Full RLHEV, the paper's combined human-and-engine feedback method, had the highest reported score in a bounded Unity asset-classification test. Its primary score was 0.681, with accuracy and balanced accuracy at 0.665, an F1 score of 0.733 and an AUC of 0.69.
The result came from a best-of-eight report, meaning the strongest observed performance across eight random seeds. The benchmark's primary score is a weighted composite of balanced accuracy, accuracy, F1 and AUC, so 0.681 is a combined measure rather than a single raw test.
A feedback loop built around game engines
The paper tests whether game-development traces can be used as a feedback and trajectory-data source for world-model training. RLHEV is the proposed approach: it combines dense signals from the game engine with implicit human acceptance feedback and applies that mix to AWoMo.
The evaluated AWoMo neural core is called UnifiedGameAssetModel. The released checkpoint has 2.890 billion parameters, a hidden width of 3,584, eight transformer layers, four attention heads and four gated experts.
What the Unity test actually measured
UnitySceneBench asks a narrow binary question: should an asset candidate be treated as accepted or rejected? It contains 720 training examples, 80 validation examples and 200 test examples, with the test set divided evenly between 100 accepted and 100 rejected candidates.
For this benchmark, Full RLHEV gave human reward a weight of 0.65 and engine reward 0.35. Offline RLHF used weights of 1 and 0, while Engine-based RLVR used 0 and 1; the reported setup also required engine gates. In other words, the comparison separated a human-only weighting, an engine-only weighting and a fused setting.
On the generation side, Full RLHEV reached a quality score of 0.8106 within training budgets of up to 640 instances. At the full 720-instance budget, its score was 0.8197, compared with 0.7934 for Engine-based RLVR.
The process may carry information the final file misses
The paper also compared features from a final snapshot with protocol traces, records from the development process. Spearman correlation, a measure of how closely two rankings line up, was 0.159 plus or minus 0.168 for snapshot-only features and 0.719 plus or minus 0.094 when protocol traces were included.
At a source size of 720 Unity instances, the corresponding correlations were 0.141 plus or minus 0.084 and 0.758 plus or minus 0.044. Across the source sizes reported, the trace advantage ranged from 0.56 to 0.62 correlation points.
Transfer results use a model-based judge
Generalization tests used 5,895 source examples and 1,254 target examples for a Unity distribution shift. Each cross-engine source-and-target split contained 1,000 examples, divided into 720 training, 80 validation and 200 test examples. The main outcome was a normalized score from a multimodal language-model judge, scaled from 0 to 1.
After target adaptation, the reported score was 0.75 versus 0.25 on the Unity shift. For Unity-to-Unreal transfer, the scores were 0.35 versus 0.25 against scratch, and for Unity-to-Godot they were 0.35 versus 0.15. The result is a positive transfer signal, but the measure is a judge proxy rather than a directly comparable engine-native score.
Embodied tests remain diagnostics
In separate embodied experiments, AWoMo was used as a profile-guided data-augmentation workflow, not as a standalone policy architecture. The comparisons covered the original baseline, naive augmentation and AWoMo-augmented training on R2R, Gymnasium MuJoCo and D4RL Gym-MuJoCo.
Relative to the original baseline, AWoMo-augmented training showed reported gains of 0.79% in R2R success rate, 9.96% in Gymnasium MuJoCo rollout return and 48.43% in D4RL Gym-MuJoCo normalized score.
These are diagnostic benchmark comparisons. They do not show that game-engine rewards transfer to the physical world or that the method has been tested with real scans, real robots or a genuine real-to-sim-to-real loop.
A bounded result with clear open questions
The strongest Unity number should be read in context. UnitySceneBench is an accepted-versus-rejected asset task, not a measure of complete game quality or broad human-preference alignment. The cross-engine scores also depend on the normalized model-based judge, and the classification headline number is the best observed seed rather than a seed-averaged result. No inferential uncertainty is provided for the reported comparisons.
The preprint identifies itself as arXiv version 1 dated 26 August 2026, and its abstract says agentic artifacts have been released for reproduction. The reported evidence supports a bounded computational case for combining human acceptance, engine checks and development traces, but it does not establish recursive self-improvement at production scale.
Paper data and sources
Original title: Agentic Game Development as a Verifiable Trajectory Data Engine for Scaling World Models
Authors: Pengfei Zhou, Hexin Wang, Zhengfeiyang Zhang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text