Spending more computation when the model answers — thinking longer, trying multiple attempts — to buy accuracy without training a bigger model.
For years the way to a smarter model was bigger training runs. Test-time compute is the other axis: at answer time, let the model reason longer, sample many candidate solutions and keep the best, or check its own work. Accuracy scales with how much you spend per question.
This reframes the economics — capability becomes a per-request dial rather than a fixed property of the model. A hard bug might be worth 10 minutes of thinking; a routine rename is not. Much current research maps where this scaling holds and where it flattens.