SpatialASM
A research framework for spatial reasoning in vision-language models. It equips models with active viewpoint planning and scene imagination capabilities through spatial simulation during training and Active Spatial Mentalization (ASM) supervision signals.
Timeline (last 90 days)
SpatialASM is a scalable framework designed to internalize Active Spatial Mentalization (ASM) capabilities into vision-language models during the training stage
The core idea of SpatialASM is spatial simulation at training time, without relying on additional tools at inference time
The SpatialASM framework includes a non-egocentric novel-perspective generator for anchoring hypothetical viewpoints to real visual evidence
SpatialASM produces two types of complementary supervision data: explicit ASM reasoning trajectories and automatically constructed perspective tasks
Explicit ASM trajectories provide dedicated Chain-of-Thought supervision for spatial reasoning, demonstrating complete thinking paths
The research team evaluated SpatialASM on seven benchmarks, covering tasks ranging from spatial understanding to multi-image, cross-view, and perspective transformation reasoning
SpatialASM yielded spatial reasoning performance improvements across four different LVLM backbone models
SpatialASM enhances spatial capabilities while preserving the model's original general visual capabilities
SpatialASM achieves improved average accuracy across all seven spatial benchmarks
SpatialASM provides a scalable path to alleviate the scarcity of spatial reasoning training data by synthesizing ASM trajectories
All Facts (10)
SpatialASM is a scalable framework designed to internalize Active Spatial Mentalization (ASM) capabilities into vision-language models during the training stage
50%UnverifiedThe core idea of SpatialASM is spatial simulation at training time, without relying on additional tools at inference time
50%UnverifiedThe SpatialASM framework includes a non-egocentric novel-perspective generator for anchoring hypothetical viewpoints to real visual evidence
50%UnverifiedSpatialASM produces two types of complementary supervision data: explicit ASM reasoning trajectories and automatically constructed perspective tasks
50%UnverifiedExplicit ASM trajectories provide dedicated Chain-of-Thought supervision for spatial reasoning, demonstrating complete thinking paths
50%UnverifiedThe research team evaluated SpatialASM on seven benchmarks, covering tasks ranging from spatial understanding to multi-image, cross-view, and perspective transformation reasoning
50%UnverifiedSpatialASM yielded spatial reasoning performance improvements across four different LVLM backbone models
50%UnverifiedSpatialASM enhances spatial capabilities while preserving the model's original general visual capabilities
50%UnverifiedSpatialASM achieves improved average accuracy across all seven spatial benchmarks
50%UnverifiedSpatialASM provides a scalable path to alleviate the scarcity of spatial reasoning training data by synthesizing ASM trajectories
50%