This post introduces EVOKE, an interactive world model (CV) by Yuanyao Yin, Gongxuan Wang, Yifan Zhan, Chuanhao Li, Kaipeng Zhang, and Feng Zhao, published on arXiv as 2608.13546.
The Problem
Interactive world models must support persistent memory, responsive interaction, and long-horizon generation — but these requirements conflict:
- Maintaining history in the denoiser context or key-value cache incurs ever-growing cost, forcing a trade-off between session length and retained memory.
- Low-latency interaction relies on few-step generation whose capabilities are bounded by its teacher model.
- With bounded context and recurrent external memory, EVOKE supports open-ended, continuously evolving generation.
- On a single H200 at 384×640 resolution, each 1.5-second chunk is generated in 2.11 seconds.
- 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.
Key Contributions
1. Externalized persistent world state: Scene geometry is maintained in an external, camera-indexed world state bank. Only view-relevant information is retrieved, keeping the denoiser context bounded as the session grows.
2. Teacher redesigned for long-horizon supervision: Rather than treating the teacher as a fixed generator, its sparse attention combines chunk-wise grouping, retrieval of selected distant frames, and a linear-attention global state — yielding linear growth in memory and compute while enabling supervision over long horizons. This supervision exposes content drift that stays locally plausible within short windows, while per-chunk conditioning enables prompt changes and event control throughout the sequence.
3. Distillation with a 30-second distribution-matching objective: Applied under self-forced rollouts, it transfers both capabilities to a three-step student that uses no classifier-free guidance, improving resistance to long-term drift while preserving responsive conditioning.