Explore model reasoning across images using ground truth annotations.