Spatial Reasoning
The capacity of an AI model to perform logical inference over two-dimensional or multi-dimensional structured data, encompassing positional state maintenance, cross-region constraint reasoning, and grid manipulation. It serves as an important dimension for assessing a model's structured reasoning capability.
Timeline (last 90 days)
In static analysis evaluations, common spatial reasoning mistakes by models include: writing relative positions as absolute coordinates, causing offset calculations to be wrong; getting the order of rotation and translation wrong; and ignoring parent-child hierarchy, causing overall misalignment.
Spatial reasoning refers to the ability of an agent to represent and infer the position, orientation, size, occlusion relationships, and motion trajectories of objects in a three-dimensional environment.
Fold animations require the model to correctly calculate the rotation axis and surface deformation of two screen segments at different fold angles, serving as a test of spatial reasoning ability.
Capabilities such as video understanding and spatial reasoning currently lack a unified, authoritative evaluation framework.
Current large language models and vision models remain relatively weak at reasoning about three-dimensional spatial relationships.
Current large language models and vision models have made significant progress in understanding two-dimensional space, but their reasoning about three-dimensional spatial relationships remains relatively weak.
All Facts (6)
In static analysis evaluations, common spatial reasoning mistakes by models include: writing relative positions as absolute coordinates, causing offset calculations to be wrong; getting the order of rotation and translation wrong; and ignoring parent-child hierarchy, causing overall misalignment.
50%UnverifiedSpatial reasoning refers to the ability of an agent to represent and infer the position, orientation, size, occlusion relationships, and motion trajectories of objects in a three-dimensional environment.
50%UnverifiedFold animations require the model to correctly calculate the rotation axis and surface deformation of two screen segments at different fold angles, serving as a test of spatial reasoning ability.
50%UnverifiedCapabilities such as video understanding and spatial reasoning currently lack a unified, authoritative evaluation framework.
50%UnverifiedCurrent large language models and vision models remain relatively weak at reasoning about three-dimensional spatial relationships.
50%UnverifiedCurrent large language models and vision models have made significant progress in understanding two-dimensional space, but their reasoning about three-dimensional spatial relationships remains relatively weak.
50%