In the no-image test, where the original image is replaced with an unrelated random image, model performance still significantly exceeded random guessing, indicating that models primarily rely on text priors rather than genuinely observing the image.