Game engines could be the cheap, unfoolable verifier spatial world models need

Spatial world models are stuck in an imitation loop. They train on scraped video, score outputs with CLIP or Fréchet Video Distance, and call that supervision. The central problem with this approach is that those metrics are noisy, biased, and gameable — optimizing against them can increase proxy scores while degrading actual quality. The paper's core argument is that game development already contains the missing ingredient: a dual verification channel where engines check structural correctness cheaply and developers judge whether the result actually works.

The analogy to code agents is precise and carries real weight here. Code is executable, so compilers and test suites provide dense, low-cost reward signals. Spatial generation has no equivalent — until you treat a game scene as an executable specification. A Unity, Unreal, or Godot project can be queried for collision penetration, navmesh reachability, physics stability, and bounded playability. These checks are grounded in ways that CLIP similarity is not: a collision query cannot be fooled by a plausible texture.

The proposed system, AWoMo, embeds a world model inside a developer workflow: propose an edit, execute it in the engine, collect localized failure signals, repair, and submit for human review. The human reviewer handles what the engine cannot — whether the scene fulfills design intent. Both signals are stored as structured traces under a protocol called UWDP, which links stable object identifiers across intent, intermediate states, engine outputs, rendered evidence, and accept/reject decisions. Rejected traces are kept too, supplying the negative half of the signal that final-artifact datasets discard.

The training objective, RLHEV, combines engine gates as hard constraints with human acceptance as the terminal reward. On UnitySceneBench, a 200-example Unity asset-edit evaluation, the full human-engine combination outperforms engine-only and fuzzy-proxy baselines in a best-of-eight comparison. Generalization results are more modest but directionally consistent: pretraining on Unity source data and adapting to a shifted Unity distribution raises the MLLM-as-judge score from 0.31 to 0.58; cross-engine transfer to Unreal and Godot shows smaller but positive gains. AWoMo-augmented training also improves embodied policy metrics on R2R, Gymnasium MuJoCo, and D4RL Gym-MuJoCo.

The paper is candid about what these experiments do not prove. The results are described as diagnostic pilots, not evidence of closed-loop embodied deployment or real-world transfer. The sim-to-real gap is acknowledged directly: current experiments contain no real scans, real robots, or genuine real-to-sim-to-real loops. Engine rewards can also be gamed if used in isolation, which is why the design pairs verifier ensembles with held-out checks and human review.

What makes this worth taking seriously is the structural argument, not the benchmark numbers. The claim is falsifiable — the paper states explicit falsifiers for each experimental claim — and the mechanism it proposes mirrors what already made code and reasoning scale: a feedback channel that is cheap, grounded, and hard to fool.