Why Game Development Matters for Scaling World Models

A common strategy for scaling world models is to train on more crawled video with more compute. We argue that this strategy is inefficient: scaling world models also requires a recursive data engine that offers grounded reward signals. The success of code agents illustrates why this matters. As code is executable, compilers and runtimes can provide high-quality rewards for Reinforcement Learning post-training of LLMs. By contrast, spatial generation still relies largely on fuzzy proxies such as CLIP scores and MLLM-as-judge. These signals are fuzzy and biased, making them hard to support RL post-training. We argue that game development provides a 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, scripts, budgets, and bounded playability, while the developer provides the global verification signal by judging whether the scene should be accepted. We therefore propose Reinforcement Learning with Human-Engine Verification (RLHEV), a post-training paradigm that combines dense engine signals with implicit human acceptance feedback from the development process. We apply this training objective to our proposed Agentic World Model (AWoMo): a world-building agent that proposes scene edits, observes human-engine verification, and converts accepted or repaired multimodal traces into training data. Our position is that game development provides a scalable source of executable supervision, where human spatial understanding is mapped into executable world specifications and verification signals. Our proposed roadmap is also supported by controlled experiments. On UnitySceneBench, a 200-example Unity asset-edit evaluation, our RLHEV obtains the highest score. In generalization, transfer learning helps with out-of-distribution shifts and gives positive signals in Unreal and Godot cross-engine experiments. AWoMo-augmented training also improves the embodied performance of the policy on R2R, Gymnasium MuJoCo, and D4RL Gym-MuJoCo. Agentic artifacts are released for reproduction: https://position-opensource.cardinal-agi.com.