Test-time compute is the idea that models can keep improving by spending more compute at inference, not only at training. OpenAI o1 is the release that crystallized the approach: the model effectively thinks longer per query, generating more tokens to reach a better answer. It is the technical foundation of the Reasoning Models category and adds a new axis to the Scaling Laws debate. Long Thinking and Extended Thinking are productized expressions of the same concept.
MEVZU N°124ISTANBULYEAR I — VOL. III
Glossary · Intermediate · 2024
Test-time Compute
An approach where a model spends more compute at inference time, not just during training, to lift performance.
- EN — English term
- Test-time Compute
- TR — Turkish term
- Test-Anı Hesaplama