Multiverse of Greatness: Generating Story Branches with LLMs

TL;DR

It is shown that simply providing the LLM with a summary leads to a subpar story compared to additionally providing the LLM with the proper context of the story, and a consistent observation with previous studies that LLMs are biased towards certain words, even with a different LLM family.

Abstract

This paper presents Dynamic Context Prompting/Programming (DCP/P), a novel framework for interacting with LLMs to generate graph-based content with a dynamic context window history. While there is an existing study utilizing LLMs to generate a visual novel game, the previous study involved a manual process of output extraction and did not provide flexibility in generating a longer, coherent story. We evaluate DCP/P against our baseline, which does not provide context history to an LLM and only relies on the initial story data. Through objective evaluation, we show that simply providing the LLM with a summary leads to a subpar story compared to additionally providing the LLM with the proper context of the story. We also provide an extensive qualitative analysis and discussion. We qualitatively examine the quality of the objectively best-performing generated game from each approach. In addition, we examine biases in word choices and word sentiment of the generated content. We find a consistent observation with previous studies that LLMs are biased towards certain words, even with a different LLM family. Finally, we provide a comprehensive discussion on opportunities for future studies.

Figures (10)

• Figure 1: An overview of our framework showing the generation of story data (top) and a story chunk using the DCP/P approach (bottom).
• Figure 2: These figures show all nodes representing each chunk of the game story of the best-performing game generated using DCP/P. The purple node is the root node representing the story data, and the immediately connected node is the first chunk, i.e., the beginning of the story. All leaf nodes represent possible endings of the story.
• Figure 3: Screenshots of the best game generated using the baseline approach. Each screenshot is taken from various parts of the story during its progression.
• Figure 4: Screenshots of the best game generated using the DCP/P approach. Each screenshot is taken from various parts of the story during its progression.
• Figure 5: A synopsis of the best-performing story generated using the baseline approach: The Chronicles of Zephyr.