Velox: Spatial Memory Makes World Models Efficient

arXiv 2026 34.4 method

TLDR

Velox introduces a training-free attention router using camera frustum slabs to speed up video world model rollouts while preserving quality on WBench.

Reasoning

The paper addresses a clear bottleneck in video world models—growing KV caches—and reports strong speedups with modest quality loss on a benchmark. However, the method relies on geometric camera assumptions and the abstract does not demonstrate generalization beyond the evaluated navigation setting.

Read-first score

Read-first score 34.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 44.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 25%

62.9

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=metric

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Graph prestige 15%

10.9

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0031

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 292.

Keyword Scores

world model

video world model

generative world model

interactive world model

world simulator

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Training-free attention router (Velox) for region-level context selection in chunked video diffusion world models.
Represents historical blocks as camera frustum slabs with depth intervals and obtains current-block depth through recency-priority multiwarp.
Softly modulates each layer's model-derived block importance based on frustum slab overlap for well-covered regions, while uncovered regions retain the model-derived route.
Uses fixed token-block granularity compatible with FlashAttention-style tiled execution to reduce attention cost growing with rollout length.
Runs frustum slab construction asynchronously with denoising on the same H200.

Methodology

Velox is a training-free attention router for chunked video diffusion world models. It constructs camera frustum slabs with depth intervals for historical blocks, estimates current-block depth via recency-priority multiwarp, and modulates model-derived block importance using frustum slab overlap in covered regions. Fixed token-block granularity matches FlashAttention-style tiled execution, and frustum slab construction runs asynchronously with denoising.

Key Results

At chunk 48, Velox accelerates attention by 4.36× and denoising by 2.43×. On the full 158-case WBench navigation split, Velox remains within 0.15 API-free average points of dense LingBot-Fast-Cam.

Resources

Paper