@AnnemarieBridy the way i understood the paper, it wouldn’t change much, but there’s a lot of variables, like the increased data efficiency also means there’s less training data to reference, but theoretically without increasing overfitting (quoting the source)