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

Summary

Shows that prompting large models with worked examples that include intermediate reasoning steps greatly improves their performance on arithmetic, commonsense, and symbolic reasoning tasks.

Key points

• The effect appears only at sufficient model scale.
• A few step-by-step examples in the prompt are enough.
• Gains are largest on multi-step problems.

Why it matters

The starting point for "thinking" models, which now produce such reasoning internally and are trained to use it.