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

> arXiv: 2605.28816 | Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players > Authors: Fangfu Liu, Kai He, Tianchang Shen, Tianshi Cao, Sanja Fidler, Yueqi Duan, Jun Gao, Igor Gilitschenski, Zian Wang, Xuanchi Ren > Published: 2026-05-27

Why single-agent world models break down

Most current world models (e.g., Genie-style systems) assume the future is determined by a single control signal — one agent acts, everything else is background. Real scenarios rarely work that way:

• Multi-player games: every player decides independently
• Robot collaboration: warehouse and surgical robots must avoid each other
• Autonomous driving: decisions at an intersection affect all vehicles
• Virtual social spaces: every avatar acts autonomously

Existing multi-agent attempts have hard limitations:

1. Fixed agent slots — lack permutation symmetry; agents behave differently depending on their assigned position, and adding agents requires retraining. 2. All-to-all attention — fair but quadratic in cost: O(n²) compute explodes as agent count grows. 3. Predefined interaction patterns — cannot cover improvised, open-ended real-world interactions.

Gamma-World targets three core challenges: independent controllability, permutation symmetry, and efficient inference.

Simplex Rotary Agent Encoding

Gamma-World's first contribution extends Rotary Position Embeddings (RoPE) into 3D space and represents each agent as a vertex of a regular simplex:

• In 2D, an equilateral triangle (3 equidistant vertices); in 3D, a regular tetrahedron; in n dimensions, n+1 equidistant vertices.
• Each agent occupies a unique phase in rotary angle space, but all agents are mathematically equivalent — no "player one" or "player two."

This yields permutation symmetry (swapping agents changes nothing), smooth scalability from 2 to 4, 8+ agents, and the encoding is parameter-free — uniqueness and symmetry are guaranteed by geometric structure, not learned layers.

Sparse Hub Attention

Instead of all-to-all communication, Gamma-World introduces learnable hub tokens — relay nodes that aggregate and redistribute information:

• All-to-all: n agents require n(n−1)/2 pairwise exchanges.
• Sparse hub: each agent reports to a hub; hubs communicate with each other — reducing complexity from O(n²) to O(n).

Because hubs are learnable, the model discovers which information needs global coordination and which can be handled locally. A unified generation framework keeps all agents on a shared time axis, preserving temporal consistency.

Real-time inference via distillation

Diffusion world models are slow: multi-step denoising per frame makes interactive use impractical. Gamma-World uses teacher-student knowledge distillation:

1. Teacher: a full-context diffusion model that sees the entire time series and produces high-quality futures — but slowly. 2. Student: a causal model relying only on past and current context, accelerated with KV caching, generating temporal blocks at 24 FPS in real time.

24 FPS is the smoothness threshold for games and interactive applications, meaning Gamma-World can generate *interactive* multi-agent worlds, not just offline simulations.

Experimental results

Evaluated in multi-player virtual environments on video fidelity, action controllability, and inter-agent consistency, against slot-based and dense-attention baselines:

• Slot-based methods collapse beyond 2 agents — fixed encodings cannot handle dynamics.
• Dense attention is acceptable at 2 agents but computationally explosive and degraded at 4.
• Gamma-World maintains high quality at both 2 and 4 agents with controlled cost.

The headline result: Gamma-World is trained only on two-agent data yet generalizes zero-shot to four-agent scenarios, showing the simplex encoding learns a general mathematical principle of multi-agent interaction rather than memorizing specific agent counts.

Significance

Gamma-World shifts world models from "single-player" to "multi-player" generation — enabling simulation of traffic, pedestrian flows, ecosystems, and economic systems where every participant is an independent agent. Combined with 24 FPS real-time generation, it points toward general world models that understand and generate worlds of many interacting intelligences.

> Reference: Liu, F., He, K., Shen, T., et al. (2026). *Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players*. arXiv preprint arXiv:2605.28816.