News · 2026-07-19

Two New Papers Split World Models Into Simulators and Streamers

Two new papers stake out opposite ends of where interactive video AI is heading, and together they suggest the field of 'world models' is splitting in two. One, 'From Pixels to States', argues that a real interactive world needs the discipline of a game engine — explicit state, persistent rules, real-time operation — not just a video model predicting the next frame. The other, 'Video = World + Event Stream', reframes any video as a stable world plus a stream of events, and uses that split to build a real-time, full-duplex audio-visual agent.

Key facts

• Paper one: arXiv 2607.14076, 'From Pixels to States: Rethinking Interactive World Models as Game Engines', a cs.CV submission from July 15, 2026 with seven authors.
• Paper two: arXiv 2607.15038, 'Video = World + Event Stream', submitted July 16 and revised July 17, 2026, with an official Wan Streamer v0.3 project page.
• The data engine: paper one builds a scalable data engine on the game Black Myth: Wukong, collecting over 90 hours of gameplay with frame-aligned actions, ground-truth game states, and observations.
• The system spec: Wan Streamer v0.3 runs 640x368 video at 25 fps with a 160 ms streaming unit and about 550 ms total interaction latency.

Some background. A world model is an AI system that learns how an environment behaves so it can predict what happens next, and the term has been stretched to cover everything from pretty video generators to robot simulators — which is exactly why researchers recently wrote a 58-page definition to pin it down. These two papers are the next move: instead of defining the term, they propose two different engineering directions for making video systems genuinely interactive.

'From Pixels to States' makes a pointed argument. Modern video generators can predict future frames conditioned on a user's action, which looks like interactivity, but the paper insists a true game world also needs rules, persistent consequences, and real-time operation — the classic loop of action, then an explicit state update, then a rendered observation. Its concrete contribution is a data engine that mines over 90 hours of Black Myth: Wukong gameplay, capturing not just the pixels but the frame-aligned player actions, the ground-truth game state, and structured annotations. The analogy is that a video model is like an artist painting plausible next moments, while a game engine is a bookkeeper that actually tracks where every object is and what the rules allow — and interactivity needs the bookkeeper.

'Video = World + Event Stream' comes at it from the other side. It proposes that any video decomposes into a world — the persistent context like environment, subjects, ambient sound, and voice characteristics — plus an event stream, the time-varying changes like behavior, speech, and scene shifts. The training trick is to establish the world once and then learn what happens next over time, which turns ordinary video into pretraining data. The team instantiates this as Wan Streamer v0.3, a real-time full-duplex audio-visual agent where the event stream is the agent's own speech and free-form behavior, holding roughly 550 ms total interaction latency under a tight network budget. Its project page notes that while v0.3 broadens the system toward a general video learner, this release focuses on real-time conversation.

Why they belong together: both are explicitly about adding a stronger notion of persistent world state to interactive video, but they diverge on how. One track makes world models more like simulators with explicit, inspectable state; the other makes video models more like streaming agents that carry persistent context and emit events. That is a useful map of a field that has been muddled by a single overloaded term.

Why it matters: interactive world models sit under the ambitions for game generation, robot simulators, and embodied agents, so whether the field converges on 'simulators with state' or 'streaming video agents' shapes what gets built next. The honest caveat is that these are early framework and system papers with the authors' own reported results — the Black Myth data engine and the Wan Streamer latency numbers are real contributions, but neither paper is a finished, independently benchmarked product.

Primary source, verified: read the paper → (arXiv 2607.14076)

Ground Truth. (2026, July 19). Two New Papers Split World Models Into Simulators and Streamers. Ground Truth. https://groundtruth.day/news/world-models-split-into-simulators-and-streamers.html

@misc{groundtruth:world-models-split-into-simulators-and-streamers,
  title  = {Two New Papers Split World Models Into Simulators and Streamers},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/world-models-split-into-simulators-and-streamers.html}
}

Topics: research · world-models · video-generation · game-engines · interactive-ai