Scaling laws that apply at inference rather than training: more compute per query (longer chains of thought, repeated sampling, verifier-guided search) yields better outputs. Foundation of o1/o3 and DeepSeek-R1 reasoning models. Distinct from training-time scaling.