Alaya-EVOKE Externalizes World State Into a 'Camera-Indexed State Bank' to Get Linear-Cost Long-Horizon Video on One H200

arXiv (via HuggingFace Daily Papers)·medium signalXBlueskyLinkedInCopy link

Yuanyang Yin, Kaipeng Zhang, Feng Zhao and co-authors submitted Alaya-EVOKE on August 13 (arXiv 2608.13546), an interactive world model that generates each 1.5-second chunk in 2.11 seconds on a single H200 at 384×640, using a three-step student with no classifier-free guidance. The architectural move is externalizing persistent world state into a camera-indexed state bank and combining chunk-wise grouping, retrieval of selected distant frames, and a linear-attention global state so memory and compute grow linearly rather than quadratically with session length. It reports state-of-the-art on WBench while staying competitive on VBench-Long and VBench-2.0.