Asking a model to think step by step改 changed what it could solve. A 2022 finding showed that prompting large models to spell out intermediate reasoning sharply improved performance on math and logic. The gain emerged only at sufficient scale.
Show the steps
The prompt matters. Encouraging explicit intermediate reasoning boosts accuracy on multi-step problems. The model externalizes its work.
Emergent with scale
Size is a threshold. The benefit appears only in sufficiently large models. Smaller ones gain little.
Broad tasks
Reach is wide. Arithmetic, commonsense, and symbolic tasks all improve. The effect generalizes.
Self-consistency
Sampling helps. Aggregating several reasoning paths and voting raises reliability. Diversity reduces error.
Limits remain
It is not magic. Models still make confident logical errors. Reasoning is improved, not solved.
A research wave
Follow-ups proliferated. Tree- and graph-structured reasoning extended the idea. The area expanded fast.
The bottom line
Chain-of-thought prompting elicits step-by-step reasoning that sharply improves multi-step tasks. The gain is emergent with scale. It launched a wave of reasoning research.
