DK-VLM: Geometry as Position for Spatial Reasoning in Vision-Language Models
DK-VLM introduces Dual-Kernel Attention to compute attention using both a token's input position and its 3D geometric position in a scene. This is paired with Reference-Frame Composition supervision to train the model to update its spatial reference frame across multiple judgments. The result is state-of-the-art performance on six spatial benchmarks, showing geometry as direct positional information for better reasoning.