World Models Cannot Scale on Crawled Video Alone: Game Engines Supply the Missing Reward

The standard recipe for scaling world models is more crawled video plus more compute. A paper submitted to arXiv on August 26 challenges that consensus directly: the strategy is inefficient, and scaling world models also requires a recursive data engine that offers grounded reward signals (arXiv:2608.25518). The paper topped the Hugging Face Daily Papers list on August 28 with 136 upvotes. It was written by eight authors including Pengfei Zhou and Yang You, with the National University of Singapore shown as the affiliated institution on the Hugging Face page.

The Root Problem: Spatial Generation Lacks a Compiler

The starting point is the success of code agents. Code can support RL post-training of LLMs precisely because it is executable: compilers and runtimes provide high-quality rewards.

Spatial generation takes a different path. The authors point out that the field still relies largely on fuzzy proxies such as CLIP scores, and these signals are fuzzy and biased, making them hard to support RL post-training. In other words, image and video models have long been trained without a hard referee — rewards come from whether something looks right, not whether it is physically valid.

The Fix: A Game Engine Is an Executable World Specification

The core claim is that game development provides the missing reward environment for spatial world models. A scene encoded by a game engine is an executable world specification: the engine can efficiently check collision, physics, navigability and bounded playability, while the developer provides the global verification signal by judging whether the scene should be accepted.

The path also yields a critical byproduct: game development produces real-world long-horizon trajectory data, exactly what RL post-training needs. Reward environment and training data arrive together in one pipeline.

RLHEV: Combining Engine Signals with Human Acceptance

On this basis the paper proposes RLHEV (Reinforcement Learning with Human-Engine Verification), a post-training paradigm that combines dense engine signals with implicit human acceptance feedback from the development process. Local fact-checking goes to the engine; the global judgment of whether a level is any good goes to humans. Rewards no longer depend on fuzzy similarity scores.

Why It Is Worth Watching

Beyond the specific method, the more notable signal here is a methodology transfer: verifiable rewards, proven effective on code agents, are being carried into the spatial intelligence race, and the game engine is the ready-made verifier. If this path works, "crawl more video + stack more compute" stops being the standard answer to scaling world models, and the combination of a data engine with executable verification may rewrite the next generation of scaling narratives. For teams working on world models, game AI or spatial intelligence, this is a paper worth reading closely.