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

Researchers introduced Gamma-World, a generative world model for interactive video that supports multiple agents, not just two players.

It uses simplex rotary agent encoding and sparse hub attention to preserve permutation symmetry and scale agent interaction efficiently.

A diffusion teacher is distilled into a causal student, enabling real-time rollout at 24 FPS with action controllability.

The work pushes interactive video generation closer to practical multi-agent simulation in multiplayer virtual environments.