In a new research paper, Hugging Face introduces MASS (Multiplayer world models with Authoritative Shared State), a novel approach to world modeling that tackles the challenges of multi-agent environments. Traditional video world models often struggle in such settings because they conflate world state with view-dependent visual latents, leading to redundant computation, inconsistent views, and poor scalability.
MASS draws inspiration from multiplayer game architectures by explicitly separating world dynamics from view rendering. A learned Logic Engine advances a global, authoritative typed state based on joint actions, without requiring any hand-written transition function. This shared state serves as the sole recurrent memory and synchronization reference. From this state, a learned Rendering Engine generates independent and consistent views for any requested camera on demand.
This disentanglement allows MASS 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. Notably, the model can advance predicted worlds with 1,024 concurrent players for 10,000 recurrent steps.
The authors argue that explicit, authoritative state modeling provides a practical foundation for scalable and consistent multi-agent world simulation, potentially paving the way for more efficient and realistic virtual environments.