Inference-time compute

Spending more computation at answer time (longer reasoning, multiple samples, search) to improve quality — the scaling axis that reasoning models exploit.