Alaya-EVOKE: From Linear-Scaling Supervision to Endless World

Authors: Yuanyang Yin, Gongxuan Wang, Yifan Zhan, Chuanhao Li, Kaipeng Zhang, Feng Zhao

arXiv ID: 2608.13546

Problem: Interactive world models need persistent memory, responsive interaction, and long-horizon generation, but these pull against each other: keeping history in the denoiser context or key-value cache makes cost grow with session length, while low-latency few-step generation is bounded by its teacher's capabilities. Long horizons also expose content drift that stays locally plausible within short windows.

Key Methodology:

Externalizes persistent world state into a camera-indexed world-state bank; only view-relevant information is retrieved, keeping the denoiser context bounded as the session grows, with per-chunk conditioning enabling prompt changes and event control throughout the sequence.
Redesigns the teacher for long-horizon supervision: sparse attention combining chunk-wise grouping, retrieval of selected distant frames, and a linear-attention global state yields linear growth in memory and compute while supervising over long horizons.
Distills the teacher's drift resistance into a three-step student (no classifier-free guidance) via a 30-second distribution-matching objective applied under self-forced rollouts, transferring long-term drift resistance while preserving responsive conditioning.

Key Results:

With bounded context and recurrent external memory, Evoke supports open-ended, continuously evolving generation: on a single H200 at 384x640, each 1.5s chunk is generated in 2.11s.
As a three-step world model, Evoke achieves state-of-the-art performance on WBench while remaining competitive on VBench-Long and VBench-2.0.

Applied Context: The design splits into two transferable patterns. First, the serving shape: persistent state lives outside the context and is retrieved per view, so session length stops taxing the hot path - the same external-state architecture agent-memory research has been converging on, demonstrated at the end where context cost grows per frame. Second, the training pattern: a linear-scaling long-horizon teacher supervises a cheap few-step student under self-forced rollouts, which is how an interactive surface gets long-horizon robustness without paying long-horizon inference costs at runtime. Caveats: results are vendor-reported on public benchmarks with no independent replication yet, and the world-state bank's retrieval quality will bound how far the bounded context can stretch.