human-engine-verification

This paper proposes using game engines as a verifiable trajectory data engine for scaling world models, introducing RLHEV to combine dense engine signals with human feedback for reinforcement learning post-training.