Abstract
Large language models have demonstrated their capabilities in storyline creation and human-like character role-playing. Current language model agents mainly focus on reasonable behaviors from the level of individuals, and their behaviors might be hard to constraint on the level of the whole storyline. In this paper we introduce IBSEN, a director-actor coordinate agent framework that generates drama scripts and makes the plot played by agents more controllable. The director agent writes plot outlines that the user desires to see, instructs the actor agents to role-play their characters, and reschedules the plot when human players participate in the scenario to ensure the plot is progressing towards the objective. To evaluate the framework, we create a novel drama plot that involves several actor agents and check the interactions between them under the instruction of the director agent. Evaluation results show that our framework could generate complete, diverse drama scripts from only a rough outline of plot objectives, meanwhile maintaining the characteristics of characters in the drama.
Introduction
While Large Language Models (LLMs) excel at individual role-playing, multi-agent simulations often suffer from "narrative drift." Interactions between distributed agents may be realistic at a local level but frequently lack a global purpose or a cohesive plot. This challenge is particularly acute in interactive scenarios where human player intervention can derail the intended storyline. IBSEN (Interactive-Based playhouSe for agENts) addresses this by introducing a centralized "Director" to regulate distributed "Actors," ensuring narrative coherence and objective completion.
The IBSEN Framework
IBSEN employs a hierarchical architecture that separates narrative management from character execution. This division of labor allows for both global plot control and local character autonomy.
The Director Agent
The Director acts as the narrative's central controller, following a continuous loop to drive the story forward:
- Outline Generation: Creates a high-level plot continuation based on predefined objectives.
- Script Translation: Converts the outline into a rough dialogue script to determine the flow of conversation.
- Actor Instruction: Provides high-level instructions (keywords and synopses) to Actors rather than exact lines. This "soft control" preserves the Actor's unique voice.
- Objective Checking: Evaluates whether the current plot goal has been met before proceeding to the next objective.
The Actor Agent
Actors focus on believable role-playing through specialized memory management:
- Divided Memory: Memory is partitioned into an Actor Profile (static traits), a Memory Database (past events), and a Character Database (relationships).
- First-Person Monologues: Memories are stored as first-person reflections, allowing the LLM to interpret events through the character's emotional lens, which significantly improves consistency.
Interactive Rescheduling
A key feature of IBSEN is its ability to handle human players. When a human player takes an unexpected action, the Director immediately regenerates the storyline and script. This dynamic rescheduling incorporates the player's influence while steering the simulation back toward the original narrative objectives.
Evaluation and Insights
The framework was tested using Hedda Gai, a novel interactive drama adapted from Henrik Ibsen’s Hedda Gabler. The scenario consisted of 7 acts with 14 plot objectives.
Key Results
- Objective Completion: Most objectives were met within a 9-turn limit, with less than 10% requiring forced completion.
- Plausibility: The system achieved high scores for storyline coherence (3.63/4) and logicality (2.96/4) as evaluated by ChatGPT.
- Director Accuracy: The Director’s objective-checking capability achieved an F1 score of 0.77 compared to human judgment.
Narrative Challenges
The study highlighted the "Positivity Bias" hurdle: standard LLM safety alignments often make agents "too nice." In dramas requiring manipulative or flawed characters, agents sometimes struggled to maintain these traits, tending toward more helpful or upright behavior.
Conclusion
IBSEN demonstrates that purposeful drama requires a centralized regulator to coordinate distributed agents. By using an "instruction buffer" and monologue-style memory retrieval, the framework achieves a balance between narrative control and character believability. Future work will focus on integrating IBSEN into game engines like Unity or RPG Maker to provide visual interfaces and more complex environmental interactions.
Summary updated