LLMs
Temperature
Temperature rescales logits before sampling to control distribution sharpness.
How it works
Lower temperature concentrates probability on high-logit tokens, while higher temperature spreads probability more broadly. A value of zero is commonly implemented as greedy selection rather than literal division.
A concrete example
Illustrative: logits [2, 1] yield probabilities about [0.73, 0.27] at temperature 1 and [0.88, 0.12] at 0.5.
Do not confuse it with Top-p
temperature reshapes all logits; top-p truncates to a cumulative probability mass.