Spatial Mental Modeling from Limited Views

People intuitively construct mental models of space beyond what they directly perceive, but can large visual-language models (VLMs) do the same with partial observations like limited views? We identify this significant gap for current VLMs via our new MINDCUBE benchmark with 17,530 questions and 2,919 images, evaluating how well VLMs build robust spatial mental models, representing positions (cognitive mapping), orientations (perspective-taking), and dynamics (mental simulation) for what if movements, to solve spatial reasoning on unseen space that lies beyond immediate perception. We explore three approaches to approximating spatial mental models in VLMs: (1) View interpolation to visualize mental simulation, which surprisingly offers little benefit, highlighting the challenge of reasoning from limited views; (2) Textual reasoning chains, which effectively guide model thinking when supervised; and (3) Structured representations like cognitive maps, where ground truth maps help little, but training VLMs to generate and reason over their own maps yields substantial gains—even if the maps are imperfect. Training models to reason over these internal maps raises accuracy from 38.3% to 61.7% (+23.5%). Adding reinforcement learning further improves performance to 76.1% (+37.8%). Our key insight is that no scaffolding of spatial mental models, actively construct-ing and utilizing spatial mental representations with flexible reasoning chains or processes, significantly improves understanding of unobservable space.