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

TL;DR

WhatELSE addresses the challenge of authoring controllable AI-bridged interactive narratives by shaping narrative spaces at configurable abstraction levels. It introduces a narrative space editor with three views (Pivot, Outline, Variants) and a bidirectional pipeline that transforms narrative instances into outlines and unfolds outlines into executable plots via LLM-based narrative planning grounded in a game environment. In a user study with 12 participants and a technical evaluation, the approach improved authors' perception and editing of narrative space, preserved authorial intent during gameplay, and produced more engaging, responsive game events compared with a baseline. The work demonstrates a practical pathway to balance authorial control and player agency in AI-driven storytelling and suggests integration opportunities with traditional IN tools and game engines.

Abstract

Generative AI significantly enhances player agency in interactive narratives (IN) by enabling just-in-time content generation that adapts to player actions. While delegating generation to AI makes IN more interactive, it becomes challenging for authors to control the space of possible narratives - within which the final story experienced by the player emerges from their interaction with AI. In this paper, we present WhatELSE, an AI-bridged IN authoring system that creates narrative possibility spaces from example stories. WhatELSE provides three views (narrative pivot, outline, and variants) to help authors understand the narrative space and corresponding tools leveraging linguistic abstraction to control the boundaries of the narrative space. Taking innovative LLM-based narrative planning approaches, WhatELSE further unfolds the narrative space into executable game events. Through a user study (N=12) and technical evaluations, we found that WhatELSE enables authors to perceive and edit the narrative space and generates engaging interactive narratives at play-time.

Figures (7)

• Figure 1: An illustration of the Narrative Space Editor interface, including the pivot, outline, and variants view. Users can (A) generate outline from pivot or variants with an abstraction ladder to configure the abstraction level. They can (B) fine-tune sentence or word-level abstraction using an abstraction tooltip. They can also (C) generate variants from outline specifying the number of variants in the variants view. They can use (D) narrative progression slider to visualize the variants' dynamic distance from the pivot (star).
• Figure 2: An overview of the technical pipeline of WhatELSE. (1) The system transforms narrative instances to an outline using the LLM to summarize their commonalities, generate outlines at different levels of abstraction, and review the outline based on user specifications in the Abstraction Ladder. (2) The Interactive Narrative Compiler unfolds the outline into (3) a sequence of character actions to act out the events in the outline. (4) The Game Environment executes the actions and updates the world states. (5) The player (or a simulated player) can interfere with the game by changing the world states. Finally, the Game Environment sends the updated world states and outline back to the compiler for the next iteration.
• Figure 3: An example workflow that shows (a) an author uploads a story draft in WhatELSE to (b) generate an outline. The system unfolds the outline into (c) an executable game plot with (d) a pre-loaded story domain, which supports branching storylines based on the player actions. If the player chooses to (e1) save the deer from the hunter, this action fulfills the "brave assistance" event in the outline defined by the author (shown as the orange star). If the player chooses to (e2) ask another character (e.g. a witch) for help, the witch will instead save the deer, demonstrating "brave assistance" to fulfill the event. Alternatively, if the player does not choose to save the deer at all, the system will choose a character from the story domain to save the deer as a demonstration of "brave assistance". This example shows how the game plot is dynamically adjusted based on the player actions to fulfill the outline. (f) The author can play the game plot to better understand the player experience.
• Figure 4: An illustration of the baseline system that uses a prompt-based approach for (1) the outline generation and (2) IN instantiation with (3) the pre-loaded Fairytale Forest story domain. (4) During runtime, LLM acts as a text adventure game engine, dynamically generating game plots based on the outline and responding to player input in the baseline system.
• Figure 5: Questionnaire results comparing WhatELSE with the baseline.