A new preprint on arXiv, DK-VLM, addresses a known weakness in vision-language models: understanding 3D spatial relations across different views and updating those judgments as the observer moves. The authors note that current models struggle with these tasks, despite progress in other areas of visual reasoning.

The paper describes a common existing strategy: fusing 3D reconstruction features into visual tokens. However, the abstract indicates that aligning these features is problematic. DK-VLM appears to take a different route by using geometry as position, though the abstract cuts off before explaining the full method.

Because the source is a single preprint with an incomplete abstract, the exact mechanism and experimental results are not yet clear. The significance lies in the proposed shift away from feature fusion toward a positional encoding of geometry, which could offer a new direction for spatial reasoning in multimodal models.