Multi-Agent World Model

Existing world models have mainly focused on controlling the actions or viewpoint of a single agent. However, real-world simulation environments require multi-agent settings where multiple players or robots interact simultaneously.

$\gamma$-World is a generative multi-agent world model designed for such environments. Its core technology, Simplex Rotary Agent Encoding, extends 3D RoPE by representing agents as vertices of a regular simplex in rotation angle space. This gives agents an order-agnostic, scalable identity.

In addition, Sparse Hub Attention maximizes the efficiency of communication between agents. Learnable hub tokens act as mediators, reducing the attention cost with respect to the number of agents from the conventional quadratic complexity to linear complexity.

Finally, through a distillation process using bidirectional teacher models, the model achieves a real-time response speed of 24 FPS. Experimental results show that $\gamma$-World maintains high video fidelity and consistency even when scaling the number of players from 2 to 4 without additional training.