Building a Game with LLM Randomness: From Play to Personas

BSc Thesis.

This project builds on a previous master’s thesis, Prompting Luck: Bias and Manipulability in LLM-Generated Randomness, that explored replacing traditional random number generators (RNG) in games with large language models (LLMs). The idea is that an LLM can produce “intelligent” random numbers, for example the outcome of a slot machine spin, that take context into account rather than being drawn from a uniform distribution. The goal of this project is to experimentally validate the findings of that thesis by building an actual game, connecting it to an LLM, and putting the idea to the test with real players.

Research questions include:

• Can the idea be put into practice? Is it technically feasible to run a real game on LLM-generated randomness, and what are the trade-offs (e.g., latency, cost, stability of outcomes)?
• Does it improve gameplay? Do LLM-generated random numbers make the game more fun or engaging than traditional RNG (e.g., by prolonging player interaction)?
• How much context should the LLM receive? What is the effect of feeding it differently sized contexts, from no context up to rich histories of previous outcomes and player actions?
• Can personas be deduced from player behavior? The previous thesis went from persona to outcome (given a player persona, have the LLM generate suitable random states); here we turn that around and ask whether the way a player behaves lets us reliably infer which persona they belong to.

Project goals include:

• Building one or two playable games with LLM-generated randomness, where the game can be played online and switch between traditional RNG and LLM-generated outcomes.
• Feeding the LLM differently sized contexts and measuring the effect on feasibility, gameplay, and player experience.
• Analyzing game data and player behavior to deduce personas, and evaluating this against the personas used in the previous thesis.

This project is ideal for bachelor students interested in building and analyzing a functional game. It offers opportunities to explore both the technical and the user-experience aspects of integrating LLMs into software.