Modifying a model's learned parameters to change its behavior, used here to construct better sampling proposals.