SpatialASM: Internalizing Active Mentalization for Spatial Reasoning
arXiv cs.AI · 10 Oct 2026, 02:39 AEST
SpatialASM introduces a framework to enhance Large Vision-Language Models' spatial reasoning through Active Spatial Mentalization. The approach improves performance on various spatial benchmarks while maintaining general vision capabilities.
Key facts
- When solving complex spatial reasoning tasks, ordinary humans naturally revisit the scenes from task-relevant viewpoints and mentally reconstruct the spatial layouts from those viewpoints.
- We term this ability Active Spatial Mentalization (ASM), an intrinsic component of visual intelligence that remains substantially underdeveloped in current LVLMs.
- SpatialASM employs an allocentric novel-view generator to ground hypothetical viewpoints in visual evidence and produces two complementary forms of supervision.
- On seven benchmarks spanning relatively perception-oriented spatial understanding to more challenging tasks requiring multi-image, cross-view and viewpoint-transformation reasoning, SpatialASM yields performance gain on LVLMs of four LVLM backbones in spatial reasoning while retaining their original general vision capabilities.
Read aloud
SpatialASM introduces a framework to enhance Large Vision-Language Models' spatial reasoning through Active Spatial Mentalization. The approach improves performance on various spatial benchmarks while maintaining general vision capabilities. When solving complex spatial reasoning tasks, ordinary humans naturally revisit the scenes from task-relevant viewpoints and mentally reconstruct the spatial layouts from those viewpoints. We term this ability Active Spatial Mentalization ASM, an intrinsic component of visual intelligence that remains substantially underdeveloped in current LVLMs. SpatialASM employs an allocentric novel-view generator to ground hypothetical viewpoints in visual evidence and produces two complementary forms of supervision. On seven benchmarks spanning relatively perception-oriented spatial understanding to more challenging tasks requiring multi-image, cross-view and viewpoint-transformation reasoning, SpatialASM yields performance gain on LVLMs of four LVLM backbones in spatial reasoning while retaining their original general vision capabilities.
Sources in this cluster
- SpatialASM: Internalizing Active Mentalization for Spatial Reasoning
arXiv cs.AI · 10 Oct 2026, 02:39 AEST