Agentic Procedural Content Generation

[Julian Togelius]

Hi all, I’d like to shamelessly advertise one of our new papers. Mostly because we’re proud of it ourselves, but also because we think this is a fertile new research direction with lots of low-hanging fruit.

We call it agentic PCG. It started with the observation that, sure, LLMs can create game levels, but they are not particularly good at it. In particular, they are not good at counting things (e.g. how many player characters or keys there are). But they can also not check if a clear path exists between two points, and obviously not do any actual playtesting. In general, they are just not very good at spatial reasoning.

However, we have code for all that. There are so many PCG methods out there, and so many game-playing agents, and so many metrics. Many basic functions are highly reusable between games, even though the actual PCG solution is typically hand-crafted for a particular game.

So, we thought, what if we chop up all our nice PCG solutions into their components, and give those components to an LLM to play with? Essentially, we pour PCG code into a toolbox so that the language model can use it as tools. We also force-feed the model all the game level metrics we can think of. Then we write some nice prompts, put it in a loop, and call it an agent. As one does these days.

This turns out to be very powerful. Through iteration, the LLMs can generate levels that fit almost any specifications you can express using your metrics. And many you can express in free-form English. You can create Mario levels that tell little stories and also are playable by certain agents and satisfy various fine-grained requirements.

We think there’s a lot of mileage in this approach, and what we’re really excited about is how to easy it is to integrate existing PCG code, and how fast it could be to create good level generators for new games by building on this old code.

Anyway, check out various videos, read the paper (if you want to) and the code (if you really want to):

https://zehua-jiang.github.io/AgenticPCG/

And let us know what you think!

Julian Togelius
Professor of Computer Science and Engineering, NYU
Head of AI, Nof1
jul...@togelius.com | http://julian.togelius.com