Temperature Sharpening is the specific application of temperature scaling to sharpen (reduce entropy of) prediction distributions — a key component in semi-supervised learning and knowledge distillation, where the temperature parameter $T$ controls the softness or hardness of the output distribution.

Temperature Effects

• $T ightarrow 0$: Distribution becomes one-hot (hard label). Maximum confidence.
• $T = 1$: Standard softmax. No modification.
• $T > 1$: Distribution becomes softer/more uniform. Used in knowledge distillation.
• $T < 1$: Distribution becomes sharper. Used in semi-supervised learning for pseudo-labels.

Why It Matters

• Two Use Cases: $T > 1$ for distillation (soft targets), $T < 1$ for semi-supervised (sharpen pseudo-labels).
• Confidence Control: Provides a continuous knob between soft uncertainty ($T$ high) and hard commitment ($T$ low).
• Universal: Used in MixMatch, FixMatch, knowledge distillation, and contrastive learning (InfoNCE temperature).

Temperature Sharpening is the confidence knob — a single parameter that controls how decisive or uncertain the model's predictions appear.