StoryVerse: Towards Co-authoring Dynamic Plot with LLM-based Character Simulation via Narrative Planning

Overview

• This paper explores a novel approach to automated plot generation for games using Large Language Models (LLMs) to drive the behavior of virtual characters, allowing narratives to emerge from their interactions.
• Traditional symbolic narrative planning methods are limited in scale and complexity, requiring extensive knowledge engineering work.
• The proposed workflow mediates between a writer's authorial intent and the emergent behaviors from LLM-driven character simulation, using a novel authorial structure called abstract acts.

Plain English Explanation

The paper presents a way to automatically create engaging stories for video games by having the characters interact with each other and their environment in a lifelike way. Previous approaches relied on pre-written scripts, which limited the complexity of the stories.

The new method uses advanced language models to control the characters' behaviors, allowing the story to unfold naturally as the characters respond to each other and the game world. This creates a more immersive and adaptable narrative experience that can change based on the player's actions.

However, this emergent approach makes it harder for writers to direct the plot progression. To address this, the authors introduce "abstract acts" - high-level story outlines that are then translated into specific character actions by the language model. This allows writers to maintain creative control while still taking advantage of the dynamic storytelling capabilities of the language model.

The result is a "living story" that adapts to the game environment and the player's choices, with the narrative co-created by the writer, the character simulations, and the player.

Technical Explanation

The paper proposes a novel plot creation workflow that combines a writer's authorial intent with the emergent behaviors from LLM-driven character simulation. This is achieved through the use of "abstract acts" - high-level plot outlines defined by the writer that are then transformed into concrete character action sequences by an LLM-based narrative planning process.

The workflow works as follows:

1. The writer defines abstract acts - high-level story beats or plot points.
2. An LLM-based narrative planning system takes the abstract acts and the current state of the game world, and generates a sequence of character actions that fulfill those plot points.
3. The character actions are executed in the game environment, with the characters' behaviors driven by the LLM, resulting in a dynamically unfolding narrative.

This approach allows for greater authorial control than pure emergent storytelling, while still leveraging the dynamic and adaptable nature of LLM-driven character interactions. The authors present StoryVerse as a proof-of-concept system to demonstrate this plot creation workflow.

Critical Analysis

The proposed workflow offers a promising solution to the challenge of balancing authorial intent with the emergent nature of LLM-driven narratives. By introducing the concept of "abstract acts," the authors provide a structured way for writers to guide the overall plot progression while still allowing for dynamic and adaptive storytelling at the character level.

However, the paper does not delve deeply into the technical details of the LLM-based narrative planning process or the specific challenges involved in translating abstract acts into concrete character actions. Additionally, the authors do not provide a comprehensive evaluation of the system's performance or the quality of the generated narratives.

Further research is needed to explore the scalability of this approach, as well as its ability to handle more complex and nuanced narrative structures. The integration of synchronized video storytelling techniques could also enhance the overall player experience by aligning the visual and narrative elements of the game.

Conclusion

This paper presents a novel approach to automated plot generation for games that combines a writer's authorial intent with the emergent behaviors of LLM-driven character simulations. By introducing the concept of "abstract acts," the authors have developed a workflow that allows for greater creative control while still leveraging the dynamic and adaptable nature of language models.

The proof-of-concept system, StoryVerse, demonstrates the potential of this approach to create living, player-driven narratives that adapt to the game environment and the player's choices. While further research is needed to fully realize the potential of this technology, the paper offers a compelling vision for the future of interactive storytelling in games.