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

Fangfu Liu 1,2∗ Kai He 1,3,4∗ Tianchang Shen 1 Tianshi Cao 1 Sanja Fidler 1,3,4 Yueqi Duan 2 Jun Gao 1 Igor Gilitschenski 3,4† Zian Wang 1† Xuanchi Ren 1† 1 NVIDIA 2 Tsinghua University 3 University of Toronto 4 Vector Institute · 2026

Read Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players by Fangfu Liu 1,2∗ Kai He 1,3,4∗ Tianchang Shen 1 Tianshi Cao 1 Sanja Fidler 1,3,4 Yueqi Duan 2 Jun Gao 1 Igor Gilitschenski 3,4† Zian Wang 1† Xuanchi Ren 1† 1 NVIDIA 2 Tsinghua University 3 University of Toronto 4 Vector Institute (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-28816.txt.

Description

Fangfu Liu and nine colleagues present Gamma-World, a generative multi-agent world model. Most world models cover single agents, but many worlds need several players or robots at once. Simplex Rotary Agent Encoding, a tuning-free spin on RoPE, puts agents at simplex corners. It gives each agent a distinct phase, keeping all agents swap-equal. Sparse Hub Attention uses hub tokens to mediate token traffic, cutting cross-agent cost from squared to linear. A taught student generates blocks in order with KV caching at 24 FPS.

Catalog record

CREATORFangfu Liu 1,2∗ Kai He 1,3,4∗ Tianchang Shen 1 Tianshi Cao 1 Sanja Fidler 1,3,4 Yueqi Duan 2 Jun Gao 1 Igor Gilitschenski 3,4† Zian Wang 1† Xuanchi Ren 1† 1 NVIDIA 2 Tsinghua University 3 University of Toronto 4 Vector Institute
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE50119
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.28816

Browse the full library · About the Tower of Babel