"Relative probabilities assigned to incorrect classes can reveal how a trained model generalizes beyond the correct label."
The distillation paper explains that probabilities assigned to wrong answers contain useful information about how a model generalizes.
View Source