Test-time compute trades inference cycles for quality: longer chains of thought, sampling multiple candidates and voting, search over reasoning paths. Reasoning models (o1, DeepSeek-R1) bake this in — they emit extended “thinking” before answering, and accuracy climbs with thinking budget. It complements train-time scaling: when pretraining gains flatten, spending compute per query is the other lever.
Worked example: spending more computation at inference — longer chains of thought, sampling many candidates, search/verification — to get a better answer from the same model, trading latency and cost for quality. Gotcha: it is a real axis (a smaller model that ‘thinks’ longer can beat a bigger one answering instantly on hard reasoning), but returns diminish and cost scales with the extra tokens/samples; the skill is spending compute only where the task is hard enough to need it, not on every request.