Changes to how a model's internal activations encode information, which can degrade performance on tasks not explicitly trained on.