A regularization technique that limits how far a model's distribution can drift from a reference distribution during training.