COLM 2025: Hidden in plain sight: VLMs overlook their visual representations