Ziqi Cai, Siqi Yang, Yimu Wang, Zixian Gao, Yunheng Liu, Shuchen Weng, Erwin Wu, Kaipeng Zhang, Boxin Shi
4 min
Traditional video world models often struggle in multiplayer environments because they entangle the world state with view-dependent visual latents. This leads to redundant computation, inconsistencies between different players' views, and poor scalability. The authors ask: can we build a scalable, consistent multiplayer world model by separating the simulation of the world from the rendering of individual player views?
Inspired by traditional multiplayer game server architectures, the authors propose MASS. Instead of using pixels or visual latents as the recurrent memory, MASS uses an explicit, typed state.
In a matched multiplayer Snake benchmark, MASS significantly outperformed state-of-the-art baselines. It achieved a state recovery score of 0.764, compared to 0.128 for the best video-based baseline. Because every view is decoded from the same authoritative state, MASS eliminates cross-view inconsistency (0.000 disagreement). The system demonstrated the ability to simulate 1,024 concurrent players for 10,000 recurrent steps, proving that explicit state modeling provides a robust foundation for scalable multi-agent simulation.
By decoupling world dynamics from rendering, MASS solves the scalability bottleneck where simulation cost previously scaled with the number of views. This approach allows for consistent, high-fidelity multiplayer simulations that remain stable over long rollouts, providing a practical framework for building complex, multi-agent virtual environments.
Current video world models struggle in multiplayer environments because they entangle world state with view-dependent visual latents, leading to redundant compute, view inconsistencies, and poor scalability. We propose MAS (Multiplayer world models with Authoritative Shared State) to resolve this limitation. Inspired by multiplayer game architectures, MAS disentangles world dynamics and view rendering. A learned Logic Engine advances a global, authoritative typed state from joint actions without any hand-written transition function, acting as the sole recurrent memory and synchronization reference. From this shared state, a learned Rendering Engine generates independent and consistent views for any requested camera on demand. This explicit disentangling allows MAS to achieve superior state accuracy and lower cross-view inconsistency compared to state-of-the-art multi-view baselines on a matched multiplayer Snake benchmark. It advances predicted worlds with 1,024 concurrent players for 10,000 recurrent steps. Our results show that explicit, authoritative state modeling provides a practical foundation for scalable and consistent multi-agent world simulation.
Alex: And when a player does need to see something—how does the list of facts become an image?
Sam: That's the second component: the Rendering Engine. It takes that official record and translates it into a visual image from a specific player's point of view. Because this step is entirely separate from the simulation, a thousand players can each get their own view of the same world—from different angles, at different moments—while the underlying world state is only being calculated once. You're not duplicating the work; you're just presenting the same truth differently to each person.
Alex: It's a bit like a theatre production. The script—the truth of the story—exists once. But every audience member experiences it from their own seat.
Sam: That's a useful parallel. And it's why the architecture scales in a way that video-based models simply can't.
Alex: Did the researchers actually test whether this holds up in practice?
Sam: They did. They ran comparisons between MASS and existing video-based models using a multiplayer version of the game Snake. MASS was significantly more accurate at tracking the true game state over time, and it remained stable with over a thousand players running simultaneously—something the video-based approaches couldn't reliably manage.
Alex: So the central insight is really about separation. Keep the truth of the world in one place, and let the visuals be derived from it—rather than trying to bake both together.
Sam: That's the heart of it. By treating the world state as a single authoritative object, MASS behaves much more like a real, scalable game server than like a video prediction system. It's a meaningful shift in how we think about what a world model actually needs to do.
Alex: Thanks for walking us through that, Sam. And thank you for listening to ResearchPod.