Vision-language models have made impressive strides in understanding 2D images, but their ability to reason about 3D space remains limited. A new preprint on arXiv attributes this gap to the passive, view-centric way these models perceive the world: they see a fixed image or video rather than actively exploring a scene. The authors argue that this paradigm prevents VLMs from building the kind of spatial understanding that comes naturally from moving through an environment.
To address this, the paper proposes an approach called active exploration over 3D worlds. Rather than passively consuming views, the model would interact with a 3D environment, selecting where to look and how to move to gather the information needed for spatial reasoning. The abstract does not include experimental results, so the effectiveness of the proposed method is not yet demonstrated.
The preprint is part of a growing line of work that connects language models to embodied, interactive perception. If the approach holds up, it could push VLMs beyond 2D benchmarks toward more robust 3D understanding. As with all preprints, the claims await peer review and further validation.