Why it matters

Temperature is core sampling knob. Understanding shapes decoding.

Advertisement

The architecture

Softmax(logits / T).

T = 0: greedy (argmax).

T = 1: unmodified distribution.

Temperature flowLogitsfrom modelDivide by Ttemperature scaleSoftmaxsampling distributionT = 0.7-1.0 typical for chat; T = 0 for deterministic tasks
Temperature.
Advertisement

How it works end to end

T = 0: greedy decoding.

T = 0.7: chat balanced.

T = 1.0: model's calibrated.

T > 1: more random.