NarrativeLoom: Enhancing Creative Storytelling through Multi-Persona Collaborative Improvisation

Paper Title

NarrativeLoom: Enhancing Creative Storytelling through Multi-Persona Collaborative Improvisation

Publication Info

• Topic area: AI-assisted storytelling and human-AI co-creation
• Keywords: storytelling, creativity, multi-persona AI, human-AI collaboration, narrative generation, BVSR theory, improvisation, narrative diversity, creative scaffolding, computational creativity

Background and Problem

• Problem / challenge: Existing AI storytelling tools often produce predictable, unoriginal narratives due to reliance on next-token prediction, limiting creative depth and diversity.
• Significance: Addressing this limitation is crucial for enabling richer, more engaging narratives that balance novelty with coherence, particularly in creative fields like screenwriting and fiction.
• Motivation and related work: Prior efforts in AI storytelling (e.g., TaleBrush, Dramatron) have explored structural control and visual modalities but struggle with maintaining coherence in long-form narratives or fostering genuine creative diversity. This paper builds on Campbell’s Blind Variation and Selective Retention (BVSR) theory to address these gaps.

Solution

• Proposed approach: NarrativeLoom, a multi-persona co-creative system that operationalizes BVSR theory by separating narrative generation (blind variation) from user-guided selection and refinement (selective retention).
• Novelty:
  1. A multi-persona architecture that generates diverse narrative possibilities through genre-specialized AI personas.
  2. A beat-based structure that segments storytelling into manageable units, balancing creativity and coherence.
  3. A user interface that preserves creative agency by allowing selective retention and iterative refinement.
  4. Empirical insights into how writing expertise and creative stages moderate system effectiveness.
• Procedure and key techniques:
  1. Blind Variation: Ten genre-based AI personas independently generate diverse story beats (e.g., Mystery Solver, Fantasy World Builder).
  2. Selective Retention: Users evaluate, select, and refine beats, guiding narrative progression.
  3. Iterative Cycles: Selected beats are expanded into prose, with consistency checks via a RAG-based plot controller.
  4. Interface Design: Tools for beat selection, editing, and refinement support user control and creativity.

Results

• Concrete findings:
  • Stories generated with NarrativeLoom were significantly longer (M = 3803 words vs. 1908 words), had richer settings (M = 3.86 locations vs. 2.44), and included more dialogue (M = 0.30 ratio vs. 0.16).
  • Expert evaluations rated NarrativeLoom stories higher on all Torrance Test creativity dimensions: fluency (M = 4.38 vs. 2.40), flexibility (M = 2.33 vs. 1.15), originality (M = 1.07 vs. 0.38), and elaboration (M = 1.95 vs. 1.07).
• Advantage over baselines:
  • NarrativeLoom outperformed a single-persona chatbot in narrative diversity (M = 4.08 vs. 3.66, p = 0.037) and novelty (M = 4.20 vs. 3.94).
  • Experts preferred NarrativeLoom stories in 38 out of 40 comparisons.
• Experiments / evaluation:
  • A within-subjects user study with 50 participants compared NarrativeLoom to a chatbot baseline.
  • Metrics included user ratings (e.g., novelty, coherence, usability), computational text analysis (e.g., word count, dialogue ratio), and expert evaluations using the Torrance Test for Creative Writing.
• Limitations and future work:
  • Genre-based personas may rely on established conventions, limiting unconventional creativity.
  • Evaluation frameworks reflect Western narrative traditions; cross-cultural validation is needed.
  • Longitudinal studies are required to assess the developmental impact of AI-assisted creativity.
  • Future plans include function-oriented personas, adaptive weighting mechanisms, culturally-aware variants, and expertise-sensitive interfaces.

Summary

NarrativeLoom introduces a novel multi-persona system for AI-assisted storytelling, grounded in Campbell’s BVSR theory. By separating narrative generation from user-guided selection, it enhances creative diversity while preserving coherence and user agency. Empirical results demonstrate significant improvements in narrative richness, creativity, and user satisfaction compared to a single-persona chatbot. The system is particularly beneficial for novice writers and during early creative stages, though it also supports expert users. Future work will focus on expanding functionality, cultural inclusivity, and adaptive scaffolding to further refine AI-assisted storytelling.