Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
ReadAsking a large model to write intermediate steps before its answer raises accuracy on reasoning tasks. A short paper that reframes prompting as controlling computation.
My notes
Short, readable, and the idea is almost too simple. The interesting bit is that it only helps above a certain model size — below that, the chains are confidently wrong.
Why I Think It Works
- Each generated token is more compute. Writing steps gives the model more forward passes before it commits to an answer.
- It makes errors inspectable — you can see where the reasoning went wrong, which is gold for debugging.
- Emergent with scale: roughly 100B+ parameters in their experiments.
In agent work I treat this as the "think" half of ReAct. Reasoning models today bake it into training rather than the prompt.