ReactiveGWM: Steering NPC in Reactive Game World Models

Model Overview

ReactiveGWM (Reactive Game World Models) is an interactive game world modeling framework developed by researchers at National University of Singapore (NUS) in collaboration with Tencent, HKPolyU, and HKUST-GZ.

It addresses a key limitation in current world models by decoupling player action controls from Non-Player Character (NPC) autonomy. Using an additive bias for player actions and cross-attention interaction modules for high-level NPC strategies (e.g. Offense, Defense, Control), ReactiveGWM enables steerable, prompt-aligned NPC behavior and zero-shot strategy transfer across different games without domain retraining.

---

Key Features

• Decoupled Control Architecture: Separates player action inputs from NPC behavior control, eliminating static/passive NPC rendering.
• Strategy-Grounded Cross-Attention: Dedicated cross-attention interaction modules ground high-level NPC intents (Offense, Control, Defense) directly into generated video frames.
• Zero-Shot Strategy Transfer: Interaction modules encode game-agnostic logic, allowing them to plug directly into unannotated game world models.
• Real-Time Steerable Video Generation: Maintains precise player action responsiveness while generating reactive NPC behaviors frame-by-frame.
• Multi-Game Causal Forcing Benchmarks: Evaluated on Street Fighter II: Champion Edition and Street Fighter III.

---

Verified Project Links

• Project Website: https://inv-wzq.github.io/ReactiveGWM/
• arXiv Paper: https://arxiv.org/abs/2605.15256
• GitHub Repository: https://github.com/INV-WZQ/ReactiveGWM
• Hugging Face Model: https://huggingface.co/INV-WZQ/ReactiveGWM-Models

---

Performance & Evaluation

• Evaluated across FVD, PSNR, SSIM, LPIPS, Player Action Accuracy, and NPC Strategy Alignment Score on Street Fighter II and Street Fighter III.