Rotary Positional Embedding (RoPE) is widely used to encode relative positions in Transformers. In multi-view settings, however, patch coordinates are defined independently for each image, so the same 3D point can therefore be represented differently across views. This inconsistency makes it harder for attention mechanisms to relate patches that correspond to the same physical location.
The paper proposes WorldRoPE, a probabilistic world-centric rotary embedding for multi-view attention. Rather than relying on per-image patch coordinates, WorldRoPE aims to encode positions in a shared world frame, allowing the model to reason about spatial relationships across images more coherently.
The available abstract focuses on the motivation and the proposed approach; it does not include experimental results. The work is positioned as a way to make multi-view attention more consistent by grounding positional information in the world rather than in individual images.