Multics Scales Multi-Player World Models with Hierarchical Communication

Introduction

A world model for multiple players must do more than generate several videos at once. Each view should describe the same shared environment, while individual players still need independent actions and camera control. As the number of players grows, two problems become increasingly difficult: communication becomes expensive, and unrelated events can contaminate one another through a shared representation.

Multics, proposed by Tencent ARC Lab, addresses this tension with a hierarchical communication architecture. Instead of allowing every player to interact directly with every other player, it structures information exchange across player, group, and global levels.

Key ideas

• Group tokens as an intermediate layer. Group tokens collect information from locally related players. They provide a compact representation of a local interaction or visual context before information is sent to a broader part of the model.
• Hierarchical information flow. Player features communicate within groups, while group tokens connect to global tokens for cross-group exchange. This topology limits unnecessary direct mixing between unrelated events.
• Shared parameters across groups. The parameters used by group tokens are shared across group slots. As a result, the same communication module can handle different player arrangements without adding group-specific parameters.
• Budget-constrained dynamic grouping. Recent camera trajectories are used to identify players that are likely to be relevant to one another. The grouping process is also constrained by a communication budget, allowing the model to avoid maintaining expensive connections that are unlikely to matter.

Results and implications

According to the paper, Multics achieves strong performance in four- and six-player settings, with improvements in cross-player consistency, player-specific action and camera controllability, and robustness to interference. The model can also be trained with four players and used for inference with six, eight, or ten players without retraining. Its self-attention computation is lower than that of fully connected communication schemes.

The broader contribution is architectural rather than merely incremental. A centralized representation can make unrelated local events compete for the same capacity, while an all-to-all design scales poorly as the player count increases. Multics treats these issues as a communication-structure problem: local information is exchanged densely where it is useful, and broader coordination is routed through a smaller number of shared representations.

The available material does not establish how well the method will perform in substantially more complex scenes, longer rollouts, or much larger player populations. It also remains important to test whether camera trajectories are a sufficient signal for grouping in interactions that are visually distant but semantically connected. Even so, Multics offers a practical direction for making multi-agent video world models more scalable without discarding player-specific control.

Source: Hugging Face Daily Papers