A multimodal large language model (an artificial intelligence system that reads images and text together) can give different answers about the same object when the camera angle changes.
Lie routing addresses this by using a pretrained geometry extractor to guide small rotations of the model's visual features. The rotated features are blended back with the original ones, so the model keeps its prior knowledge while adjusting to the new angle.
It improves several three-dimensional scene tasks compared with simply adding geometry features, and its answers stay steadier as the camera moves. This matters because camera angle should not change what a system says is there.