Not really. Rather, the model is still overconfident in what it has learned, the question is if it is trained only to do counting without relying on knowledge, can it do this?