WorldMind: Decoupled Game World Model for State-Aware NPC Behavior
WorldMind makes NPC decisions a separate step before rendering the next gameplay frames.
The paper splits game world modeling into understanding, decision, control, and generation layers. Its goal is to ground NPC behavior in the evolving game state instead of leaving it buried inside video generation or driven by outside control signals. The authors also introduce BOSS-140K, a gameplay dataset paired with internal game states. In experiments, WorldMind was preferred over baselines in about 70% of pairwise comparisons for more tactically appropriate and coherent NPC behavior. HF Daily Papers' note
The paper splits game world modeling into understanding, decision, control, and generation layers. Its goal is to ground NPC behavior in the evolving game state instead of leaving it buried inside video generation or driven by outside control signals. The authors also introduce BOSS-140K, a gameplay dataset paired with internal game states. In experiments, WorldMind was preferred over baselines in about 70% of pairwise comparisons for more tactically appropriate and coherent NPC behavior. HF Daily Papers' note
score 5