The Cartesian Shortcut: Re-evaluate Vision Reasoning in Polar Coordinate Space
This research discovered a 'Cartesian Shortcut' where multimodal models exploit grid layouts by converting them into text coordinates, shifting visual reasoning onto text deduction. The study created Polaris-Bench to test this by rephrasing tasks in Polar space. Results show a massive performance drop when moving from Cartesian to Polar layouts, proving current models lack topology-invariant visual reasoning.