Did You Get the Level You Asked For? Bridging LLMs and PCGRL for Mario Level Generation
Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment · 08.11.2026 · 22 (1) · 290-299
DOI: 10.1609/aiide.v22i1.43587 ↗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 Bruno Lemos und 1 weiteren Autor:innen.
Abstract (Originalsprache)
Designers describe levels in concrete terms, such as a flat, empty stretch or a difficult climb across gaps, but turning such requests into faithful, playable content remains challenging. Corpus-driven generators are bounded by their source patterns, whereas direct language generators must both interpret the request and place every tile. We present a controllable Super Mario Bros. generator that separates these concerns. GPT-5.2 maps a request to numeric targets for eight properties: enemies, coins, blocks, pipes, gaps, verticality, path smoothness, and platform usage. A language-independent MaskablePPO policy first assembles annotated corpus chunks and then applies validated macro-edits, reverting edits that break geometry or reachability. This two-phase design reaches targets outside corpus support, including zero enemies or gaps, while retaining corpus structure. A deterministic polish stage tightens target error when metric interactions permit and reports per-dimension residuals otherwise. The release pipeline also requires an external A* agent from the Mario AI Framework to complete each accepted level. At the medium tier, three independently trained policies outperform beam search over the same chunk space while reducing single-sample inference time; one 10M-step final-tier policy further reduces intent error. The language bridge maps requests to targets with mean error 0.053 on the normalized 0–1 scale. What the system offers is bounded structural control over a fixed eight-dimensional design vocabulary rather than open-vocabulary synthesis, and we evaluate target fidelity, corpus style, and agent-verified playability rather than designer satisfaction.
Zitieren
APA
Lemos, B., & Chaimowicz, L. (2026). Did You Get the Level You Asked For? Bridging LLMs and PCGRL for Mario Level Generation. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, 22(1), 290-299. https://doi.org/10.1609/aiide.v22i1.43587
Vancouver
Lemos B, Chaimowicz L. Did You Get the Level You Asked For? Bridging LLMs and PCGRL for Mario Level Generation. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment. 2026;22(1):290-299. doi:10.1609/aiide.v22i1.43587
BibTeX
@article{Lemos2026,
author = {Lemos, Bruno and Chaimowicz, Luiz},
title = {Did You Get the Level You Asked For? Bridging LLMs and PCGRL for Mario Level Generation},
journal = {Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment},
year = {2026},
volume = {22},
number = {1},
pages = {290--299},
doi = {10.1609/aiide.v22i1.43587}
}
EUVIMED