Loading…
Glossary term
Extra computation a model uses while answering, such as longer reasoning, to improve results.
Test-time, or inference-time, compute is the computation a model spends while answering, as opposed to while training. Letting a model think longer, sample several answers or search over options all use more of it.
Spending more at answer time can raise accuracy on hard questions without making the model bigger.
It opens a second way to improve results besides scaling up training, and it changes the cost of running models.