GENEVA: GENErating and Visualizing branching narratives using LLMs (2311.09213v3)

Abstract: Dialogue-based Role Playing Games (RPGs) require powerful storytelling. The narratives of these may take years to write and typically involve a large creative team. In this work, we demonstrate the potential of large generative text models to assist this process. \textbf{GENEVA}, a prototype tool, generates a rich narrative graph with branching and reconverging storylines that match a high-level narrative description and constraints provided by the designer. A LLM, GPT-4, is used to generate the branching narrative and to render it in a graph format in a two-step process. We illustrate the use of GENEVA in generating new branching narratives for four well-known stories under different contextual constraints. This tool has the potential to assist in game development, simulations, and other applications with game-like properties.

• The paper presents a novel framework using GPT-4 to generate and visualize interactive branching narratives based on specified constraints.
• It employs textual prompts and graph encoding via D3JS to transform narrative text into directed acyclic graphs representing alternative story paths.
• Case studies on classic stories validate the system's effectiveness in integrating thematic twists while highlighting challenges in narrative variability.

Branching Narratives in Game Design with GENEVA

Introduction

The paper "GRIM: GRaph-based Interactive narrative visualization for gaMes" (2311.09213) investigates the application of large generative text models, specifically GPT-4, to enhance narrative creation in dialogue-based role-playing games (RPGs). Traditional RPGs demand extensive effort from creative teams to craft engaging narratives capable of providing immersive player experiences. The GENEVA system leverages the capabilities of LLMs to streamline this process by generating narrative graphs with branching storylines according to specified constraints, enabling game designers to focus on high-level narrative design.

The GRIM System

GRIM operates by constructing interactive narrative graphs that encompass branching storylines grounded in a high-level narrative description. Game designers can specify constraints such as the number of storylines, their starting and ending points, and thematic elements particular to the story setting. The system empowers designers to iteratively modify the narrative graph, expanding or contracting storylines through the addition or deletion of nodes and edges. This allows designers to dynamically refine the narrative without abandoning the core structure initially set out.

The generated graphs are visualized as directed acyclic graphs (DAGs) where each node represents a scene, and directed edges depict the alternative paths players can follow. This visualization aids designers in understanding and modifying the narrative structure effectively.

Utilization of GPT-4

The inclusion of GPT-4 is pivotal to GRIM's functionality. The system formulates specific prompts to GPT-4 to produce storylines initially in text format before transforming them into a graph visualization. Two primary steps are involved:

1. Generating Storylines: GPT-4 extracts textual descriptions of narrative sequences based on constraints, ensuring each storyline is distinct yet coherent. Storylines share common beats to maintain narrative consistency while introducing variability.
2. Graph Encoding: Subsequently, GPT-4 recodes the textual storylines into graph data, ready for visualization and further designer input using D3JS for web-based graph rendering.

A notable example is the adaptation of "Frankenstein" to a 21st-century setting, where malign ingenuity manifests as genetic engineering, transforming the tale by integrating modern scientific themes while maintaining its original narrative architecture (Figure 1).

Figure 1: Narrative graph of branching storlyines generated by GRIM for the story Frankenstein but grounded in the 21st century. Additional constraints on the graph includes one start, two endings, and four storylines.

Case Studies

Four familiar stories—Dracula, Frankenstein, Jack and the Beanstalk, and Little Red Riding Hood—served as case studies. These narratives were reimagined in contexts such as Minecraft, 21st-century urban settings, Ancient Rome, and theoretical quantum realms. GRIM's flexibility and adaptability in preserving core story elements while introducing context-specific variations were tested, revealing certain strengths and limitations:

• Strengths: GRIM adeptly integrates thematic elements from the chosen settings, enriching narratives with culturally and contextually relevant twists. It adheres to structural constraints for narrative complexity, evidenced by innovative twists in familiar stories.
• Weaknesses: Certain narrative diversify aspects require refinement. Variability among storylines is sometimes limited, which constrains imaginative breadth. Furthermore, the richness of setting integration relies significantly on the extensiveness of available cultural data in the LLM training corpus.

Implications and Future Directions

GRIM's introduction signifies a shift towards automating narrative creation processes in RPG design, suggesting its potential for broad applicability in interactive storytelling. The integration of AI tools like GRIM enhances creative freedom, allowing game designers to focus on macro-narrative elements while relying on AI to manage micro-narrative complexity.

The continual evolution of AI models promises greater sophistication in narrative generation, including deeper emotional nuance and more adaptive storyline branching benefiting from richer data sources. Future research may focus on integrating real-time player feedback mechanisms to enable adaptive narrative progression in response to gameplay dynamics.

Conclusion

GENEVA, through the GRIM prototype, showcases the utility of LLMs in generating structured, interactive narratives within digital games. By empowering designers to efficiently manage story complexity and incorporate dynamic interaction possibilities, GRIM illustrates the transformative capability of generative AI in modern game development. Building upon this foundation, future narrative systems may advance to encompass real-time generative narratives and further enrich player engagement with dynamically adaptive storylines.