Do spatial foundation models actually work everywhere?

Most spatial foundation models shine on their test sets but stumble when facing different viewpoints, scene types, or data densities. SpatialBench—a 19-dataset benchmark spanning 5 domains—evaluates 41 models across 546 scenes to measure true generalization. Key finding: models need full-context attention for accuracy and bounded memory for long sequences, but dataset quality and domain alignment matter far more than pure scale. Authors release DA-Next-5M dataset and DA-Next baseline model.