A model that reads images and text together may spot every object but lose track of which ones it has already considered. This reference gap is the failure to keep words connected to the right objects or places while reasoning through a task.
The researchers compared identical images and questions with different visual aids, including numbered objects, shading, and grid lines. They tested tasks such as counting shapes and comparing sizes, changing the aid while keeping the scene and correct answer fixed.
The right aid improved accuracy by up to 25%, while poorly matched marks often did little or made answers worse. A system that chooses a suitable aid without extra model training could make image reasoning more reliable.