Multics launches generative multi‑player world model with hierarchical communication
| Source: HF Papers | Original article
Researchers introduce Multics, a generative multi‑player world model that uses hierarchical communication to overcome the quadratic attention costs of fully‑connected approaches.
A new research project called **Multics** has been unveiled, offering a generative multi‑player world model that structures communication hierarchically rather than relying on fully‑connected attention. The framework, described in a GitHub repository maintained by researchers from the University of Hong Kong and Tencent’s ARC Lab, organises information exchange at three levels – individual player, group, and global – to curb the quadratic cost that plagues existing approaches such as Solaris.
The development tackles a core obstacle in interactive video generation: scaling robustly as the number of participants grows. By limiting each player’s direct attention to a subset of peers and aggregating signals through intermediate group nodes, Multics promises more efficient handling of shared environments while preserving the fidelity of each participant’s perspective.
The significance of this work lies in its potential to broaden the applicability of AI‑driven simulation in gaming, virtual collaboration, and training scenarios. Current multi‑player models often hit a performance wall when the player count rises, restricting real‑time interactivity. Multics’ hierarchical scheme could enable richer, larger‑scale sessions without the prohibitive compute overhead that has limited prior systems.
The research team has made a static project page available, but no software package or build instructions have been released yet. The next steps to watch include the publication of quantitative benchmarks, possible open‑source code releases, and integration tests with existing interactive platforms. If the hierarchical communication proves effective in practice, it may set a new standard for scalable, AI‑generated multi‑player worlds.
Sources
Back to AIPULSEN