4.36x Faster World Model Attention with Zero Training: NUS Introduces Velox Spatial Memory, Alongside Nankai's DeltaStreamer and HKPolyU's CoWorld Reasoner

Core Background & Industry Pain Points

As interactive Video World Models (VWMs) and streaming video agents power embodied simulation and real-time video assistants, three systemic bottlenecks limit long-horizon deployment: (1) chunked autoregressive video diffusion models retain historical context in continuously growing KV caches, causing quadratic attention slowdowns as rollout chunks accumulate; (2) streaming video agents compress historical frames before knowing future user queries, losing critical event transitions with no way to recover omitted details; and (3) Vision-Language Models (VLMs) invoke world models through rigid single-purpose pipelines while lacking standardized benchmarks to evaluate dynamic state persistence across repeated camera departures and revisits.

Architectural Highlights & Core Mechanisms

Four complementary papers published today introduce geometric KV routing, replay-based streaming memory, self-evolving VWM harnesses, and memory diagnostics:

• NUS's Velox: Training-Free Frustum-Slab Attention Router (arXiv:2610.22311): Represents historical video chunks as 3D camera frustum slabs with depth intervals (constructed asynchronously alongside denoising on the same H200 GPU via recency-priority multiwarp). For spatially covered regions, 3D frustum overlap modulates model-derived token-block importance at a fixed tile granularity matching FlashAttention, pruning historical KV attention overhead without retraining.
• Nankai University's DeltaStreamer: Change-Aware Memory & Targeted Frame Replay (arXiv:2610.19849, 39 upvotes): Continuously records visual state transitions in structured memory prior to query arrival; once a question is posed, a query-driven planner identifies missing visual evidence and triggers targeted replay of historical raw frames to ground answers in exact visual details.
• HKPolyU's CoWorld Reasoner & ZJU's WorldMemBench (arXiv:2610.21910 & arXiv:2610.19063): CoWorld Reasoner wraps a frozen VLM and VWM in a self-evolving VLM-VWM Harness that iteratively refines visual imagination tools, skills, and execution code from feedback. Concurrently, Zhejiang University's WorldMemBench introduces 260 multi-revisit test cases evaluating both static retention (identity/geometry) and dynamic state evolution across 9 video world models.

Authoritative Benchmarks & Empirical Comparisons

• Velox Achieves 4.36x Attention & 2.43x Denoising Speedup (arXiv:2610.22311): At chunk 48 on an NVIDIA H200, Velox accelerates attention by 4.36x and overall diffusion denoising by 2.43x, while remaining within 0.15 API-free average points of dense LingBot-Fast-Cam across the 158-case WBench navigation split.
• DeltaStreamer on OVO-Bench & StreamingBench (arXiv:2610.19849): Achieves 66.07% overall accuracy on OVO-Bench and 59.53% on StreamingBench, outperforming representative streaming video understanding baselines.
• CoWorld Reasoner Across 7 Spatiotemporal Benchmarks (arXiv:2610.21910): Improves over direct VLM reasoning across all 7 benchmarks—including relative accuracy gains of +11.8% on MMSI and +13.0% on ERQA—and beats all five evolving baselines across five benchmarks.

Developer Practical Guide & Quickstart

Developers can explore Velox (https://velox-world.com/velox/), clone DeltaStreamer (https://github.com/BBBBchan/DeltaStreamer) and CoWorld Reasoner (https://github.com/ModalityDance/CoWorld-Reasoner), and benchmark video world memory persistence using WorldMemBench (https://github.com/ZJU-REAL/WorldMemBench).