Velox: Spatial Memory Makes World Models Efficient
To address this bottleneck, we introduce Velox, a training-free attention router for region-level context selection using frustum slab correspondences from streaming geometry.
ProofPaper ↗
Key points
- Video world models turn actions and camera motion into long visual rollouts, making video a natural interface for interactive simulation.
- Chunked video diffusion models usually keep this context in growing KV caches, so later chunks become progressively slower.
- Velox represents historical blocks as camera frustum slabs with depth intervals and obtains current-block depth through recency-priority multiwarp.
- Uncovered regions retain the model-derived route without geometric bias.
Sources (2)
- [1]Velox: Spatial Memory Makes World Models EfficientHugging Face Daily Papers · Oct 15, 12:00 AM
To address this bottleneck, we introduce Velox, a training-free attention router for region-level context selection using frustum slab correspondences from streaming geometry.
Video world models turn actions and camera motion into long visual rollouts, making video a natural interface for interactive simulation.
- [2]Velox: Spatial Memory Makes World Models EfficientarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 05:50 PM · same content
Extractive summary: sentences quoted from the sources.
Before this
- Oct 14, 2026StateCraft: Augmenting Video World Models with Abstract-State Reasoning
- Oct 14, 2026SonoWorld: Towards Interactive Pose-Controllable Ultrasound World Models
- Oct 14, 2026Glance Forcing: Making Bidirectional Diffusion Autoregressive at a Glance
- Oct 14, 2026The Design Space of Tri-Modal Masked Diffusion Models
- Oct 13, 2026Gemini 4 Argon for Robotics
- Oct 13, 2026LongForcing: Scaling Temporal Manifold Matching for Long-Horizon World Rollout
