← All papers
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Wei, Wang, Schuurmans et al. · 2022 · NeurIPS 2022
PromptingRead on arXiv
Showed that prompting models with intermediate reasoning steps dramatically improves performance on reasoning-heavy tasks. Chain-of-thought is now a fundamental technique in prompt engineering.
Key Idea
Simply adding intermediate reasoning steps to few-shot examples causes LLMs to generate their own step-by-step reasoning, leading to dramatically better performance on math, logic, and commonsense reasoning tasks.
How It Works
- Standard prompting: Q: "Roger has 5 tennis balls..." → A: "11"
- Chain-of-thought prompting: Q: "Roger has 5 tennis balls..." → A: "Roger started with 5 balls. He bought 2 cans of 3 balls each, so 2 × 3 = 6 new balls. 5 + 6 = 11. The answer is 11."
- The model learns to decompose problems by following the demonstrated reasoning pattern
Key Findings
- CoT is an emergent ability: it only helps models above ~100B parameters. Smaller models produce incoherent reasoning chains that hurt performance
- Improvements are largest on complex reasoning tasks (arithmetic, commonsense, symbolic)
- Zero-shot CoT also works: simply adding "Let's think step by step" to the prompt triggers reasoning chains (Kojima et al., 2022)
Why It Matters
- Unlocked a class of tasks that LLMs previously couldn't solve (multi-step math, logic puzzles)
- Established prompting as a first-class research area alongside model architecture and training
- Led to Tree-of-Thought, Graph-of-Thought, and other structured reasoning extensions
- Now standard practice in all serious LLM applications
Key Takeaways for Interviews
- CoT works by eliciting step-by-step reasoning through demonstration
- It's emergent, only effective at large model scales
- Zero-shot CoT ("Let's think step by step") is surprisingly effective
- CoT improves accuracy but increases token usage (cost/latency tradeoff)