When a trained model's behavior gradually diverges from its intended target during continued training.