test-time compute

Test-time compute is the idea that a model can get better answers simply by being allowed to work harder when you ask, rather than by being bigger or retrained. Give it a larger thinking budget, more room to write out reasoning, try several approaches, and check its work, and accuracy on hard problems often rises. It is the machine version of taking an exam with five minutes versus an hour: the same student, more time, a better paper.

This matters because it opens a second dial alongside model size. For years the main way to improve a model was to scale training, parameters and data. Test-time compute lets you trade extra inference effort for quality on demand, problem by problem: spend little on easy questions, spend a lot on the rare hard one. The cost is real, since more thinking means more tokens, more money, and more waiting, and the returns eventually flatten, so the practical art is matching the budget to how much the problem deserves.