Reasoning / Standard term
Test-time compute
Computation used while producing or selecting an answer, rather than while training the model.
Test-time compute can include a longer internal reasoning process, several sampled candidates, search, tool calls, or a verifier that ranks possible answers. These methods sometimes improve results on tasks with useful intermediate structure or a reliable scoring signal. The effect depends on the model, task, procedure, and budget; additional computation can plateau, amplify a bad approach, or reduce quality.
Builder example
Inference-time computation is a configurable design choice, not a universal quality dial. A workflow may spend more time on difficult code changes or proofs and less on retrieval or classification, but the allocation should come from measured accuracy, latency, and cost on representative cases.
Common confusion: A larger reasoning budget does not mean the model understood the problem better. Bad context, weak tools, misleading verification, and tasks that do not benefit from search can make extra computation unhelpful or harmful.