Test-Time Compute
Test-time compute is the amount of computation an AI model uses while answering a query, as opposed to during training.
What is Test-Time Compute?
Researchers found that giving models more time to reason at inference improves results, creating a second scaling axis alongside training compute. This shifts costs and energy use towards inference.
How it works
Extra compute is spent generating longer reasoning, sampling several answers or searching and verifying candidate solutions.
← All Artificial Intelligence & Superintelligence definitions · ← Back to the A–Z definitions