A regularization technique that keeps a fine-tuned policy close to its original pretrained distribution.