Learning Compositional and Transferable Mechanics in Generative World Models
Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment · 08.11.2026 · 22 (1) · 12-22
DOI: 10.1609/aiide.v22i1.43560 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, online erschienen am 08.11.2026. Verfasst von Shinichi Arimura und 5 weiteren Autor:innen.
Abstract (Originalsprache)
Generative world models can now simulate playable games frame by frame, yet it is unclear whether they learn game mechanics as reusable, manipulable objects or merely memorize game-specific transition statistics. We study this by training a rule-conditioned latent diffusion world model on 2D, fully observable General Video Game AI (GVGAI) transitions collected under controlled atomic rule interventions. We encode rules as shared embeddings that can be switched on with a multi-hot vector. This lets us test composition, where unseen rule combinations are activated together, and transfer, where a rule is applied to a game that never contained that rule in training. Our results demonstrate that rule composition can emerge from limited single-rule training data, even without synthetically generated rule-combination examples. While all tested rules transfer to a visually similar target, transfer to a visually distant target varies, with some rules transferring fully, some weakly, and some not at all.
Zitieren
APA
Arimura, S., Basnet, S. J., Gandhi, P. S., Simpson, W., Merino, T., & Togelius, J. (2026). Learning Compositional and Transferable Mechanics in Generative World Models. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, 22(1), 12-22. https://doi.org/10.1609/aiide.v22i1.43560
Vancouver
Arimura S, Basnet SJ, Gandhi PS, Simpson W, Merino T, Togelius J. Learning Compositional and Transferable Mechanics in Generative World Models. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment. 2026;22(1):12-22. doi:10.1609/aiide.v22i1.43560
BibTeX
@article{Arimura2026,
author = {Arimura, Shinichi and Basnet, Soyuj Jung and Gandhi, Pranit Singh and Simpson, Wes and Merino, Timothy and Togelius, Julian},
title = {Learning Compositional and Transferable Mechanics in Generative World Models},
journal = {Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment},
year = {2026},
volume = {22},
number = {1},
pages = {12--22},
doi = {10.1609/aiide.v22i1.43560}
}
EUVIMED