ReactiveGWM

ReactiveGWM is a research game world model from authors at Tencent and several universities (National University of Singapore, HK PolyU, HKUST and others). Existing game world models usually bind player and NPC roles to fixed characters. ReactiveGWM instead assigns control modes at initialization and jointly simulates externally controlled players and reactive NPCs in the same scene.

It uses two components. Spatial Role Binding grounds learned character handles in initial-frame regions using instance masks, which also lets it distinguish characters with identical appearance. Unified Agency Conditioning encodes player action sequences and conditional NPC rules (a trigger plus a response) into control-token groups, each bound to a character handle through cross-attention. Causal self-attention over past latent frames lets the model interpret NPC rules as interactions unfold, and training uses flow matching over future video latents. A YouTube summary says it is built on Wan 2.2 and that code is released, though no repository link appears in the provided material.

The project is studied in two shared-view 2D games, Street Fighter III and Hikou no mizu, with demos such as an NPC anti-airing a jump-in or parrying a projectile. It is a research release with a project page and demos; scope is limited to those games as presented.