Velox: Spatial Memory Makes World Models Efficient

Velox. Spatial Memory Makes World Models Efficient. By Qiuhong Shen and two others, from the National University of Singapore. Published on the fifteenth of October. No code repository is linked to it.

Video world models turn actions and camera motion into long visual rollouts, which makes video a natural interface for interactive simulation. Interaction of that kind needs responsive generation. A coherent rollout, though, also has to retain visual memory of everything it has already generated. Chunked video diffusion models usually hold that context in key-value caches that keep growing, and so each later chunk becomes slower than the ones before it. Responsiveness and memory pull against each other, and the longer the rollout runs, the harder they pull.

Velox is a training-free attention router. Rather than attending to the whole history, it selects context at the level of regions, and it decides which regions matter using frustum slab correspondences taken from streaming geometry. The history stops being a flat pile of tokens and becomes a set of places, each of which either overlaps what the camera is looking at now or does not.

Each historical block is represented as a camera frustum slab with depth intervals. Depth for the current block is obtained through what the authors call recency-priority multiwarp. Where a region is well covered, the overlap between frustum slabs softly modulates the block importance the model itself derives, layer by layer. Where a region is uncovered, the model-derived route is kept as it is, with no geometric bias added. Two practical choices sit underneath. The fixed token-block granularity matches FlashAttention-style tiled execution, which reduces the attention term that grows with rollout length. And slab construction runs asynchronously with denoising, on the same H200.

At chunk forty-eight, Velox accelerates attention by four point three six times, and denoising by two point four three times. Quality is checked on the full WBench navigation split, one hundred and fifty-eight cases, where Velox stays within zero point one five API-free average points of dense LingBot-Fast-Cam.

Velox retrains nothing, and it does not throw away the model's own sense of what matters either. Geometry acts as a soft nudge where the scene is well observed, and the model's routing stands untouched where it is not. That pairing is what lets later chunks stop paying for the whole of their past while staying close to the dense model they started from.