Artificial Intelligence

Test-Time Compute

Allocating additional computation during inference (not training) to improve output quality. Techniques include chain-of-thought, self-consistency, and iterative refinement.

Why It Matters

Test-time compute offers a way to improve AI output quality without retraining. You can make any model better by giving it more time to 'think.'

Example

Running chain-of-thought reasoning plus self-consistency (10 samples, majority vote) at inference time, improving math accuracy from 70% to 90% at the cost of 10x more compute.

Think of it like...

Like giving a student more exam time — the same knowledge, but more time to think, check work, and reconsider, leading to better answers.

Related Terms