Prompt-Guided Level Generation Shyam Sudhakaran 1 , Miguel González-Duque 1 , Claire Glanois 1 , Matthias Freiberger 1 , Elias Najarro 1 , Sebastian Risi 1,2 1 IT University of Copenhagen 2 modl.ai, Copenhagen Denmark {shsu,migd,clgl,matfr,enaj,sebr}@itu.dk ABSTRACT Automated generation of complex and diverse environments can be achieved through the use of Procedural Content Generation (PCG) algorithms. However, generating content that is both meaningful and refective of specifc intentions and constraints remains a chal- lenge. Recent advances in Large Language Models (LLMs) have demonstrated their efectiveness in various domains. These mod- els can be fne-tuned and information can be reused to accelerate training for new tasks. Our study presents MarioGPT, a fne-tuned GPT2 model that has been trained to generate tile-based game levels for Super Mario Bros. The results demonstrate that MarioGPT can generate diverse levels and can be text-prompted for controllable level generation, addressing a critical challenge in current PCG techniques. CCS CONCEPTS · Computing methodologies → Artifcial intelligence. ACM Reference Format: Shyam Sudhakaran 1 , Miguel González-Duque 1 , Claire Glanois 1 ,, Matthias Freiberger 1 , Elias Najarro 1 , Sebastian Risi 1,2 . 2023. Prompt-Guided Level Generation. In Genetic and Evolutionary Computation Conference Companion (GECCO ’23 Companion), July 15ś19, 2023, Lisbon, Portugal. ACM, New York, NY, USA, 4 pages. https://doi.org/10.1145/3583133.3590656 1 INTRODUCTION Procedural Content Generation (PCG) algorithms allow the auto- mated creation of game content, such as levels, maps, or charac- ters [19]. Adding PCG to a game can lead to reduced production costs and an increase in the game’s replayability. Recently, PCG researchers have started to incorporate machine learning-based approaches into their systems. One such example is the use of Gen- erative Adversarial Networks (GANs) [7] that can be trained to generate levels for games as diverse as Doom [5] or Super Mario Bros, using a level from the Video Game Level Corpus [26]. A drawback of current approaches that combine PCG with Ma- chine Learning (a feld now referred to as Procedural Content Gener- ation via Machine Learning (PCGML) [22]) is that they often rely on costly searching inside of the latent space of the underlying neural Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for proft or commercial advantage and that copies bear this notice and the full citation on the frst page. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s). GECCO ’23 Companion, July 15ś19, 2023, Lisbon, Portugal © 2023 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-0120-7/23/07. https://doi.org/10.1145/3583133.3590656 (a) "many pipes, many enemies, little blocks, low elevation" (b) "little pipes, little enemies, many blocks, high elevation" (c) "many pipes, some enemies" (d) "no pipes, no enemies, many blocks" Figure 1: Prompt-conditioned generations from a single seed block. MarioGPT is able to create diverse levels solely based on a text prompt in natural language. networks. Ideally, we would like to directly condition a generator to create levels with certain properties in natural language. The presented paper addresses this challenge, introducing Mar- ioGPT (Figure 2), a fne-tuned GPT-2 model trained to generate Mario levels. The model shows that LLMs can be combined with PCG techniques to efectively create new and diverse levels through natural language prompts. Showing the efectiveness of the method, a surprisingly high percentage (88%) of MarioGPT generated levels are in fact playable. 2 BACKGROUND AND RELATED WORK 2.1 Procedural Content Generation Procedural Content Generation (PCG) algorithms [19] deals with the automatic creation of game content (e.g. for level design, charac- ter generation, environment modeling, etc.). As reviewed in [19, 27], earlier works often focused on evolutionary computation [2], solver- based methods [20] or constructive generation methods (such as cellular automata, grammar-based methods), etc. More recently, research on deep learning for PCG [12, 22] has proposed a promis- ing approach to generate high-quality game content śnot only aesthetically pleasing but also functional and challengingś in a data-driven manner, while reducing the manual efort required from game developers. Diversity, originality and playability of the 179