StatePlay: State-Aware Game World Models for Mechanics-Consistent Generation
StatePlay adds explicit game-state prediction so generated rollouts obey mechanics, not just visuals.
The paper argues that game world models can look plausible while breaking rules tied to health, skill meters, timers, and termination states. StatePlay jointly predicts frames and internal game states, using a mixture-of-transformers setup to keep visual and state branches specialized while letting them interact. The authors report average normalized L1 state-prediction distance below 0.06 and an 18.6% mechanics-fidelity gain over models without explicit state modeling. HF Daily Papers' note
The paper argues that game world models can look plausible while breaking rules tied to health, skill meters, timers, and termination states. StatePlay jointly predicts frames and internal game states, using a mixture-of-transformers setup to keep visual and state branches specialized while letting them interact. The authors report average normalized L1 state-prediction distance below 0.06 and an 18.6% mechanics-fidelity gain over models without explicit state modeling. HF Daily Papers' note
score 5