NUS' Velox routes attention via frustum slabs, speeding world model video 4.36x
NationalUniversityofSingapore · hf · 2026-10-16
Researchers at the National University of Singapore (Qiuhong Shen, Shizun Wang, Xinchao Wang) introduce Velox, a training-free attention router that makes long video rollouts in world models efficient.
- Problem: chunked video diffusion models keep a growing KV cache for visual coherence, so later chunks get progressively slower
- Method: represents historical blocks as camera frustum slabs with depth intervals from streaming geometry; recovers current-block depth via recency-priority multiway merging, softly modulating model-derived block importance in well-covered regions while uncovered regions keep the model's own route
- Results: at chunk 48, Velox accelerates attention by 4.36x and denoising by 2.43x; frustum slab construction runs asynchronously with denoising on the same H200, preserving visual coherence
More from Research
- Hironobu Fujiyoshi publishes deck: from AI agents to Agentic Robotics — HirokatuKataoka · 2026-10-16
- Is math just a language? A debate on why LLMs are good at mathematics — LucaAmb · 2026-10-16
- Alibaba open-sources OpenGCUA, a full-stack generalist computer-use agent; +19.6 points on OSWorld-v2 across 8 models — Tongyi-MAI · 2026-10-16
- CoDA co-evolves simulation and robot policies via shared failure memory, lifting OOD success from 5.3% to 27.6% — PLAN-Lab · 2026-10-16
- willcb: higher-level behavior shaping via RL envs is 'almost practical' — willcb · 2026-10-16
- Cambridge NLIP seminar: Danae Sanchez Villegas on multimodal reasoning beyond final answers — nikaletras · 2026-10-16