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

World models for interactive video generation have largely focused on single-agent settings, where future observations are generated from a single control signal. However, many generated environments require multi-agent interaction: multiple players, robots, or embodied agents act simultaneously within a shared space. Scaling world models to such settings requires a principled multi-agent design: agents should remain independently controllable, permutation-symmetric, and support efficient inference while maintaining consistency across time and perspectives. In this paper, we present our generative multi-agent world model for interactive simulation. It introduces Simplex Rotary Agent Encoding, a parameter-free extension of 3D RoPE that represents agents as vertices of a regular simplex in rotary angle space. This gives each agent a distinct phase while making all agents permutation-equivalent, enabling scalable agent identity without learned per-slot identities or a fixed agent ordering. To avoid dense all-to-all attention across agents, we further propose Sparse Hub Attention, where learnable hub tokens mediate token interaction across agents, reducing cross-agent attention cost from quadratic to linear in the number of agents. For real-time rollout, we distill a full-context diffusion teacher into a causal student that generates temporal blocks sequentially with KV caching, enabling action-responsive generation at 24 FPS. Experiments in multiplayer virtual environments show that our model improves video fidelity, action controllability, and inter-agent consistency over slot-based and dense-attention baselines, while generalizing from two to four players without additional training.

📋 한 줄 요약

**[Multi-Agent World Model / Video]** Gamma-World가 Simplex Rotary Agent Encoding과 Sparse Hub Attention으로 다중 에이전트 video world model 구성 — agent 수 선형 확장, 24 FPS 인터랙티브 생성, 2→4 player 추가 학습 없이 일반화.

🎯 핵심 기여도

인터랙티브 video generation의 world model이 주로 single-agent 환경에 집중 — single control signal에서 future observation 생성에 머무름.
다수 생성 환경이 multi-agent interaction을 요구함을 지적 — 여러 플레이어·로봇·embodied agent가 공유 공간에서 simultaneously 동작.
이런 환경으로 world model을 확장하려면 principled multi-agent 설계 필요 — agent independently controllable·permutation-symmetric·효율 추론·시간·관점 간 일관성 유지.
Gamma-World — generative multi-agent world model 제안. Simplex Rotary Agent Encoding(3D RoPE의 parameter-free 확장)으로 agent를 rotary angle 공간 regular simplex의 vertex로 표현 — agent에 distinct phase 부여하면서 permutation-equivalence 유지, learned per-slot identity·고정 ordering 없이 scalable; Sparse Hub Attention으로 학습 가능 hub 토큰이 agent 간 토큰 상호작용 매개해 cross-agent attention 비용을 agent 수에 quadratic→linear; real-time rollout 위해 full-context diffusion teacher를 causal student로 distill, temporal block sequential 생성·KV caching으로 24 FPS action-responsive generation; multiplayer virtual environment 실험에서 video fidelity·action controllability·inter-agent consistency를 slot 기반·dense-attention baseline 대비 향상, 추가 학습 없이 2→4 player 일반화.

💡 핵심 아이디어

다중 에이전트 video world model의 확장은 (1) agent identity의 parameter-free·permutation-equivalent 인코딩(Simplex Rotary), (2) cross-agent attention의 linear 축소(Sparse Hub), (3) full-context teacher→causal student distillation의 결합으로 scalability·real-time·일관성을 동시 달성한다.

🔬 기술적 접근법

**방법론**: Gamma-World — multi-agent video world model.
**핵심 기법**: (1) Simplex Rotary Agent Encoding — 3D RoPE의 parameter-free 확장, agent를 simplex vertex로, (2) Sparse Hub Attention — 학습 가능 hub 토큰이 cross-agent 토큰 상호작용 매개, attention 비용 linear, (3) Full-context diffusion teacher → causal student distillation, (4) Temporal block sequential 생성 + KV caching으로 24 FPS, (5) Action-responsive generation.

📊 주요 결과

24 FPS action-responsive generation 달성.
Multiplayer virtual environment에서 video fidelity·action controllability·inter-agent consistency 향상(vs slot 기반·dense-attention baseline).
2→4 player로 추가 학습 없이 일반화.
Agent 수 증가에 attention 비용 linear 유지.

💭 의의 및 한계

**의의**: Multi-agent video world model의 첫 principled 설계 — identity·attention·distillation 3축 동시 해결, real-time interactivity 달성, agent 수 일반화의 design-level 보장. **한계**: 가상 환경 검증으로 실제 로봇·복잡 물리 일반화 추가 검증, 더 많은 agent(예: 10+)에서의 video fidelity, causal student의 long-horizon drift 가능성.

🚀 실용적 활용

멀티플레이어 게임·시뮬레이션의 real-time world model.
로봇·embodied agent multi-agent 학습 환경.
인터랙티브 콘텐츠 생성·메타버스.