When VLMs Destroy What They Know: Spatial Representations Fall Below Lexical Priors

In 30 Sekunden

Was passiert ist

Researchers found that vision-language models represent spatial relationships far less faithfully than word meanings, despite performing well on compositional benchmarks.

Warum es zählt

Matters for engineers building or evaluating VLMs for tasks requiring accurate scene understanding, object relationships, or spatial reasoning.

Achtung

The models lose spatial information during aggregation into global representations, so output-based evaluation alone misses this systematic failure.