Large Language Models and Games: A Survey and Roadmap

TL;DR

This survey maps how large language models intersect with games, identifying nine roles that LLMs can play—from in-game players and NPCs to designers and commentators—and provides concrete examples across text-based, board, and modern video games. It presents a taxonomy, surveys existing work, and highlights promising directions such as procedural design assistance, long-horizon memory, and streamer-facing analytics, while candidly addressing limitations like hallucinations, context loss, and cost. The paper also discusses ethical, legal, and societal issues including copyright, explainability, privacy, and bias, offering a roadmap for responsible advancement and benchmarking in this nascent field. Overall, it lays a foundation for leveraging LLM capabilities in games while outlining practical constraints and critical research priorities for researchers and industry alike.

Abstract

Recent years have seen an explosive increase in research on large language models (LLMs), and accompanying public engagement on the topic. While starting as a niche area within natural language processing, LLMs have shown remarkable potential across a broad range of applications and domains, including games. This paper surveys the current state of the art across the various applications of LLMs in and for games, and identifies the different roles LLMs can take within a game. Importantly, we discuss underexplored areas and promising directions for future uses of LLMs in games and we reconcile the potential and limitations of LLMs within the games domain. As the first comprehensive survey and roadmap at the intersection of LLMs and games, we are hopeful that this paper will serve as the basis for groundbreaking research and innovation in this exciting new field.

Figures (6)

• Figure 1: Screenshot of the promotional video for AI people, where a player can interact via their avatar's chat with other NPCs and watch how their text has consequences between NPC relationships, in this case. Used with permission from goodai2023blog.
• Figure 2: In 1001 Nightssun2023language, the player uses free-form text to trick the king (role-played by an LLM) into uttering the name of a particular weapon (which will then materialize, allowing the player to defeat him). Image used with permission.
• Figure 3: In Infinite Craft, the player combines what begins as a simple set of atomic elements into increasingly complex entities, with an LLM dictating the product resulting from arbitrary combinations. Image used with permission.
• Figure 4: A level for the puzzle game Sokoban generated by GPT-3, visualized with the Griddly tileset bamford2021griddly. Image used with permission.
• Figure 5: A screenshot of a CrawLLM game instance, generated for the theme of Ancient Egypt zammit2024crawllm. In CrawLLM, the LLM generates themes, stories, characters, and locations for a card-based dungeon crawler game while Stable Diffusion generates the visuals. Image used with permission.