Knowledge MaxxingTest-time compute
Test-time compute is additional computation a model spends while answering a specific question, such as generating and comparing multiple reasoning attempts, rather than during training.
Prerequisite: Model serving
What matters
Spending more compute at answer time can improve accuracy on hard problems, trading response speed and cost for a better chance of a correct answer, separate from any change to the trained model itself.
Example
A model might generate five different step-by-step solutions to a math problem and pick the most consistent answer among them, spending extra time only at that moment.
Common misconception
Test-time compute is not the same as making the model itself bigger or smarter; it only changes how much work is done per question, not the underlying model.
Recall question
What is being traded off when a system increases test-time compute?