WhatELSE: Shaping Narrative Spaces at Configurable Level of Abstraction for AI-bridged Interactive Storytelling

Research Background and Issues

• Issues or Challenges: Traditional interactive narrative (IN) creation requires authors to meticulously design all possible game scenarios, a complex and time-consuming task. With the development of generative AI, particularly large language models (LLMs), it has become possible to generate content in real-time that adapts to player choices. This new approach enhances interactivity but also makes it difficult for authors to control the scope of potential narratives. Challenges include expressing abstract narrative spaces and reflecting authorial intent.
• Significance: Interactive narrative is a vital form of digital storytelling, widely applied in gaming, education, and other fields. Generative AI not only enhances player engagement but also supports dynamic story generation, surpassing traditional fixed branching designs with greater freedom. However, this approach is limited by authors' difficulty in grasping the boundaries of narrative spaces and ensuring consistency in content generation.
• Research Motivation and Related Work: Existing studies have proposed real-time content generation methods based on LLMs, but authors face limitations when expressing intent solely through abstract prompts. Additional challenges include ensuring consistency in LLM-generated content and aligning it with the causal dynamics of game mechanics. These issues highlight the need for a tool that helps authors shape narrative spaces while developing storylines.

Solution

• Method or Solution: The proposed interactive narrative creation system, WhatELSE, allows authors to import example stories to generate narrative spaces, which can then be explored and controlled through three views: Core Instance View, Outline View, and Variant View. The system leverages linguistic abstraction to edit narrative spaces and uses LLMs combined with narrative planning methods to expand narrative spaces into executable game events.
• Innovations:
  1. Introduces a new mechanism for interactive narrative creation based on multi-level linguistic abstraction to control narrative spaces.
  2. Provides a multi-perspective structure with "Core Instances," "Outline," and "Variants" to help users understand and edit narrative spaces.
  3. Develops a technical pipeline enabling bidirectional conversion between narrative instances and outlines, integrating LLMs with narrative planning to ensure causality and behavioral validation.
• Implementation Steps and Key Techniques:
  1. Users provide example stories, which the system converts into narrative spaces containing core instances.
  2. Linguistic abstraction generates narrative outlines, allowing users to adjust abstraction levels to balance detail and player freedom.
  3. Narrative variants are developed by integrating external game environments and character behavior models, dynamically adjusting story progression based on player actions.
  4. Event sequences are generated using LLM-based narrative planning methods, with causal dynamics and consistency validated in simulation environments.

Research Outcomes

• Specific Outcomes:
  1. WhatELSE enables users to flexibly create narrative spaces and edit them through linguistic abstraction.
  2. The system generates game events with high responsiveness to player behavior and consistency, providing more interactive storylines.
• Advantages Over Existing Solutions:
  1. Overcomes the limitations of traditional tools in offering customizable narrative space abstraction.
  2. Compared to simple prompt-based generation using LLMs, WhatELSE provides greater controllability and user agency.
  3. Achieves bidirectional conversion between story instances and outlines, along with dynamic event generation.
• Experimental or Evaluation Results:
  1. User studies show that participants feel a stronger sense of control and higher satisfaction when shaping and editing narrative spaces using the system.
  2. Technical evaluations demonstrate the system's ability to generate more abstract outlines and expand them into diverse storylines, while excelling in responding to player behavior.
  3. Causality in game story generation and character behavior logic were validated through external environment simulations.
• Limitations and Future Directions:
  1. Current research primarily targets simple narrative domains and novice users; future work should expand to complex storylines and professional creators.
  2. Further exploration is needed on generating branching outlines from multiple narrative instances and enabling users to intuitively perceive narrative space boundaries.
  3. Enhancements are required to address long-term consistency issues and improve interface design for easier user operation and understanding of abstract concepts.

By integrating narrative space editing, linguistic abstraction, and real-time dynamic generation, WhatELSE effectively addresses some of the bottlenecks in traditional interactive narrative creation. It provides new insights into AI-assisted content creation while promoting a better balance between authorial intent and player interaction.