World models for multiplayer settings need to anticipate how independent actions change a shared environment, including areas outside a player's view. The authors note that point-of-view video only reveals changes once they enter the frame, while panoramic video captures the full surroundings. However, no existing Minecraft dataset offered synchronized POV and panoramic footage with actions.

To fill this gap, the paper presents PAWA, a dataset and framework. It pairs each player's POV video with world-stabilized panoramic video and aligned actions. The collection pipeline separates CPU gameplay capture from GPU rendering, scales beyond 64 simultaneous players, and can re-render recorded sessions under new observation configurations without replaying the game.

The authors also build a dual-tower model that predicts POV and panoramic video jointly, coupling the streams through cross-attention that aligns tokens observing the same direction in a shared world frame. A quadratic-phase player encoding offers more player codes than existing rotary codes at the same embedding dimension. Experiments show the model outperforms state-of-the-art multiplayer world models in POV prediction quality and also produces high-fidelity panoramic predictions.

The dataset and pipeline are open source. Since this is a single paper, there are no conflicting sources to compare; the claims are the authors' own.