Situating Large Language Models Within the Landscape of Digitial Storytelling

Abstract

Introduction

Large language Models (LLMs), such as ChatGPT, have been used to generate complex human-like text, including stories, with remarkable success. However, limitations of reliability and explainability are inherent to such systems. These limitations have also surfaced in the domain of interactive storytelling [1]. Overall, storytelling is a promising avenue for investigating these issues, as shown by ongoing research efforts in the field, that aim to merge LLMs with other, less opaque, approaches to AI [2].

Research Goal and Methods

The research project is exploratory in nature. Our primary research goal is to gain a better understanding of the field of digital storytelling and LLM’s place in it. This is accomplished through a survey of important literature, as well as hands-on familiarization with different storytelling systems.

Theoretical Framework

A crucial distinction we apply is the one between stories and storytelling. This distinction can be explicated in terms of the difference between Marr’s Type 1 and Type 2 systems [3].

A Type 1 system can be divided into distinct sub-parts relatively cleanly. Hence, Type 1 systems are algorithmic in nature; we can understand them by identifying and understanding their individual sub-parts. The production of a closed, completed story can be seen as a Type 1 system, because the story is composed of clearly identifiable sub-parts, like the structure of the story-arc, the characters that all serve some purpose, etc.

Contrarily, a Type 2 system is complex, rather than algorithmic, with components of the system interacting and affecting each other, such that disentanglement becomes impossible. Thus, in order to understand a Type 2 system, we cannot solely rely on analysis, but need to consider the entire system holistically. We propose that the ongoing activity of telling a story, for example in live improvisation or role-play games, be best viewed as a Type 2 system, because its constitutive parts, like the developing psychological situation between participants, cannot be meaningfully disentangled.

This theoretical framework should enable us to interpret LLM’s success in story-generation and their simultaneous challenges in the field in an informative manner.

References

[1] J. Kelly, M. Mateas and N. Wardrip-Fruin, "Towards Computational Support with Language Models for TTRPG Game Masters," in FDG '23: Proceedings of the 18th International Conference on the Foundations of Digital Games, no. 78, April 2023. doi:10.1145/3582437.3587202

[2] J. Sung Park, J. C. O'brien, C. J. Cai, M. R. Morries, P. S. Liang and M. S. Bernstein, "Generative Agents: Interactive Simulacra of Human Behavior," arXiv:2304.03442, April 2023. doi:10.48550/arXiv.2304.03442

[3] D. Marr, Vision, A Computational Investigation into the Human Representation and Processing of Visual Information, San Francisco: W.H. Freeman and Company, 1982.