Available at: https://digitalcommons.calpoly.edu/theses/3377
Date of Award
6-2026
Degree Name
MS in Computer Science
Department/Program
Computer Science
College
College of Engineering
Advisor
Rodrigo Canaan
Advisor Department
Computer Science
Advisor College
College of Engineering
Abstract
Procedural Content Generation (PCG) has had much research poured into it in recent years, as it is a way to more easily create content and levels for player replayability. Historically, this type of level generation has been done by algorithms with strict rules and black-and-white decision making. However, in recent years Large Language Models (LLMs) have been fine-tuned for this task in research in games such as Super Mario Bros and other 2D platformers. In this thesis, we apply a similar approach to level generation for a vertical platformer called Kid Icarus. We create a tile-based level generation system that allows for varying both tile densities and gameplay properties associated with the solution path.
To do this, we fine-tune a GPT-2 Medium LLM on a set of six annotated Kid Icarus levels from the Video Game Level Corpus (VGLC). Our per-tile encoding system allows us to get consistent level tile density which is in line with user expectations from their prompting. Levels are generated in segments from the bottom of the level up to the top, using a seeded player spawn section at the bottom.
Rather than enforcing playability through generation, we use a repair algorithm to go through the level after generation is completed to make small adjustments. These changes are verified by game physics and close any gap in the solution path to the top of the level. We finalize our system by allowing users to set parameters on difficulty, punishability, and route branching which give them the ability to control the feel of the solution path for the level, rather than solely being based on tile generation densities