SPACENUM: Revisiting Spatial Numerical Understanding in VLMs

In three linesSpaceNum evaluates spatial numerical understanding in VLMs through bidirectional tasks (Num2Space, Space2Num). Current models largely fail to ground numbers in spatial meaning, performing near random chance. They rely on shallow spatial cues and fail to build stable coordinate-aware representations.

Summary generated by Claude — human-verified