Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players

Code | Paper | Website

Gamma-World (γ-World) is a generative multi-agent world model that renders a single shared world in which several independently-controlled agents stay causally connected. It supports action-responsive streaming rollouts and generalizes from two-player training to four-player (and beyond) interaction without additional training. Gamma-World is built on the NVIDIA Cosmos stack (Cosmos-Predict2.5).

Citation

@article{gammaworld2026,
 title={Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players},
 author={Fangfu Liu and Kai He and Tianchang Shen and Tianshi Cao and
 Sanja Fidler and Yueqi Duan and Jun Gao and Igor Gilitschenski and
 Zian Wang and Xuanchi Ren},
 journal={arXiv preprint arXiv:2605.28816},
 year={2026}
}

Base model