NVIDIA Launches Gamma-World, a Multi-Agent Model Supporting 4-Player Collaboration at 24 FPS

ME AI News: According to monitoring by Beating, researchers from NVIDIA, Tsinghua University, the University of Toronto, and the Vector Institute have introduced Gamma-World, a multi-agent generative world model that breaks the long-standing bottleneck in virtual environment simulation, which was previously limited to single or dual-player interactions. The team has now released the project page and paper, with code and weights planned for open-sourcing soon. The model introduces two key mechanisms—high-dimensional generalization of rotary positional encoding and informational mediator tokens—enabling independent control of multiple players while achieving zero-shot generalization from two-player to four-player collaboration without retraining. The primary challenge in multi-player world models lies in ensuring each player maintains independent control without conflicting actions. The research team designed Simplex Rotary Agent Encoding, extending the classic Rotary Positional Encoding (RoPE) into a high-dimensional angular space. This new encoding grants all players full physical symmetry, eliminating reliance on fixed player IDs and enabling more natural individual referencing and control. To prevent computational costs from quadratically increasing with more players, the solution introduces Sparse Hub Attention, which uses learnable hub tokens to mediate interaction information, reducing the attention computation cost between players to linear scale. In terms of generation speed, the team distilled a high-latency diffusion model teacher into a causal student model, combined with KV Cache, achieving real-time action response output at 24 FPS. 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 inter-player consistency. (Source: BlockBeats)