SpatialASM: Internalizing Active Mentalization for Spatial Reasoning
SpatialASM trains vision-language models to plan imagined viewpoints before answering spatial questions.
The paper defines that skill as Active Spatial Mentalization, modeled on how people revisit a scene from task-relevant angles. Its framework uses a novel-view generator to create training supervision for view planning, scene imagination, and perspective-taking tasks. Across seven spatial reasoning benchmarks and four LVLM backbones, the authors report higher average accuracy while preserving general vision performance. HF Daily Papers' note
The paper defines that skill as Active Spatial Mentalization, modeled on how people revisit a scene from task-relevant angles. Its framework uses a novel-view generator to create training supervision for view planning, scene imagination, and perspective-taking tasks. Across seven spatial reasoning benchmarks and four LVLM backbones, the authors report higher average accuracy while preserving general vision performance. HF Daily Papers' note
score 5