NUS's Velox Routes Video World Model Attention via Frustum Slabs, 4.36x Attention and 2.43x Denoising Speedup at Chunk 48

Shizun Wang and Xinchao Wang (NUS) propose Velox, a training-free attention router that treats a video world model's history as camera frustum slabs with depth intervals and selects context per region using streaming geometry. On an H200 at chunk 48, Velox speeds attention up 4.36x and denoising 2.43x against dense LingBot-Fast-Cam while staying within 0.15 average points on the 158-case WBench navigation split.