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

Researchers found that vision-language models often entangle vertical image position with perceived distance, revealing a persistent spatial bias across model families.

Using minimal contrastive pairs, the team tested how spatial information is represented internally rather than relying only on benchmark scores.

The internal vertical-distance entanglement predicted drops in accuracy and robustness, even when models looked strong on standard tasks.

To surface the problem more clearly, they introduced SpatialTunnel, a synthetic benchmark that removes common correlations from natural images.