[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-80571":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":10,"languages":10,"totalLinesOfCode":10,"stars":11,"forks":12,"watchers":13,"openIssues":14,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":16,"stars7d":17,"stars30d":18,"stars90d":15,"forks30d":15,"starsTrendScore":19,"compositeScore":20,"rankGlobal":10,"rankLanguage":10,"license":10,"archived":21,"fork":21,"defaultBranch":22,"hasWiki":21,"hasPages":21,"topics":23,"createdAt":10,"pushedAt":10,"updatedAt":29,"readmeContent":30,"aiSummary":31,"trendingCount":15,"starSnapshotCount":15,"syncStatus":32,"lastSyncTime":33,"discoverSource":34},80571,"Gamma-World","nv-tlabs\u002FGamma-World","nv-tlabs","Implementation of Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players","https:\u002F\u002Fresearch.nvidia.com\u002Flabs\u002Fsil\u002Fprojects\u002Fgamma-world\u002F",null,604,7,9,1,0,3,19,456,14,6.71,false,"main",[24,25,26,27,28],"aigc","multi-agent","robotics","video-game","worldmodel","2026-06-12 02:04:04","\u003Cdiv align=\"center\">\n\n\u003Cimg src=\"assets\u002Fnvidia-logo.png\" width=\"260\" alt=\"NVIDIA\">\n\n\u003Ch1>\n    ✨Gamma-World: Generative Multi-Agent World Modeling\u003Cbr>\n    Beyond Two Players✨\n\u003C\u002Fh1>\n\n\u003Cp align=\"center\">\n    \u003Ca href=\"https:\u002F\u002Fliuff19.github.io\u002F\">Fangfu Liu\u003C\u002Fa>\u003Csup>1,2*\u003C\u002Fsup>    \n    \u003Ca href=\"https:\u002F\u002Fwww.cs.toronto.edu\u002F~hekai\u002F\">Kai He\u003C\u002Fa>\u003Csup>1,3,4*\u003C\u002Fsup>    \n    \u003Ca href=\"https:\u002F\u002Fwww.cs.toronto.edu\u002F~shenti11\u002F\">Tianchang Shen\u003C\u002Fa>\u003Csup>1\u003C\u002Fsup>    \n    \u003Ca href=\"https:\u002F\u002Fwww.linkedin.com\u002Fin\u002Ftianshi-cao-a23270b1\u002F\">Tianshi Cao\u003C\u002Fa>\u003Csup>1\u003C\u002Fsup>    \n    \u003Ca href=\"https:\u002F\u002Fwww.cs.utoronto.ca\u002F~fidler\u002F\">Sanja Fidler\u003C\u002Fa>\u003Csup>1,3,4\u003C\u002Fsup>    \n    \u003Ca href=\"https:\u002F\u002Fduanyueqi.github.io\u002F\">Yueqi Duan\u003C\u002Fa>\u003Csup>2\u003C\u002Fsup>    \n    \u003Ca href=\"https:\u002F\u002Fwww.cs.utoronto.ca\u002F~jungao\u002F\">Jun Gao\u003C\u002Fa>\u003Csup>1\u003C\u002Fsup>\n    \u003Cbr>\n    \u003Ca href=\"https:\u002F\u002Fdiscover.research.utoronto.ca\u002F32914-igor-gilitschenski\">Igor Gilitschenski\u003C\u002Fa>\u003Csup>3,4†\u003C\u002Fsup>    \n    \u003Ca href=\"https:\u002F\u002Fwww.cs.utoronto.ca\u002F~zianwang\u002F\">Zian Wang\u003C\u002Fa>\u003Csup>1†\u003C\u002Fsup>    \n    \u003Ca href=\"https:\u002F\u002Fxuanchiren.com\u002F\">Xuanchi Ren\u003C\u002Fa>\u003Csup>1†\u003C\u002Fsup>\n    \u003Cbr>\n    \u003Cbr>\n    \u003Csup>1\u003C\u002Fsup>NVIDIA    \n    \u003Csup>2\u003C\u002Fsup>Tsinghua University    \n    \u003Csup>3\u003C\u002Fsup>University of Toronto    \n    \u003Csup>4\u003C\u002Fsup>Vector Institute\n\u003C\u002Fp>\n\n\u003Ca href=\"https:\u002F\u002Fresearch.nvidia.com\u002Flabs\u002Fsil\u002Fprojects\u002Fgamma-world\u002F\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProject-Page-Green\">\u003C\u002Fa>     \n\u003Ca href=\"assets\u002Fgamma-world.pdf\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPaper-PDF-b31b1b.svg\">\u003C\u002Fa>     \n\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.28816\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2605.28816-b31b1b.svg\">\u003C\u002Fa>     \n\u003Ca>\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-Apache--2.0-blue\">\u003C\u002Fa>\n\n![γ-World teaser](assets\u002Fteaser.png)\n\n\u003C\u002Fdiv>\n\n\u003Cstrong>γ-World:\u003C\u002Fstrong> We introduce a generative multi-agent world model that rolls out a single shared environment for multiple independently controllable agents. γ-World supports permutation-symmetric agent conditioning with \u003Cstrong>Simplex Rotary Agent Encoding\u003C\u002Fstrong>, efficient cross-agent communication with \u003Cstrong>Sparse Hub Attention\u003C\u002Fstrong>, real-time \u003Cstrong>24 FPS\u003C\u002Fstrong> streaming with a distilled block-causal student, and zero-shot generalization from two to four players.\n\n## 📢 News\n\n- 🚀[05\u002F28\u002F2026] We release γ-World with the project page, paper, videos, qualitative results, and method overview.\n- 🔜[Coming Soon] We will release the code and distilled streaming checkpoints with KV cache support.\n- ⏳[Planned] Training scripts and dataset preparation tools will be released in a future update.\n\n## 🌟 Overview\n\nγ-World interactively generates coherent future frames from multi-agent actions while preserving shared-world consistency, scaling from multiplayer virtual games to real-world multi-robot environments.\n\nhttps:\u002F\u002Fgithub.com\u002Fuser-attachments\u002Fassets\u002F11a81855-5b51-4117-bfcd-ef07246e0a4e\n\n## 📖 Abstract\n\nWorld 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.\n\nWe present \u003Cstrong>γ-World\u003C\u002Fstrong>, a generative multi-agent world model for interactive simulation. γ-World introduces \u003Cstrong>Simplex Rotary Agent Encoding\u003C\u002Fstrong>, 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.\n\nTo avoid dense all-to-all attention across agents, we further propose \u003Cstrong>Sparse Hub Attention\u003C\u002Fstrong>, 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 \u003Cstrong>24 FPS\u003C\u002Fstrong>.\n\n## 🖼️ Gallery\n\n### Two-Agent Interaction\n\nEach agent is independently controllable while sharing the same evolving world.\n\n![Two-agent interaction](assets\u002Ftwo-agent-interaction.png)\n\n### Four-Agent Generalization\n\nBenefiting from the permutation-symmetric simplex agent encoding, γ-World generalizes from two to four players \u003Cstrong>without additional training\u003C\u002Fstrong>.\n\n![Four-agent generalization](assets\u002Ffour-agent-generalization.png)\n\n### Real-World Robotics Coordination\n\nγ-World extends to real-world multi-robot coordination scenarios, demonstrating applicability beyond virtual environments.\n\n![Real-world robotics coordination](assets\u002Frobotics-coordination.png)\n\n## 🧠 Method\n\n![Method overview](assets\u002Fmethod-overview.png)\n\nγ-World takes synchronized observations and actions from multiple agents as input, tokenizes each agent stream with shared visual and action encoders, and generates future multi-agent rollouts with a causal multi-agent DiT. The model formulates the input with an explicit synchronized agent axis, encodes exchangeable agent identity using Simplex Rotary Agent Encoding, and routes cross-agent information through Sparse Hub Attention. During streaming inference, the causal student uses KV caches for past visual tokens and hub states to preserve block-causal generation while scaling efficiently with the number of agents.\n\n### Simplex Rotary Agent Encoding\n\nSimplex Rotary Agent Encoding is a parameter-free extension of 3D RoPE. Instead of assigning agents scalar indices or learned identity vectors, γ-World places them at the vertices of a regular simplex in rotary angle space. All agents have equal pairwise distances, so every pair is permutation-equivalent while each agent retains a distinct rotary phase.\n\n### Sparse Hub Attention\n\nSparse Hub Attention routes cross-agent communication through a small set of learnable hub tokens. Agent tokens attend to their own stream and to the hubs; the hubs aggregate information across agents and broadcast it back. This preserves a shared communication pathway without dense pairwise interaction, reducing cross-agent attention from \u003Cstrong>O(N²)\u003C\u002Fstrong> to \u003Cstrong>O(N)\u003C\u002Fstrong>.\n\n### Efficiency\n\n![Sparse Hub Attention timing](assets\u002Fsparse-hub-timing.png)\n\nSparse Hub Attention achieves significantly lower latency and FLOPs as the number of agents increases, making γ-World more scalable beyond two players.\n\n## 📚 Citation\n\nIf you find γ-World useful for your research or applications, please cite our paper:\n\n```bibtex\n@article{gammaworld2026,\n    title={Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players},\n    author={Fangfu Liu and Kai He and Tianchang Shen and Tianshi Cao and\n            Sanja Fidler and Yueqi Duan and Jun Gao and Igor Gilitschenski and\n            Zian Wang and Xuanchi Ren},\n    journal={arXiv preprint arXiv:2605.28816},\n    year={2026}\n}\n```\n\n## Acknowledgements\n\nThe authors would like to thank Product Managers Aditya Mahajan and Matt Cragun for their valuable support and guidance, Jingnan Gao for proof discussion, and Yixin Hong for demo creation.\n\n## License\n\nγ-World will be released under the Apache License 2.0. Final license terms will be confirmed at the code release.\n","Gamma-World 是一个生成式多智能体世界模型，能够为多个独立可控的智能体提供单一共享环境。该项目通过Simplex Rotary Agent Encoding技术实现智能体条件的置换对称性，并利用Sparse Hub Attention机制支持高效的跨智能体通信。此外，它还能够在24帧每秒的速度下实现实时流媒体传输，并且可以从两名玩家零样本泛化到四名玩家。这些特性使得Gamma-World非常适合应用于需要多智能体交互模拟的场景，如视频游戏开发、机器人协作研究以及复杂系统建模等领域。",2,"2026-06-11 04:01:15","CREATED_QUERY"]