Agentic Game Development as a Verifiable Trajectory Data Engine for Scaling World Models
The paper argues game engines can supply the verifiable reward signals spatial world models lack.
The authors say crawled video plus compute is an inefficient path for scaling world models. They propose game development as an executable environment where collision, physics, navigability, and bounded playability can be checked directly. Their RLHEV framework combines those dense engine checks with human acceptance signals from the development process. HF Daily Papers' note
The authors say crawled video plus compute is an inefficient path for scaling world models. They propose game development as an executable environment where collision, physics, navigability, and bounded playability can be checked directly. Their RLHEV framework combines those dense engine checks with human acceptance signals from the development process. HF Daily Papers' note
score 4