Multics: a multi-player generative world model with hierarchical communication
A paper titled "Multics: Generative Multi-Player World Model with Hierarchical Communication" proposes a multi-player video generation framework that structures information exchange across player, group and global levels. It introduces group tokens into self-attention as intermediate representations aggregating locally relevant player information, with those tokens connecting to global tokens for cross-group communication, and shares group-token parameters across group slots so the same module can handle varying player counts. The authors also propose budget-constrained dynamic grouping based on recent camera trajectories, and report state-of-the-art results in 4- and 6-player settings with improved cross-player consistency, action and camera controllability, and interference robustness, plus generalization from 4-player training to 6-, 8- and 10-player inference without retraining and reduced self-attention computation versus fully-connected communication.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.17201 (cs)
[Submitted on 12 Oct 2026]
Title:Multics: Generative Multi-Player World Model with Hierarchical Communication
Authors:Kaiyi Huang, Wangbo Yu, Yukun Huang, Jiwen Yu, Kunhao Liu, Donglin Yang, Wenbo Hu, Ying Shan, Xihui Liu
View a PDF of the paper titled Multics: Generative Multi-Player World Model with Hierarchical Communication, by Kaiyi Huang and 8 other authors