Exploring Creator-Centric Methods for LLM-Assisted Interactive Storytelling

Paper Title

Exploring Creator-Centric Methods for LLM-Assisted Interactive Storytelling

Publication Info

• Topic area: Interactive storytelling with AI assistance
• Keywords: Large language models, interactive storytelling, creator-centered design, narrative tools, node-graph editing, moral complexity, user simulation, co-creation, ripple-effect analysis, human-AI collaboration

Background and Problem

• Problem / challenge: Existing tools for interactive storytelling fail to adequately address creators' needs for managing complex narrative structures, maintaining thematic coherence, and balancing AI assistance with authorial control. Current LLM-based systems often prioritize generation efficiency over creator workflows, leading to misalignment with creative processes.
• Significance: Addressing these gaps is crucial for enabling creators to produce immersive, multi-layered narratives while preserving their creative sovereignty. Improved tools can reduce cognitive load, enhance narrative quality, and support morally complex storytelling.
• Motivation and related work: Prior research has focused on system-oriented perspectives, emphasizing technical capabilities and user behavior modeling, but lacks empirical validation of creators' workflows and pain points. Existing tools often fail to integrate structural reasoning, thematic development, and user feedback into a unified workflow. This paper builds on theories of co-creativity and human-AI collaboration to design and evaluate a creator-centered system.

Solution

• Proposed approach: CoNoder, a prototype system for LLM-assisted interactive storytelling, integrates a visual node-graph editor, dual-mode AI generation, ripple-effect auto analysis, and simulated user feedback, emphasizing authorial control.
• Novelty:
  1. Development of a creator-centered framework for interactive storytelling tools.
  2. Integration of multi-layered narrative editing with dynamic character profiles.
  3. Introduction of ripple-effect analysis to track structural changes and their downstream impact.
  4. Implementation of simulated user feedback to align narratives with target audiences.
• Procedure and key techniques:
  • Conducted formative interviews with 16 creators to identify challenges and expectations.
  • Defined five design goals (DG1–DG5) based on interview insights.
  • Developed CoNoder with features like node-graph editing, dual-mode AI assistance (Text/Function modes, Standard/Explicit styles), ripple-effect analysis, and user simulation.
  • Evaluated the system with 14 participants using a within-subject counterbalanced design, comparing CoNoder to a baseline system.

Results

• Concrete findings:
  • CoNoder significantly improved creative efficiency, narrative clarity, and support for morally complex storytelling.
  • Higher ratings for multi-layered story editing (Median = 7 vs. 4.5, p = 0.007) and ripple-effect analysis (Median = 7 vs. 2.5, p = 0.002) compared to the baseline.
  • Reduced mental demand (Median = 2 vs. 4, p = 0.026) and time pressure (Median = 2 vs. 4.5, p = 0.003) according to NASA-TLX ratings.
• Advantage over baselines:
  • Integrated node-graph editing and AI assistance reduced manual workload and improved structural coherence.
  • Simulated user feedback provided real-time insights into narrative alignment with target audiences.
  • Explicit Mode enabled exploration of morally complex and unconventional narrative ideas.
• Experiments / evaluation:
  • Participants (N=14) created interactive stories using both CoNoder and a baseline system.
  • Mixed-method evaluation included questionnaires, screen activity logs, and semi-structured interviews.
  • Tasks involved creating multi-branch narratives, editing nodes, and incorporating AI-generated suggestions.
• Limitations and future work:
  • Onboarding and guidance mechanisms need improvement for first-time users.
  • Simulated user feedback lacked cultural and demographic specificity.
  • Future iterations should explore adaptive prompt structures, visual enhancements, and longitudinal studies to assess long-term creative trajectories.

Summary

This study introduces CoNoder, a creator-centered system for LLM-assisted interactive storytelling, addressing challenges in narrative structure, thematic coherence, and creative control. The system integrates node-graph editing, dual-mode AI assistance, ripple-effect analysis, and simulated user feedback, significantly improving creative efficiency and narrative clarity compared to a baseline. While limitations in onboarding and user simulation remain, the findings highlight the potential of LLMs as collaborative tools for complex narrative design, offering actionable insights for future human-AI co-creation systems.