Yang You et al. pitch game engines as world-model reward layer

Scaling world models by training on more crawled video is inefficient, argues a position paper posted to arXiv on 26 August by Pengfei Zhou, Yang You and six collaborators, because video carries no grounded reward signal that reinforcement learning can post-train against.

The parallel the paper draws is to code agents. "As code is executable, compilers and runtimes can provide high-quality rewards for Reinforcement Learning (RL) post-training of LLMs," the authors write. "By contrast, spatial generation still relies largely on fuzzy proxies such as CLIP scores." Those proxies, the paper argues, are "fuzzy and biased," too weak to anchor RL.

Game engines, the authors contend, are the missing analogue on the spatial side. "A scene encoded by a game engine is an executable world specification," they write, whose runtime "can efficiently check collision, physics, navigability and bounded playability," while the developer supplies a "global verification signal by judging whether the scene should be accepted."

On that framing they propose Reinforcement Learning with Human-Engine Verification (RLHEV), a post-training recipe that mixes dense engine rewards with implicit acceptance feedback from the game-development process itself.

The submission is a position paper. It publishes no benchmark numbers and does not name a specific engine or game the authors have used to prototype the recipe.