Towards Enhanced Immersion and Agency for LLM-based Interactive Drama

Hongqiu Wu, Weiqi Wu, Tianyang Xu, Jiameng Zhang, Hai Zhao

Introduction

Interactive drama represents a frontier in digital entertainment, promising a fusion of cinematic storytelling and player-driven agency. While Large Language Models (LLMs) have demonstrated remarkable capabilities in role-playing and open-ended dialogue, they often struggle to maintain the structural integrity required for compelling narrative arcs. This paper introduces a novel framework designed to bridge the gap between surface-level mimicry and structured, emotionally resonant storytelling. By integrating classical narrative theory with dynamic reflection mechanisms, the proposed system enhances both player immersion and the meaningfulness of player agency.

The Agency-Immersion Paradox

A fundamental challenge in AI-driven storytelling is the "Agency-Immersion Paradox." Traditional interactive narratives often rely on rigid scripts; while these scripts ensure a high level of immersion through coherent plots and dramatic tension, they severely limit player agency—the ability to significantly influence the world. Conversely, open-ended "sandbox" interactions provide high agency but frequently suffer from narrative collapse, where the story loses its thematic focus and dramatic momentum.

The authors define Immersion as the psychological sensation of being present within a narrative world, and Agency as the player's capacity to effect change. To resolve this paradox, the framework posits that AI agents must function not just as characters, but as "playwrights" and "directors" capable of adapting the underlying plot to accommodate player behavior without sacrificing narrative structure.

Methodology: A Two-Pronged Approach

1. Playwriting-guided Generation

To move beyond the aimless generation typical of vanilla LLMs, the framework employs a structured pipeline inspired by classical dramatic theory:

  • Macro-level Structure: The system utilizes Polti’s 36 dramatic situations, sampling from eight core categories (e.g., Revenge, The Savior, Love Triangle) to establish a narrative foundation.
  • Micro-level Techniques: Specific narrative devices such as Suspense, Twists, and Irony are applied to individual scenes to maintain engagement.
  • Critic & Revise Loop: The generation process is iterative. A "Writer" LLM generates a draft, which is then evaluated by a "Critic" LLM based on the effective use of narrative techniques. The Writer then revises the content based on this feedback.
  • Refinement: Stories are expanded progressively from broad strokes to nuanced details, ensuring that the final output is both coherent and rich in texture.

2. Plot-based Reflection

To handle player agency, the system implements a Plot-based Reflection mechanism. Rather than simply reacting to the most recent user input, a "Director" agent periodically analyzes the player’s history, memories, and perceived intentions. This analysis is used to update a dynamic "plot chain." For instance, if a player attempts to extract a secret from a non-player character (NPC), the Director can insert new plot points or modify existing ones to ensure the player's choice has a tangible impact on the story's trajectory.

3. Hybrid Architecture

The framework introduces a flexible architecture that alternates between two modes:

  • One-for-All: A single agent manages multiple NPCs during narrative-heavy or expository scenes. This approach is computationally efficient and maintains high narrative consistency.
  • Director-Actor: For complex, interaction-heavy scenes, the system switches to a multi-agent setup. Here, a "Director" coordinates individual "Actor" agents, allowing for greater character depth and more nuanced responses to player actions.

Experimental Evaluation

The system was evaluated using a diverse set of data and scenarios:

  • Premise Paragraphs: 50 manually crafted story starters across genres like Crime, Fantasy, and History.
  • Interactive Script: A high-quality adaptation of Seven People In the Station (based on Detective Conan) served as the primary testbed.
  • Aggressive Agent Players: To test robustness, the researchers deployed 10 "simulated players" with challenging personalities (e.g., "Troublemaker," "Fan Girl") to determine if the system could maintain narrative coherence under pressure.

Results and Insights

The experimental results demonstrate significant improvements over standard prompting methods:

  • Narrative Quality: In human evaluations, the playwriting-guided approach significantly outperformed vanilla models. The win-rate for Suspense increased from 10% to 32%, while Emotional Tension saw a dramatic rise from 10% to 48%.
  • Agency and Influence: Plot-based reflection boosted scores for Player Influence (from 3.5 to 4.0) and Intention Following (from 3.3 to 4.0) on a 5-point scale.
  • Efficiency: The Hybrid Architecture provided a 1.49x speedup compared to a pure multi-agent model, while actually improving immersion scores.

A key insight from the study is that while LLMs are "well-read"—having processed vast amounts of literature—they are "poorly trained" as playwrights. They possess the data but lack the heuristic principles of dramatic construction. Explicitly prompting models with structural frameworks (like Aristotle’s three-act structure) proved more effective than simply increasing the volume of training data.

Limitations and Future Work

Despite its successes, the framework faces several challenges. "Script leakage" occurred when the Director agent inadvertently revealed future plot points to the player. Additionally, the reflection mechanism occasionally "hallucinated" environmental details (e.g., adding a tennis court to a train station) to satisfy player queries. Future work will focus on automating the evaluation of dramatic quality and allowing narrative techniques to adapt dynamically to different story domains.

Summary updated

BibTeX
@inproceedings{wu2025towards,
  author = {Hongqiu Wu and Weiqi Wu and Tianyang Xu and Jiameng Zhang and Hai Zhao},
  title = {{Towards Enhanced Immersion and Agency for LLM-based Interactive Drama}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure