Alaya-EVOKE: From Linear-Scaling Supervision to Endless World
Yuanyang Yin, Gongxuan Wang, Yifan Zhan, Chuanhao Li, Kaipeng Zhang, Feng Zhao
cs.CV
2026-08-14
Evoke externalizes scene memory into a geometry bank and redesigns its teacher for long supervision, letting a three-step student run two uninterrupted hours on one H200.
Interactive world models (systems that generate video in real time as users act) need to do three things that pull against each other: remember what happened earlier, respond to input with low latency, and keep running for a long time without breaking down. Storing history in the denoiser's context or KV cache makes per-step cost grow with session length, forcing a tradeoff between how long a session can be and how much it remembers. Low latency, meanwhile, requires generating in just a few denoising steps, and a few-step student model can only be as capable as the slower teacher it was distilled from. Evoke's problem statement is this systemic conflict: how to give a three-step, low-latency model both persistent memory of earlier scenes and the ability to stay stable and responsive over long, open-ended sessions.
Evoke splits the problem in two. First, persistent scene state moves out of the denoiser entirely, into an external world state bank indexed by camera pose. Each generated 1.5-second chunk gets its depth estimated and unprojected into a point cloud that's written into the bank; when the camera later returns to a previously seen location, the bank renders that view directly and feeds it in as pixel-level conditioning. This keeps the denoiser's per-step token budget fixed regardless of session length — a longer session means more recurrent calls, not a bigger call each time.
Second, the teacher model used for distillation gets redesigned. The paper's premise is that a few-step student inherits only what its teacher's supervision covers: a teacher that only scores short windows never exposes the student to slow content drift that only shows up over longer spans. Evoke's teacher uses chunk-wise sparse attention (each chunk sees a local window, a handful of distant keyframes, and a linear-attention-compressed global state), which turns attention cost from quadratic to roughly linear in sequence length and makes it affordable to score a full 30-second window at once. The same chunk structure assigns each segment its own text condition, so the teacher also experiences mid-sequence prompt changes during training. Both capabilities, long-range stability and responsiveness to new instructions mid-session, get transferred to a three-step, classifier-free-guidance-free student through distribution-matching distillation.
On WBench's navigation split (158 scenarios, identical conditions across all systems), Evoke leads on video quality, scene setting, and physical plausibility among nine evaluated models:
| Metric group | Evoke | Best competitor |
| Video quality avg. | 82.79 | 81.77 (HappyOyster) |
| Scene setting avg. | 83.76 | 77.90 (LingBot-World v2 fast) |
| Physical plausibility avg. | 72.06 | 69.05 (HappyOyster) |
Evoke ranks 1st of 10 on VBench-2.0 (66.77) and 7th of 10 on VBench-Long (85.11, behind leader IPOW's 88.26). This isn't a step-matched comparison: Evoke samples in three steps with no CFG, while competitors run their own many-step default samplers. Long-session testing (eight continuous 65.5-minute rollouts, 2,619 chunks total) shows color statistics stabilize after an initial transient rather than progressively drifting, and scene-identity cosine similarity settles at 0.523, matching what real video scores against itself 60 seconds apart. A revisit-recall test on the world state bank finds PSNR improves by 2.3-3.2 dB once the retention window covers the time spent away, plateauing at 15.4-17.8 dB, recognizable rather than pixel-faithful reconstruction. For mid-session prompt changes, a new element gets successfully introduced 67% of the time when targeting unanchored content, but only 4% of the time when the request would overwrite content already anchored by the geometry bank.
For teams building interactive content generation (games, embodied-agent simulation, long-form video), this offers a clean architectural division of labor: spatial memory belongs in an explicit geometric structure that's queryable by camera pose, while long-range temporal consistency belongs to training-time supervision that teaches the model not to drift. Neither substitutes for the other — a longer supervision window doesn't fix can't recall a place already seen, and a geometry bank can't determine how unobserved content should evolve. That division matters for anyone building long-session interactive systems rather than just scaling up parameters. Practically, generating a 1.5-second chunk in 2.11 seconds on a single H200 puts this approach within reach of usable real-time interaction latency.
The authors list three open problems themselves: the geometric world state bank currently captures coarse scene structure, and fine-grained consistency of object identity and appearance remains limited; the bank only stores static geometry, with no mechanism yet for tracking dynamic state like object motion or in-progress transitions; and true real-time interaction still needs faster VAEs and more efficient few-step generators. The long-session stability evidence is also based on a single continuous run (n=1), which the paper itself flags as a stability claim, not a fidelity claim, without repeated trials to confirm robustness. The causal link between teacher supervision-window length and drift improvement isn't fully pinned down either: controlled scoring sweeps found little additional detectability once the window exceeds a modest horizon, meaning the benefit of a long-horizon teacher can't be reduced to window length alone.