Why Far Looks Up: Probing Spatial Representation in Vision-Language Models

Researchers found a persistent spatial bias in vision-language models: vertical image position is often tied to perceived distance.

The bias resembles perspective cues in natural photos, but it creates gaps between normal and counter-heuristic cases.

As models scale and benchmark scores improve, this shortcut can persist or even strengthen, masking weak spatial reasoning.

To reveal it, the team introduced SpatialTunnel, a synthetic benchmark that strips away common image correlations.

Models with more disentangled spatial axes were more reliable across spatial benchmarks and showed better robustness.