A penalty term that measures how much a model's output distribution diverges from a reference distribution, used to preserve quality during training.