Gamma World: Generative Multi-Agent Real-Time World Model

Model Overview

Gamma-World ($\gamma$-World) is a generative multi-agent world model developed by NVIDIA Spatial Intelligence Lab (SIL) and NVIDIA Toronto AI Lab (Fangfu Liu, Kai He, Sanja Fidler, et al.).

Unlike traditional single-agent world models, Gamma-World simulates complex multi-agent environments where multiple independently controlled agents act simultaneously and interact within a shared, spatio-temporally consistent visual space. It supports real-time video rollouts at 24 FPS and generalizes zero-shot from 2-player training to 4+ players.

---

Key Features

• Simplex Rotary Agent Encoding (SRAE): Maps agents to regular simplex vertices in RoPE angle space, ensuring permutation-equivalent identities without learned embeddings.
• Sparse Hub Attention: Routes cross-agent communication through learnable hub tokens, reducing attention complexity from $O(N^2)$ to $O(N)$.
• Teacher-Student Distillation & KV-Caching: Enables 24 FPS real-time interactive rollouts with low action latency.
• Zero-Shot Multi-Agent Generalization: Scales zero-shot from 2-player training dynamics to 4+ player simulations.
• Action Controllability & Cross-View Consistency: Preserves temporal continuity and inter-agent visual consistency across long rollout horizons.

---

Verified Project Links

• Project Website: https://research.nvidia.com/labs/sil/projects/gamma-world/
• arXiv Paper: https://arxiv.org/abs/2605.28816
• GitHub Repository: https://github.com/nv-tlabs/Gamma-World
• Hugging Face Model: https://huggingface.co/chijw/Gamma-World

---

Performance Benchmarks

• FVD (Memory Protocol): 184.1 (vs. 333.8 for Solaris baseline).
• FID (Memory Protocol): 24.8 (vs. 51.7 for Solaris baseline).
• Real-Time Speed: 24 FPS interactive streaming frame rate.