Temperature

Temperature is a sampling parameter that reshapes how boldly a model picks its next token. Each logit is divided by the temperature T before softmax:

• Low T sharpens the distribution → focused, deterministic, repetitive.
• High T flattens it → diverse, creative, riskier.

Temperature adds no new knowledge; it only stretches or squashes the confidence the model already has. It’s usually combined with top-p (nucleus) sampling. Try it in Temperature and the art of the next word.