Generating Converging Narratives for Games with Large Language Models

Published: 02 Apr 2024, Last Modified: 04 Apr 2024COLING LREC 2024EveryoneRevisionsCC BY-NC 4.0
Abstract: Human authors make numerous choices in crafting narratives. In interactive stories like the $\it{Choose Your Own Adventure}$ series, authors must decide when and how readers will influence the plot. Authoring tools that leverage large language models (LLMs) to assist authors currently can generate multiple diverse story paths, but provide no way for rejoining these. Here we explore extending the use of LLMs for bringing separate story lines back together. We test various methods of combining the next-token probability distributions of two distinct story lines into a single distribution, and present samples of the resulting texts. Our working hypothesis was that the LLM would seek to “unify” content whenever possible. We found that the probability consolidation functions mattered less than one might expect and that the method was capable of rejoining narratives in a natural way for a wide variety of differences between the two incoming texts.
Loading

OpenReview is a long-term project to advance science through improved peer review with legal nonprofit status. We gratefully acknowledge the support of the OpenReview Sponsors. © 2025 OpenReview