Agentic Game Development as a Verifiable Trajectory Data Engine for Scaling World Models
Pengfei Zhou, Hexin Wang, Zhengfeiyang Zhang, Yixing Ma, Zhenglin Wan, Kaipeng Zhang, Wangbo Zhao, Yang You
cs.AI
2026-08-26
NUS trains world models with engine checks plus human acceptance. Full RLHEV hits 0.681 primary on UnitySceneBench, +0.098 over the best non-full baseline.
The default recipe for scaling world models is more crawled video and more compute. Code agents post-train because compilers and tests are cheap, reproducible rewards. Spatial generation still scores CLIP, FVD, and MLLM-as-judge, which are noisy, biased, and gameable. Human preference over final frames is too expensive and too low-bandwidth to support the loop that made coding models scale.
The paper calls this a verifiability bottleneck: the missing piece is a feedback channel that can grade structural right and wrong. Game engines already compute collision, physical stability, navmesh reachability, script errors, and whether a bounded playtest probe hits a goal. A developer still decides whether the level should ship. That split matches tests-plus-review in software.
RLHEV splits the reward. Engine checks are dense structural terms. Human accept/reject is sparse and owns qualification. The full model mixes 0.65 human and 0.35 engine. Hard gates must pass; softer diagnostics enter as penalties.
AWoMo is a world model sitting inside a developer workflow: propose a scene edit, let the engine render and localize failures, repair, stop on human review. Every loop is stored as a Unified World-Development Protocol trace: brief, object id, state, action, engine output, rendered evidence, reviewer decision, residual risk. The same object supports understanding (asset classification) and generation (scene programs).
Rewards climb a ladder: loads, physically plausible, navigable with an achievable objective, then playable. The paper treats this as a falsifiable agenda: verifiable reward should beat fuzzy proxies; human-plus-engine should beat engine-only; source traces should transfer under distribution shift and across engines.
UnitySceneBench is a 200-example Unity asset-edit eval. Learned methods train on 720 instances across eight seeds. Full RLHEV’s best-of-eight primary score is 0.681, with 0.665 accuracy, 0.665 balanced accuracy, 0.733 F1, and 0.690 AUC, +0.098 primary and +0.120 accuracy over the strongest non-full baseline. On the generation split it reaches 0.8106 at a 640-instance budget and 0.8197 at 720, versus 0.7934 for engine-only RLVR. Primary is 0.45 balanced accuracy + 0.25 accuracy + 0.20 F1 + 0.10 AUC.
For in-family OOD, pretrain on source then adapt on target lifts a judge score from 0.25 to 0.75. Cross-engine gains are smaller: Unity→Unreal 0.25→0.35, Unity→Godot 0.15→0.35. There is no engine-native scalar shared across Unity, Unreal, and Godot, so the comparison uses a human-audited MLLM rubric.
In embodied diagnostics AWoMo is not a policy architecture. It is profile-guided environment-data augmentation. Relative to the original baseline: +0.79% R2R success, +9.96% Gymnasium MuJoCo return, +48.43% D4RL Gym-MuJoCo normalized score.
Replacing “scrape more video” with “find a compiler for spatial intelligence” is the claim to track. Development traces record intent, failure, repair, and acceptance; a finished asset does not. If post-training can actually eat that dual channel, world models get a path that resembles code models.
The numbers are pilots, not a spinning data engine. Human-engine verification is strongest on Unity classification and generation. Cross-engine and R2R gains are small. The contribution is first an agenda and a protocol. Recursive self-improvement is not yet shown.
The authors list the obvious objections. Games are not reality; there are no real scans, real robots, or a sim-to-real loop. An engine is a partial verifier and a loose collision check can be gamed. Source-engine traces can overfit one runtime. The studies do not establish full game quality, human-subject validity, or embodied deployment.
Read the metrics with a discount. The headline Unity result is best-of-eight; seed means live in a separate figure. OOD and cross-engine scores use MLLM-as-judge, the same family of proxy the paper rejects as training reward, even with human audit. R2R +0.79% is near noise. D4RL’s +48.43% is large inside a diagnostic setup. The bench is 200 test items and 720 training items. A loop where one agent builds levels and another plays them is future work.