The stark performance gap between current models and humans—especially the fact that even the top proprietary model only hits 40%—highlights how underexplored a

The stark performance gap between current models and humans—especially the fact that even the top proprietary model only hits 40%—highlights how underexplored and underdeveloped multi-image spatial reasoning still is.