Inference-time compute (test-time compute) is the strategy of spending more computation when a model answers, longer reasoning chains, multiple sampled attempts
Also called test-time compute, it is the strategy of spending more computation when a model answers, longer reasoning chains, multiple sampled attempts, search over solution paths, to buy higher accuracy on demanding problems.
Because capability is no longer fixed at training time but purchasable per query. The engineering discipline is matching compute budgets to question difficulty, since spending ten times the tokens on trivial queries is pure waste.