From Role-Play to Drama-Interaction: An LLM Solution

TL;DR

This work offers a cohesive framework for LLM-based interactive drama that treats the drama as a complete narrative system rather than single-character role-play. It introduces Narrative Chain to decompose plot progression into manageable sub-narratives, Auto-Drama to automatically generate rich drama scripts from arbitrary stories, and Sparse Instruction Tuning to improve following long, multi-task instructions. A global drama script guides the LLM-driven performance, while a five-dimension evaluation (scenery, narration, transition, guidance, coherency) provides a multi-faceted assessment of drama LLMs. Empirical results show strong performance for 8B models trained with Auto-Drama and SIT, with limited gains from larger models, and case studies illustrating tone control, scene transitions, and character variation. The approach enables scalable, interactive storytelling with practical implications for education, entertainment, and AI-driven narrative design, while acknowledging modality and evaluation limitations and outlining avenues for future work.

Abstract

Drama is a form of storytelling inspired by human creativity, proceeding with a predefined storyline, carrying emotions and thoughts. This paper introduces \emph{LLM-based interactive drama}, which endows traditional drama with an unprecedented immersion, where a person is allowed to walk into it and interact with the characters and scenes. We define this new artistic genre by 6 essential elements-plot, character, thought, diction, spectacle and interaction-and study the entire pipeline to forge a backbone \emph{drama LLM} to drive the playing process, which is challenged by limited drama resources, uncontrollable narrative development, and complicated instruction following. We propose \emph{Narrative Chain} to offer finer control over the narrative progression during interaction with players; \emph{Auto-Drama} to synthesize drama scripts given arbitrary stories; \emph{Sparse Instruction Tuning} to allow the model to follow sophisticated instructions. We manually craft 3 scripts, \emph{Detective Conan}, \emph{Harry Potter}, \emph{Romeo and Juliet}, and design a 5-dimension principle to evaluate the drama LLM comprehensively.

Figures (7)

• Figure 1: Illustration of LLM-based interactive drama. Two scenes from a script for Detective Conan (Library Murder Case, Episode 50), where the player takes the role of Conan and there are other four characters, Ayumi, Mitsuhiko, Genta, and Librarian. The player can either make conversations with characters or take actions to progress the plot.
• Figure 2: Prototype of drama scripts to prompt drama LLMs. The identifier at the beginning of each input either designates a character for dialogue or indicates that the input corresponds to an action outlined in the script.
• Figure 3: Story arcs defined by the drama script and experienced by the player across different methodologies.
• Figure 4: The pipeline of Auto-Drama for drama script generation.
• Figure 5: Diagram of Sparse Instruction Tuning.