← back to paper
arxiv: 2606.02742 · 2 revisions
Consistent Yet Wrong: Evidence Insensitivity in Spatial Vision-Language Models