Inference-time scaling
Beginner
Getting better answers by letting the AI work longer on each question, without changing the AI itself.
Novice
Spending more computation when the model answers (longer chains, more samples, a checker that picks the best) rather than training a bigger model. Also called test-time scaling or test-time compute.
Expert
Trading inference FLOPs for accuracy: parallel sampling with aggregation (self-consistency, best-of-N), sequential revision, search against a process reward model, or longer chains. The best allocation depends on question difficulty.
See also: Self-consistency, Best-of-N, Budget forcing.