NVIDIA · 2026-05-27 · notable
NVIDIA's Gamma-World — Generative Multi-Agent World Model Scales Beyond Two Players With Simplex Rotary Agent Encoding and Sparse Hub Attention, Streams Shared Worlds at 24 FPS
NVIDIA's new generative world model represents agents as a simplex via 'Simplex Rotary Agent Encoding' and replaces quadratic cross-agent attention with linear 'Sparse Hub Attention,' streaming a shared interactive world at 24 FPS and zero-shot generalizing from 2 to 4 players.
NVIDIA's Gamma-World scales generative world models past the two-player ceiling with linear-cost agent attention and 24 FPS real-time rollouts.
Key specs
| Fps | 24 |
|---|---|
| Players scaled to | 4 |
| Arxiv id | 2605.28816 |
| Hf paper upvotes | 132 |
What is it?
Gamma-World is a generative video world model from NVIDIA's Spatial Intelligence Lab that simulates a shared interactive environment with multiple independently controllable agents. Authors include Sanja Fidler, Jun Gao, and Igor Gilitschenski; the paper opened the week at #1 on Hugging Face's trending papers list.
How does it work?
Two architectural moves get past the 2-player limit. Simplex Rotary Agent Encoding represents each agent as a distinct vertex of a simplex so the network stays permutation-equivariant across any agent count. Sparse Hub Attention funnels agent-to-agent communication through a shared hub, dropping cost from quadratic to linear in agent count. The full model is distilled into a causal student with KV-cache support that streams at 24 FPS and zero-shot generalizes from 2 to 4 players without retraining.
Why does it matter?
Generative world models have been stuck at single-agent or two-player settings — Gamma-World is the first credible scaling demo into four-agent shared environments at interactive frame rates. The same techniques target multi-robot real-world simulation, not just multiplayer games.
Who is it for?
Researchers in world models, embodied AI, and multi-agent simulation.
Try it
github.com/nv-tlabs/Gamma-World (code coming soon)