompose multi-step problems instead of guessing the final answer in one shot. On math word problems, CoT can improve accuracy by 20-40 percentage points for large models.
Is CoT the same as a reasoning model?
Reasoning models like o1, o3, and DeepSeek-R1 are trained (often via RL on verifiable rewards) to produce long private chains of thought before answering. So CoT is the underlying capability, but modern reasoning models do it intrinsically and often for much longer than a hand-written CoT prompt would elicit.
Is the chain of thought faithful to the real computation?
Often, but not always. Several studies show that models can produce convincing-looking reasoning that doesn't actually determine the answer — the answer was set by some other feature, and the chain is post-hoc rationalization. This matters for safety and auditing.
Sources
- Wei et al. — Chain-of-Thought Prompting Elicits Reasoning in LLMs — accessed 2026-04-20
- Kojima et al. — Large Language Models are Zero-Shot Reasoners — accessed 2026-04-20
- OpenAI — Learning to Reason with LLMs (o1) — accessed 2026-04-20