Towards Enhanced Immersion and Agency for LLM-based Interactive Drama

Conference: ACL 2025
arXiv: 2502.17878
Code: GitHub
Area: LLM NLP
Keywords: Interactive Drama, Immersion, Agency, Script Generation, Role-Playing Agent

TL;DR

This work proposes an Immersion-Agency paradigm to conceptualize LLM-based interactive drama, and designs two methods—Playwriting-guided Generation and Plot-based Reflection—to enhance story generation quality and player agency, respectively.

Background & Motivation

LLM-based interactive drama is an emerging genre of AI conversational application, where players role-play and interact with LLM agents representing other characters to experience an unfolding narrative. However, existing works suffer from the following limitations:

1. Lack of Theoretical Framework: Previous studies focused primarily on general architecture design, without deeply exploring the core dimensions of the interactive experience. This paper introduces two key concepts from classical interactive narrative theory: Immersion (the feeling of being absorbed in the story) and Agency (the player's ability to influence the story world).
2. Insufficient Story Generation Quality: Although LLMs are exposed to a vast amount of literary works during pre-training, fine-tuning processes lack emphasis on playwriting techniques. Consequently, the generated stories often lack fundamental dramatic structures and compelling conflicts. Experiments demonstrate that GPT-4o and Qwen2.5-72b rarely employ any narrative techniques without explicit prompting.
3. Neglect of Character Agency: Previous character agent architectures rarely considered how player actions could meaningfully influence character reactions and the narrative trajectory.

Method

Overall Architecture

The system consists of two major modules: (1) Script Generation, which utilizes Playwriting-guided Generation to generate high-quality dramatic stories (including plot structures and narrative techniques) from a premise provided by the player; (2) Character Agents, which employ Plot-based Reflection to allow NPCs to dynamically adjust the plot chain based on player actions, thereby enhancing agency.

Key Designs

1. Playwriting-guided Generation:

   • Defines 8 classic dramatic situations (such as love, phoenix rebirth, Cinderella, vengeance, etc.), described based on Aristotle’s three-act structure (setup, confrontation, resolution).
   • Summarizes 6 micro-narrative techniques (suspense, twists, non-linear narrative, multiple-narrative, irony, symbolism).
   • Generation workflow: Sample 1 dramatic situation + 3 narrative techniques \(\rightarrow\) Writer LLM generates the story \(\rightarrow\) Critic LLM evaluates and provides improvement feedback \(\rightarrow\) Writer revises \(\rightarrow\) Repeat 3 times to select the best formulation \(\rightarrow\) Progressive refinement of details.
   • Effect: Narrative technique utilization rate increases from $6\%\text}12\%$ in the baseline to $28\%\text{74\%$ (based on GPT-4o).

2. Plot-based Reflection:

   • The character agent performs a reflection every \(k=5\) interaction steps to analyze memories of player activities (emotions, intentions) and dynamically adjust the plot chain.
   • Each reflection is constrained to adjusting at most one incomplete plot point or inserting at most one new plot point, preventing incoherent narratives caused by excessive LLM modifications.
   • This enables characters to exhibit meaningful shifts in reactions driven by player behavior, such as leaking secrets, offering companionship, or advancing the plot in specific directions.

3. Hybrid Agent Architecture:

   • Director-Actor Architecture: A Director Agent coordinates globally while individual Actor Agents play their respective characters, which is suitable for high-interaction scenarios.
   • One-for-All Architecture: A single global agent plays all characters, yielding higher efficiency, suitable for narrative-centric scenarios.
   • The hybrid approach dynamically switches between these two architectures based on scenario characteristics, balancing performance and efficiency (accelerating inference by \(1.49\times\)).

Loss & Training

This work does not involve model training; all agents are built on prompt engineering using GPT-4o. The key strategies include: - A "Sampling-Critic-Revise" loop to ensure the correct application of playwriting techniques. - Progressive generation (adding details from coarse to fine). - A memory system that preserves all dialogue history within the prompt.

Key Experimental Results

Main Results

Story generation evaluation (50 premise paragraphs, human annotator win rates):

Method	Conflict (Best ↑ / Worst ↓)	Suspense	Emotional Tension	Character Arc	Technique Adherence Rate
Outline-First	18%/34%	10%/28%	10%/50%	18%/36%	-
Playwriting-Guided	32%/24%	32%/22%	48%/16%	34%/20%	92%
w/o Critic & Revise	24%/24%	26%/34%	18%/28%	12%/32%	66%
w/o Refinement	26%/18%	32%/26%	24%/6%	36%/12%	-

Character Agent evaluation (5-point scale, hand-crafted script "Seven at the Station", 10 human players + 10 agent players):

Architecture	Character Consistency	Attractiveness	Narrative Completeness	Progress	Influence	Intention Following	Speedup Ratio
Director-Actor	3.9	4.2	3.8	3.6	4.2	3.9	1.00x
Hybrid Architecture	4.1	3.9	4.3	4.3	4.0	4.0	1.49x
w/o Reflection	4.0	3.5	4.2	3.9	3.5	3.3	1.90x

Ablation Study

Configuration	Key Metric	Description
w/o Critic & Revise	Technique adherence rate \(92\% \rightarrow 66\%\)	Critic LLM is crucial for ensuring the correct application of playwriting techniques.
w/o Refinement	Emotional tension \(48\% \rightarrow 24\%\)	Progressive refinement contributes the most to emotional details.
w/o Plot-based Reflection	Influence \(4.0 \rightarrow 3.5\), Intention following \(4.0 \rightarrow 3.3\)	The reflection mechanism is the core of agency.
Pure Director-Actor	Progress 3.6 vs. Hybrid 4.3	Multi-agent communication causes information loss, affecting narrative progress.

Key Findings

1. Progressive refinement contributes the most to emotional tension (\(10\% \rightarrow 48\%\)), as emotions typically derive from subtle details in the text.
2. Plot-based Reflection not only enhances agency but also increases character attractiveness (\(3.5 \rightarrow 3.9\)), presumably because reflection encourages characters to display stronger empathy.
3. Aggressive agent players paradoxically scored higher in character attractiveness and influence dimensions, likely because agents interact more actively, and high-quality responses in turn left a deep impression on the annotators.
4. The algorithm can automatically match appropriate dramatic situations to different themes (e.g., romance \(\rightarrow\) love, crime \(\rightarrow\) deliverer/savior).

Highlights & Insights

• Theoretical Contribution: First to establish an Immersion-Agency evaluation paradigm for LLM-based interactive drama, providing an analytical framework from the perspectives of narratology and psychology.
• Systematization of Dramatic Techniques: Combining classical dramatic theories (Polti’s 36 dramatic situations, Aristotle’s three-act structure) with modern LLM prompt engineering, presenting an interesting practice of computational creativity.
• Human-Centric Evaluation: Rejecting automated LLM evaluators and insisting on evaluations conducted by annotators with training in the humanities, as the evaluation of literary works requires accuracy and empathy.
• Practicality of Hybrid Architecture: Dynamically selecting the architecture based on scenario characteristics achieves a favorable balance between efficiency and quality.

Limitations & Future Work

1. Dependency on GPT-4o: All agents rely on the same closed-source model, which is costly and uncontrollable.
2. Efficiency Issues: Playwriting-guided Generation is 10-12 times slower than vanilla prompting.
3. Reflection Boundary Control: LLMs tend to over-adjust the narrative. Currently, this is addressed via hard constraints, which may restrict more creative adaptation.
4. Solely Focused on Dialogue Formats: Exploration of scene generation (multimodal elements like visuals and music) for enhancing immersion is not conducted.
5. Limited Evaluation Scale: Tested with only 10 human players and one hand-crafted script; generalizability remains to be verified.

Related Work & Insights

• Mateas (2000)'s theory of Interactive Drama provides the original definitions of Immersion and Agency, which this work operationalizes into evaluable dimensions.
• Park et al. (2023)'s memory-based reflection focuses on memory synthesis, whereas Plot-based Reflection focuses on plot adaptation; the two are complementary, parallel techniques.
• Wu et al. (2024) first defined the six elements and plot chain mechanism of LLM interactive drama, upon which this work adds reflection and generation enhancement.
• This paper demonstrates the possibility of integrating classical dramatic theory into AI systems, providing broad insights for NPC AI, game narrative, educational simulation, and other fields.

Rating

Dimension	Score (1-5)
Novelty	4
Theoretical Depth	4
Experimental Thoroughness	3
Practical Value	4
Writing Quality	4
Overall Score	3.8

Related Papers

• [ACL 2025] Cross-Modal Alignment for LLM-Enhanced Spoken Language Understanding
• [ACL 2025] Boosting LLM's Molecular Structure Elucidation with Knowledge Enhanced Tree Search Reasoning
• [ACL 2025] CER: Confidence Enhanced Reasoning in LLMs
• [ACL 2025] Alignment Drift in CEFR-prompted LLMs for Interactive Spanish Tutoring
• [ACL 2025] Interactive and Expressive Code-Augmented Planning with Large Language Models