Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

TL;DR

Shows that prompting a large model to write out intermediate reasoning steps markedly improves its accuracy on multi-step problems.

Why it matters

Chain-of-thought reframed prompting as a way to unlock latent reasoning and seeded the whole line of work on test-time reasoning in modern models, while also exposing that stated reasoning is not always faithful.

Key ideas

• Ask the model to show its steps, not just the final answer.
• Reasoning gains appear mainly in sufficiently large models.
• Few-shot exemplars can demonstrate the reasoning style.
• A caution: written steps may not reflect the true computation.

Related concepts