News · 2026-08-16

A world model that keeps the map outside the model

A new interactive world model called Evoke handles persistent memory by not keeping it in the model at all. Scene geometry lives in an external bank indexed by camera position, and only the parts relevant to the current view get pulled in, so the generative model's working context stays a fixed size however long the session runs. On a single H200 at 384 by 640 resolution, each 1.5-second chunk is generated in 2.11 seconds.

Key facts

• Evoke generates 1.5 seconds of video in 2.11 seconds on one H200 GPU, using a three-step sampler with no classifier-free guidance.
• Persistent scene state is stored outside the model in a camera-indexed world state bank, keeping the denoiser context bounded as sessions grow.
• The paper reports state-of-the-art results on WBench while staying competitive on VBench-Long and VBench-2.0.
• Primary sources: arXiv:2608.13546 and the AlayaLab/Evoke repository.

The problem Evoke targets is a genuine three-way squeeze, and the paper states it cleanly: interactive world models "must support persistent memory, responsive interaction, and long-horizon generation, yet these requirements place conflicting demands on the model."

Take them one at a time. Persistent memory means that if you turn the camera away from a room and turn back, the room should still be there. The obvious way to get that is to keep the history in the model's context or its key-value cache, but that cost grows every second you play, which means a hard trade between how long a session can run and how much it remembers. Responsive interaction means the model must generate in very few denoising steps, and a few-step model is capped by the quality of the teacher it was distilled from. Long-horizon generation means avoiding the slow drift where a scene stays locally plausible frame to frame but wanders somewhere else entirely over a minute.

Evoke's answer separates the memory from the generator. The world state bank holds the scene geometry outside the model, indexed by camera position, and supports reads, writes, and evictions. When the view moves, the system retrieves only view-relevant information. The generative model never has to hold the whole world, only the part it is currently looking at, which is much closer to how a video game engine works than to how a video diffusion model normally works. The team behind it, from Alaya Lab, describes the goal as open-ended, continuously evolving generation rather than a fixed window.

The second half of the design is about the teacher. Rather than distilling a fast student from an off-the-shelf generator, the researchers rebuilt the teacher specifically for long-horizon supervision: sparse attention that groups frames into chunks, retrieval of selected distant frames, and a linear-attention global state, which together make memory and compute grow linearly with sequence length instead of quadratically. That is what lets them supervise over long enough spans to expose drift that short-window training would never catch. A 30-second distribution-matching objective under self-forced rollouts then transfers those capabilities into the three-step student.

The reason to care is that this is an architectural bet, not a scaling result. The prevailing approach to memory in generative video has been to make the context bigger and pay for it. Evoke argues that persistence belongs in an external, explicitly managed store, and that the generator should be a renderer that queries it. If that holds up, it is the same structural move that retrieval-augmented generation made for language models, applied to visual state.

The caveats are real and the timing makes them sharper. The benchmark results are self-reported, as they are for essentially every model in this field right now. The memory is bounded, not infinite, so this is managed persistence rather than the "endless world" the branding suggests. And 2.11 seconds per 1.5 seconds of video is not yet real time, though it is close enough for the prompt changes and event control the system is built to support.

Most importantly, the field just got a hard reminder that long-horizon consistency is not solved. Ground Truth reported this week on a closed-loop benchmark that caught nine world models forgetting the room once an agent pursued a goal inside the generated world instead of following a script. Evoke is not evaluated on that benchmark. The reasonable read is that interactivity is improving faster than persistence, and that these two papers are arguing about the same weakness from opposite ends: one proposing a fix, the other showing how much is left to fix.

Background reading: world models, diffusion distillation, and classifier-free guidance.

Primary source, verified: read the paper → (arXiv 2608.13546)

Ground Truth. (2026, August 16). A world model that keeps the map outside the model. Ground Truth. https://groundtruth.day/news/a-world-model-that-keeps-the-map-outside-the-model.html

@misc{groundtruth:a-world-model-that-keeps-the-map-outside-the-model,
  title  = {A world model that keeps the map outside the model},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/a-world-model-that-keeps-the-map-outside-the-model.html}
}

Topics: world-models · video-generation · memory · interactive · research