Researchers have introduced SpatialASM, a framework designed to give large vision-language models (LVLMs) a more human-like form of spatial reasoning. The work centers on what the authors call Active Spatial Mentalization (ASM): the ability to revisit a scene from task-relevant viewpoints and mentally reconstruct its spatial layout, an ability the paper argues remains substantially underdeveloped in current LVLMs.
SpatialASM internalizes this ability through training-time spatial simulation. It relies on an allocentric novel-view generator that grounds hypothetical viewpoints in visual evidence, and it produces two complementary forms of supervision. The first is explicit ASM trajectories that teach models when and how to adaptively invoke view planning and imagination during spatial reasoning. The second is automatically constructed perspective-taking tasks that further strengthen viewpoint-conditioned scene imagination.
The authors report that SpatialASM improves spatial reasoning performance across four LVLM backbones and seven benchmarks, spanning relatively perception-oriented spatial understanding as well as more challenging tasks requiring multi-image, cross-view and viewpoint-transformation reasoning. According to the paper, the approach also preserves the models' original general vision capabilities.
The work is framed as a scalable pathway to alleviate the scarcity of spatial Chain-of-Thought (CoT) supervision, which the authors say is needed to elicit active viewpoint exploration and mental scene simulation. Notably, the framework requires neither additional views nor a renderer at inference time.
The researchers describe SpatialASM as an attempt to move LVLMs beyond passively observing given images toward actively reasoning by planning task-relevant viewpoints and imagining what should be visible from them, and they invite feedback from the community.