Multics: hierarchical-communication world model scales multiplayer generation to 10 players

TencentARC · hf · 2026-10-14

HKU and Tencent ARC Lab present Multics, a multiplayer world model using hierarchical communication to scale shared-environment generation.

Problem: fully-connected approaches (e.g., Solaris) incur quadratic attention costs as players grow; centralized approaches (e.g., γ-World) suffer interference between unrelated events in a shared global representation.
Method: information exchange is structured across player, group, and global levels — group tokens in self-attention aggregate locally relevant player info and connect to global tokens; group-token parameters are shared across slots, and budget-constrained dynamic grouping adapts to camera trajectories.
Results: SOTA in 4- and 6-player settings with better cross-player consistency, per-player action/camera controllability, and interference robustness; generalizes from 4-player training to 6/8/10-player inference without retraining while reducing self-attention compute.

More from Multimodal