Looks like somethings that could always be worth a try, but as they show; it works well with some models in some applications and in other cases it doesn’t. Maybe it is actually a nudge of a model to something it hasn’t seen during initial training.