“Model distillation is highly effective because soft labels (logits) contain significantly more information per sample than one-hot labels due to their higher entropy.”