Temperature

A sampling parameter applied to the next-token probability distribution at generation time. A low value sharpens the distribution toward the single most likely token. A high value flattens it and lets less likely tokens win more often. Changing it changes how the model samples, not what it knows.

Why exams ask this

The trap answer treats temperature as a creativity switch or an accuracy switch. It is neither: a model at temperature zero can still produce a confidently wrong answer, because temperature reshapes sampling over a distribution the model already computed. It never touches what the model actually knows.

Relevant to

Related concepts

Resources

No resources linked to this concept yet.