The insight
When a language model is asked a hard arithmetic or logic question, jumping straight to an answer often fails. Chain of thought prompting asks the model to lay out intermediate reasoning steps before the final answer, which dramatically improves accuracy on multi step problems.
Why it helps
Generating intermediate steps gives the model room to work:
- Each step conditions the next, breaking a hard problem into smaller ones
- The model spends more computation per answer, in effect thinking longer
- Mistakes become visible and are sometimes self corrected along the way
A simple trigger like asking it to reason step by step can unlock this behavior in a capable model, a striking emergent property of scale.
Extensions
Self consistency samples several chains and takes a majority vote, while tree based methods explore branching paths. The cost is more tokens and latency, and the written steps are a plausible narrative, not a guaranteed faithful trace of the internal computation.
Key idea
Chain of thought prompting makes a model produce intermediate steps before answering, improving multi step reasoning at the cost of extra tokens.