Advanced Concepts
Test-time compute
Extra computation used during inference to improve the quality of a model's answer.
Example
The model uses extra test-time compute to evaluate several possible answers before responding.
Why it matters
It offers a way to improve an answer without first retraining the model.
You'll hear it
“Would giving it more time actually help with this question?”
Can you control it?
Not directly
No, not directly. Test-time compute describes a learned property, observed result, or broader concept rather than a normal user setting. You may influence it indirectly by changing the tool, data, prompt, or workflow.
Common questions
- Does extra processing change what the model learned during training?
- Generally no. It changes the effort spent on the current task, rather than retraining the model.
- Can more effort make an answer worse?
- It can. Extra attempts may follow a mistaken path, so improvement needs measurement rather than assumption.
- Is this useful for every request?
- No. A straightforward request may not benefit enough to justify additional processing.