Abstract
Drama is a form of storytelling inspired by human creativity, proceeding with a predefined storyline while carrying deep emotions and thoughts. This paper introduces LLM-based interactive drama, a new artistic genre that endows traditional drama with unprecedented immersion by allowing players to interact directly with characters and scenes. We define this genre through six essential elements—plot, character, thought, diction, spectacle, and interaction—and develop a comprehensive pipeline to create a "drama LLM." To overcome challenges such as limited drama resources and uncontrollable narrative development, we propose Auto-Drama for script synthesis, the Narrative Chain for plot control, and Sparse Instruction Tuning (SIT) for sophisticated instruction following. Our evaluation across three iconic scripts—Detective Conan, Harry Potter, and Romeo and Juliet—demonstrates the model's ability to maintain narrative coherence while providing a high degree of player agency.
Introduction
While Large Language Models (LLMs) have excelled at role-playing individual characters, they often struggle to manage the broader scope of a "drama." In standard role-play, the model simulates a single persona. In an interactive drama, the model must simultaneously act as the stage manager and the entire cast, managing the "Spectacle" (the environment), multiple "Characters" (NPCs), and the "Plot" (the narrative arc).
This transition from role-play to drama-interaction requires the model to balance player freedom with a structured storyline. We ground our framework in Aristotelian principles, replacing the traditional element of "music" with "interaction" to suit the digital age. To realize this, we address three primary hurdles: data scarcity, narrative control, and the complexity of following massive, multi-faceted instructions.
The Drama LLM Pipeline
1. Auto-Drama: Synthesizing Scripts
High-quality, structured interactive drama scripts are rare. To bridge this gap, we developed the Auto-Drama pipeline. Using GPT-3.5 as a base, we transform standard prose stories (such as Andersen’s Fairy Tales) into structured drama scripts. The pipeline follows four key steps:
- Scene Extraction: Identifying distinct locations and timeframes.
- Character and Environment Detailing: Defining the traits of NPCs and the specifics of the setting.
- Plot Flow Production: Mapping out the logical progression of the story.
- Trigger Imagination: Designing how specific player actions or dialogue choices lead to different narrative outcomes.
2. Narrative Chain: The Story's GPS
To prevent the story from stalling or veering entirely off-course, we introduce the Narrative Chain. This mechanism breaks a scene’s narrative into a sequence of smaller, manageable segments. Rather than "railroading" the player into a single path, the Narrative Chain acts like a GPS; it understands the current state of the story and provides the necessary dialogue or environmental cues to gently guide the player toward the next intended plot beat, ensuring the story reaches a satisfying conclusion without breaking immersion.
3. Sparse Instruction Tuning (SIT)
A drama script is essentially a massive instruction. However, at any given moment during gameplay, only a tiny fraction of that instruction—such as a specific character's reaction or a scene transition—is relevant. Standard fine-tuning often fails to handle this "sparsity."
Sparse Instruction Tuning (SIT) addresses this through a two-stage training process:
- Stage 1: The model learns individual sub-tasks, such as role-playing or describing scenery.
- Stage 2: The model learns to associate specific parts of a full drama script with its output using special annotations (e.g.,
/* trigger */). This "anchors" the model's attention to the relevant section of a long system prompt, significantly improving its ability to follow complex, multi-layered instructions.
Experimental Setup and Results
Datasets and Models
We trained our models on a dataset generated from over 80 stories from Andersen’s Fairy Tales. For evaluation, we manually crafted three high-quality test scripts based on popular intellectual properties: Detective Conan (mystery), Harry Potter (adventure), and Romeo and Juliet (classical drama). We fine-tuned LLaMA3-8B-Instruct and Qwen1.5-14B-Chat using the SIT method.
Performance Metrics
Models were evaluated across five dimensions: Scenery, Narration, Transition, Guidance, and Coherency.
Key findings include:
- The Necessity of SIT: Without Sparse Instruction Tuning, models struggled significantly with "Transitions"—the ability to recognize when a player's action should trigger a jump to a new scene.
- Data Efficiency: While role-play and narration skills emerged with only 10% of the Auto-Drama data, the full dataset was required to master "Guidance" (steering the player back to the plot).
- Scaling Observations: Interestingly, scaling from 8B to 14B parameters did not yield massive gains. This suggests that for interactive drama, the quality of instruction following and the structure of the training data are more critical than raw parameter count.
Conclusion and Future Work
The transition from role-play to interactive drama represents a significant step forward in digital storytelling. By utilizing the Auto-Drama pipeline and Sparse Instruction Tuning, we have created a backbone LLM capable of orchestrating complex narrative experiences.
Future work will focus on expanding this framework beyond text. Integrating images, sound, and video would enhance immersion, while more complex "world-state" logic would allow for even more open-ended gameplay. Additionally, as these drama engines become more sophisticated, we aim to develop more robust, human-centric evaluation methods to better capture the nuances of player experience.
Summary updated