Is it really surprising that text model can't solve graphical quizzles?