Wouldn't a model always underperform when you ask it to do an aligned task it wasn't specifically trained on?