SpatialASM: Internalizing Active Mentalization for Spatial Reasoning
The paper proposes training LVLMs to plan imagined viewpoints before answering spatial questions.
SpatialASM uses a novel-view generator to anchor hypothetical perspectives in visual evidence, then creates supervision for view planning and viewpoint-conditioned scene imagination. The authors say this targets Active Spatial Mentalization, a capability they argue is still weak in current large vision-language models. Across seven spatial reasoning benchmarks, the framework improves average accuracy on four LVLM backbones while preserving general vision performance. ArXiv · AI/CL/LG's note
SpatialASM uses a novel-view generator to anchor hypothetical perspectives in visual evidence, then creates supervision for view planning and viewpoint-conditioned scene imagination. The authors say this targets Active Spatial Mentalization, a capability they argue is still weak in current large vision-language models. Across seven spatial reasoning benchmarks, the framework improves average accuracy on four LVLM backbones while preserving general vision performance. ArXiv · AI/CL/LG's note
score 5