Multimodal AI models struggle to understand other people's visual perspectives, raising concerns about their use in collaborative settings.

Researchers evaluated multimodal language models (MLMs) using two perspective-taking tasks adapted from human studies: the Director Task and the Rotating Figure Task.

MLMs show significant deficits in Level 2 perspective-taking, which requires suppressing one's own viewpoint to adopt another person's perspective.

Current MLMs have critical limitations in representing and reasoning about alternative visual perspectives, with implications for their deployment in social and collaborative applications.

Existing benchmarks rely mainly on text-based scenarios, leaving visuospatial perspective-taking largely unexplored until this study.