NVIDIA, in collaboration with researchers from Tsinghua University, the University of Toronto, and the Vector Institute, has published a multi-agent generative world model called Gamma-World. It breaks the bottleneck that has long limited virtual environment simulations to interactions involving only one or two players. The team has released a project page and paper, with code and weights planned for open source soon.
The model introduces a high-dimensional generalization of rotary position encoding and an information mediator token mechanism. While ensuring that multiple players can be independently controlled, it achieves, for the first time, zero-shot generalization from two-player to four-player cooperation without retraining.
The primary challenge for a multi-player world model is keeping each player’s actions independent and conflict-free. The team designed Simplex Rotary Agent Encoding, which extends the classic Rotary Position Embedding (RoPE) into a high-dimensional angular space. The new encoding gives all players completely equivalent physical symmetry, no longer relying on fixed player IDs, enabling more natural independent representation and control.
To prevent the computational cost from exploding quadratically as the number of players increases, they introduced Sparse Hub Attention. The system communicates interaction information through learnable hub tokens, compressing the attention cost between players to linear complexity.
In terms of generation speed, the team distilled a high-latency diffusion model teacher into a causal model student, and with KV cache, achieved 24 frames per second (24 FPS) real-time action response output. Evaluations in multi-player game environments show that the new model significantly outperforms traditional slot-based and dense attention networks in video realism, action controllability, and consistency among players.