What is temperature scaling in neural networks?

Temperature scaling adjusts logits to control prediction confidence.

Logits are divided by temperature T before softmax. Higher T produces softer probabilities; lower T makes distributions sharper.