ReactiveGWM brings steerable NPCs to video world models

Interactive video world models look incredible in short clips, but the moment you try to play them, the illusion falls apart. The characters on screen aren't really fighting back. They are just background pixels passively drifting across the canvas.

Researchers open-sourced ReactiveGWM, a reactive game world model that finally gives non-player characters a functional brain. Instead of treating the whole screen as a single diffusion soup, it lets you steer the opponent with high-level prompt strategies while you control the hero with traditional keystrokes.

Keep in mind that there is no game engine running underneath. Every frame is pure AI-generated video synthesized on the fly.

Splitting player controls from NPC tactics

Most existing world models view games strictly through a single, player-centric lens. That turns the model into an interactive video player rather than a physics-aware simulation engine. When you throw a punch, the model does not truly understand that the opponent ought to block.

ReactiveGWM solves this by explicitly decoupling player inputs from NPC actions. Keystrokes are injected into the diffusion backbone via a lightweight additive bias, keeping player movement snappy and responsive.

Meanwhile, high-level NPC strategies, such as Offense, Control, or Defense, are fed in through separate cross-attention modules. Evaluated on two Street Fighter titles, players can execute fine-grained combos while the opponent reliably shifts its posture based on the text prompt you gave it.

Plug-and-play interactive logic

What makes the system especially neat is how modular it is.

The cross-attention modules learn game-agnostic interaction rules rather than memorizing Street Fighter sprites. This unlocks zero-shot strategy transfer: you can take the trained NPC module and drop it straight into off-the-shelf, unannotated world models of different games without touching the base weights or retraining.

The project builds on top of the Wan 2.2 video foundation model. The code and setup instructions are already up on GitHub, and the researchers note that the setup is lightweight enough to run locally on consumer-grade, upper-mid-tier GPUs.

Why it matters

For generative gaming, visual fidelity has stopped being the main blocker. The real wall has always been causal logic and agency. A world model where you wander around alone is just a fancy tech demo. ReactiveGWM takes the first serious stab at multi-agent physical interaction within an end-to-end video generator.

We think decoupling the inputs is an exceptionally clever architecture choice. Instead of training an enormous, fragile end-to-end model that tries to learn player controls and enemy AI simultaneously, breaking them into lightweight bias layers and prompt-driven attention modules keeps things steerable and cheap to run.

Still, realistic expectations are in order. Prompt-driven behaviors like defense or attack are still coarse-grained guidelines, not deep competitive tactical thinking. Frame generation latency also means this won't replace a real fighting game anytime soon. But as a proof of concept for engine-less, steerable worlds, it moves the ball forward significantly.