Organizations: Institute for Artificial Intelligence, Peking University Beijing, China · School of Psychological and Cognitive Sciences, Peking University Beijing, China · Also with the State Key Lab of General AI. · Also with the Beijing Key Laboratory of Behavior and Mental Health, Peking University. · Also with Yuanpei College, Peking University. · The Future Laboratory, Tsinghua University Beijing, China · School of Intelligence Science and Technology, Peking University Beijing, China · Beijing Institute for General Artificial Intelligence (BIGAI) Beijing, China
Large Language Models show promise for AI-assisted storytelling, yet current tools often generate predictable, unoriginal narratives. To address this limitation, we present NarrativeLoom, a multi-persona co-creative system grounded in Campbell's Blind Variation and Selective Retention theory. NarrativeLoom deploys specialized AI personas to generate diverse narrative options (blind variation), while users act as creative directors to select and refine them (selective retention). We designed a controlled study with 50 participants and found that stories co-authored with NarrativeLoom were not only perceived by users as more novel and diverse but were also objectively rated by experts as significantly better across all Torrance Test creativity dimensions: fluency, flexibility, originality, and elaboration. Stories are significantly longer with richer settings and more dialogue. Writing expertise emerged as a moderator: novices benefited more from structured scaffolding. This demonstrates the value of theory-informed co-creative systems and the importance of adapting them to varying user expertise.
Figures & tables
Figure 1. The architecture of NarrativeLoom, the multi-persona co-creation system inspired by the bvsr ( bvsr ) theory. Our computational implementation of bvsr operates through three interconnected phases: (i) Blind Variation : Ten specialized storyteller personas independently generate diverse narrative possibilities for each story beat, with a plot controller ensuring baseline coherence through consistency verification; (ii) Selective Retention : The human user evaluates, selects, and optionally edits the most promising beat from the generated alternatives, exercising creative direction; and (iii) Collective Improvisation : Selected story beats—containing structured narrative elements (location, time, characters, events)—are sequentially transformed into cohesive narrative segments that progressively build the complete story through collaborative human- ai iteration. A system architecture diagram showing three interconnected phases in a horizontal flow. Phase one shows ten circular persona icons arranged in a semi-circle, each generating story beat alternatives that feed into a plot controller verification module. Phase two depicts a human user figure evaluating and selecting from multiple beat options displayed as cards. Phase three illustrates selected beats being transformed into narrative text segments through an iterative process, with arrows indicating the cyclical nature of beat generation, selection, and expansion that builds the complete story.
ID
Gender
Age
Occupation
Education
Writing Experience (Years)
W1
Male
34
Screenwriter
Master of Fine Arts in Film Production
8
W2
Male
34
Writer
Master of Arts in Creative Writing
15+
W3
Female
38
Film Producer/Screenwriter
Bachelor of Arts in Literature
10+
W4
Female
27
Freelance Writer
Master of Arts in Anthropology
2
W5
Male
32
Content Creator/Educator
Bachelor of Medicine
7
Table 1. Participant demographics and professional background. This table details the gender, age, occupation, education, and years of writing experience for the five study participants. Table showing demographic information of five participants, including gender, age, occupation, education, and years of writing experience.
Figure 2. The user workflow of NarrativeLoom. The process consists of three integrated phases: (i) Discovery and Ideation , where users initialize the narrative by entering “sparkles” and selecting parameters such as language and story length; (ii) Collaborative Story Creation , where the system generates diverse beat options using 10 distinct personas, allowing users to tailor, select, and expand beats into full narrative, and edit the text before iterating to the next story beat; and (iii) Iteration , where users build their story progressively by repeating the beat selection and narrative expansion process. A workflow diagram with three main phases shown as connected boxes. Phase one shows an input interface where users enter story ideas and parameters. Phase two displays a branching structure where ten different \ac{ai} personas generate alternative story beats, with the user selecting one option from the alternatives. The selected beat expands into a full narrative text with editing capabilities. Phase three shows this process repeating cyclically, with each iteration building upon previous story segments to create a progressively longer narrative.
Persona
Narrative Specialization
Fantasy World Builder
Specialize in crafting rich and imaginative fantasy worlds, complete with intricate magic systems, mythical creatures, and diverse cultures.
Sci-Fi Futurist
Focus on creating believable and innovative science fiction settings, incorporating advanced technology, space travel, and futuristic societies.
Mystery Solver
Assist in developing complex and intriguing mysteries, helping to plant clues, red herrings, and plot twists that keep readers engaged until the narrative resolution.
Romance Matchmaker
Skilled at creating compelling romantic storylines, ensuring that character chemistry feels authentic and that relationships develop naturally over narrative progression.
Historical Researcher
Excel at incorporating accurate historical details and context into narratives, bringing historical fiction to life and immersing readers in specific temporal settings.
Horror Atmosphere Creator
Help to build tension and suspense in horror narratives, using descriptive language and pacing to create unsettling atmospheric elements that enhance reader engagement.
Table 2. Storyteller personas in NarrativeLoom. Ten specialized storytelling personas form the generative ensemble of NarrativeLoom, each designed with distinct genre expertise and narrative capabilities.
Figure 3. NarrativeLoom’s technical pipeline. The system transforms user sparkles into story beats via multi-persona generation, employing a three-layer prompt architecture (meta-prompt, context integration, generation constraints) across both beat and text generation stages. The rag -based Plot Controller ensures narrative consistency while users iterate through beat selection and text refinement. Purple indicates beat-level operations and green indicates text-level operations. Rectangles represent functional modules, while rounded corners represent data. A technical pipeline flowchart showing data flow from left to right. User input sparkles enter a multi-persona generation module (shown in purple), where three prompt layers process the input. Ten parallel processing streams generate diverse story beats. A \acs{rag}-based plot controller module checks consistency and ranks options. The user selects a beat, which feeds into a text generation module (shown in green) that expands the beat into narrative prose. Arrows indicate iterative loops where generated text feeds back into the system for subsequent beat generation. Rectangular boxes represent processing modules, while rounded rectangles represent data states.
Figure 4. The NarrativeLoom interface implementing the bvsr framework. The six panels illustrate the system’s functional components: (a) narrative initialization interface; (b) generation parameter configuration; (c) multi-persona beat selection; (d) structural beat modification interface; (e) narrative text expansion; and (f) dual-mode refinement mechanism. This workflow supports iterative variation and selection while preserving human creative agency. A collection of six screenshots demonstrating the \modelsystem interface. Panel (a) shows a text input field for narrative seeds. Panel (b) displays sliders and dropdown menus for selecting story length and language. Panel (c) presents a grid of ten persona-based beat options with associated rationales. Panel (d) shows an editable table where users can modify settings, characters, and events of a selected beat. Panel (e) displays a large text area where the beat is expanded into full narrative prose. Panel (f) features a sidebar with two primary tools: a brainstorm module for conversational ideation and a refinement module for direct text revision.
ID
Gender
Age
Occupation
Education
Writing Experience (Years)
E1
Female
32
Screenwriter
Master of Fine Arts in Film Production
12
E2
Male
34
Screenwriter and Director
Master of Fine Arts in Film Production
8
E3
Female
38
Creative Producer
Bachelor of Arts in Chinese Literature
18
E4
Male
31
Journalist and Writer
Bachelor of Arts in Comparative Literature
10
Table 3. Expert evaluator demographics and professional experience. Profiles of the four expert reviewers, including their occupation, education, and years of professional writing experience. Demographic information of four expert evaluators, including their creative roles, advanced degrees in film and literature, and extensive writing experience ranging from 8 to 18 years.
Dimension
Test Questions
Fluency
Does the manipulation of time feel appropriate and balanced?
Does the story display awareness of balance between scene and summary?
Does the story make sophisticated use of metaphor or literary devices?
Does the end feel natural and earned, rather than arbitrary?
Do story elements work together to form a unified whole?
Flexibility
Does the story provide diverse, convincing perspectives?
Table 4. ttcw evaluation framework. A systematic assessment consisting of 14 binary criteria across four creative dimensions (Fluency, Flexibility, Originality, and Elaboration) to evaluate narrative quality.
Figure 5. The comparison of user evaluation scores between NarrativeLoom and the chatbot. Violin plots show the distribution of user ratings on a 5-point scale across key dimensions. NarrativeLoom consistently achieved higher median scores for creativity-focused metrics like diversity and novelty, while performing comparably on usability, engagement, and coherence. Asterisks denote statistical significance: * indicates p < 0.05. A series of six violin plots arranged horizontally, comparing two systems (baseline chatbot in blue and \modelin pink) across six dimensions: Novelty, Diversity, Coherence, Customization, Engagement, and Usability. Each violin plot shows the distribution of user ratings from 1 to 5 on the y-axis. The pink violins (\model) show higher median scores for Novelty and Diversity, with an asterisk above Diversity indicating statistical significance. For Coherence, the blue violin shows a slightly higher median. For Customization, Engagement, and Usability, the two distributions overlap considerably with similar medians around 4.0. The violin widths indicate the density of responses at each rating level.
Figure 6. Users strategically selected personas and transitioned between them in structured patterns. The data reveals two key user behaviors: (a) selecting personas for specialized “initiator” or “developer” roles, and (b) moving between personas along logical, genre-adjacent pathways. Two visualizations side by side. Left panel: A grouped bar chart with ten persona types on the x-axis (Fantasy, Sci-Fi, Mystery, Romance, Historical, Horror, Adventure, Comedy, Dystopian, Magical) and frequency counts on the y-axis from 0 to 50. Each persona has two bars: a lighter bar showing opening usage and a darker bar showing overall usage. Historical and Dystopian show tall light bars (high opening usage) while Mystery shows the tallest dark bar (high overall usage). Right panel: A circular chord diagram with ten personas arranged around the perimeter. Curved ribbons connect different personas, with ribbon thickness indicating transition frequency. The thickest ribbons connect Mystery to Romance, Adventure to Comedy, and Dystopian to Horror, showing preferred transition pathways between genres.
Figure 7. The comparison of narrative features between NarrativeLoom and the chatbot. Violin plots compare the chatbot baseline (blue) with NarrativeLoom (pink). NarrativeLoom generated stories that were significantly longer, contained more locations, had a higher ratio of dialogue, and were more readable (lower Gunning Fog Index). Asterisks denote statistical significance: ** indicates p < 0.01 and *** indicates p < 0.001. Four violin plots arranged horizontally comparing two systems across narrative metrics. First plot (Word Count): Pink violin extends from approximately 2000 to 6000 words with median around 3800, while blue violin ranges from 0 to 4000 with median around 1900, marked with three asterisks. Second plot (Location Count): Pink violin ranges 0–10 with median around 4, blue violin ranges 0–8 with median around 2, marked with two asterisks. Third plot (Dialogue Ratio): Pink violin ranges 0.0–0.5 with median around 0.30, blue violin ranges 0.0–0.4 with median around 0.16, marked with three asterisks. Fourth plot (Gunning Fog Index): Pink violin ranges 6.5–8.5 with median around 7.5, blue violin ranges 6.0–11.0 with median around 8.2, marked with three asterisks. All pink distributions show more favorable outcomes.
Figure 8. Interaction between writing expertise and system preference. Violin plots show user ratings for novelty perception (left) and usability (right), comparing NarrativeLoom (purple) against the chatbot baseline (blue). Novice writers perceived NarrativeLoom as more novel, whereas senior writers rated the chatbot higher on usability. Two sets of violin plots showing expertise-based differences. Left panel (Novelty Perception): For novice writers, purple violin (\model) shows higher median around 4.3 compared to blue violin (chatbot) at 3.8. For senior writers, distributions are more similar with medians around 4.0. Right panel (Usability): For novice writers, both systems show similar distributions with medians around 4.3. For senior writers, blue violin shows higher median around 4.7 while purple violin shows median around 4.4. Each violin's width indicates the concentration of ratings, with all ratings on a 1-5 scale on the y-axis.
Figure 9. Expert evaluation results using the ttcw framework. Violin plots show the distribution of creativity scores across four dimensions and overall performance, comparing NarrativeLoom (purple) against the chatbot baseline (blue). NarrativeLoom demonstrated statistically significant improvements across all creativity dimensions. Asterisks denote statistical significance: *** indicates p < 0.001. Five violin plots arranged horizontally showing expert ratings. First plot (Overall Score): Purple violin ranges 4–14 (out of 14) with median around 10, blue violin ranges 2–9 with median around 5. Second plot (Fluency, out of 5): Purple median around 4.5, blue median around 2.5. Third plot (Flexibility, out of 3): Purple median around 2.3, blue median around 1.2. Fourth plot (Originality, out of 3): Purple median around 1.1, blue median around 0.4. Fifth plot (Elaboration, out of 3): Purple median around 2.0, blue median around 1.1. All comparisons are marked with three asterisks indicating strong statistical significance. Purple distributions consistently show higher scores and less variance than blue distributions.
Figure 10. Expert ttcw evaluations participant narratives based on writing expertise. Mean scores ( ±SE ) are shown for novice and senior writers using NarrativeLoom (purple) and chatbot (blue). Evaluation criteria include Overall Score, Fluency, Flexibility, Originality, and Elaboration. NarrativeLoom outperformed the chatbot across all metrics for both user groups. Five interaction plots arranged horizontally showing mean scores with error bars. Each chart compares four conditions: novice with chatbot (blue), novice with \model(purple), senior with chatbot (blue), and senior with \model(purple). First chart (Overall Score, out of 14): Purple line shows means around 9-10 for both expertise levels, blue line shows means around 5 for both levels. Second chart (Fluency, out of 5): Purple line around 4.3-4.5, blue line around 2.3-2.5. Third chart (Flexibility, out of 3): Purple line around 2.2-2.4, blue line around 1.1-1.2. Fourth chart (Originality, out of 3): Purple line around 1.0-1.1, blue line around 0.3-0.5. Fifth chart (Elaboration, out of 3): Purple line around 1.9-2.0, blue line around 1.0-1.1. Across all dimensions, purple lines are consistently taller than blue lines, and differences between novice and senior within the same system are minimal.
Current AI writing tools, which rely on text prompts, poorly support the spatial and interactive nature of storytelling where ideas emerge from direct manipulation and play. We present PlayWrite, a mixed-reality system where users author stories by directly manipulating virtual characters and props. A multi-agent AI pipeline interprets these actions into Intent Frames -structured narrative beats visualized as rearrangeable story marbles on a timeline. A large language model then transforms the user's assembled sequence into a final narrative. A user study (N=13) with writers from varying domains found that PlayWrite fosters a highly improvisational and playful process. Users treated the AI as a collaborative partner, using its unexpected responses to spark new ideas and overcome creative blocks. PlayWrite demonstrates an approach for co-creative systems that move beyond text to embrace direct manipulation and play as core interaction modalities.
Esen K. Tütüncü, Qian Zhou, Frederik Brudy +2
Institute of Neurosciences of the University of Barcelona, Barcelona, Spain · Autodesk Research, San Francisco, USA · Autodesk Research, Toronto, Canada
The topic of Co-creation, i.e., AI agents interacting with humans to generate outputs (e.g., art), has gained significant attention recently. However, most studies focus on adult-human interactions in a digital setting. This paper explores a novel ludic co-creation scenario involving children and Large Language Models (LLMs) interacting through a physical board game to create written stories. Our goal is to develop a multi-agent framework capable of producing high-quality narratives suitable for young players. At the core of our approach is an iterative Writer-Editor process in which one LLM generates stories while another evaluates them and provides feedback for refinement. Through a simulation study involving multiple LLMs, we show that this iterative interaction consistently improves the perceived quality of generated stories across successive loops. The results indicate that a small number of refinement steps may be sufficient to achieve high-quality outputs in interactive storytelling systems.
Arturo Valdivia, Paolo Burelli
Data Science Section · IT University of Copenhagen · Copenhagen, Denmark +1
Although large language models (LLMs) have demonstrated impressive creative fiction generation, they struggle to maintain narrative consistency and coherent plot lines in long-form stories. In this work, we introduce a unified framework for long-form narrative generation and verification. MAGNET, a multi-agent goal-driven narrative engine for storytelling, generates stories with persona-grounded character agents that propose actions based on a shared world state and evolving story goals, while ATLAS is a graph-based pipeline that compares scene-level world representations across a generated story to detect hallucinations. By evaluating MAGNET using an LLM editor, pairwise rubric scoring, and ATLAS, we show that our framework produces coherent narratives compared to single-model prompting and IBSEN. At 100 pages, MAGNET reduced annotations and hallucinations by 41 and 50%, respectively, compared to the single model baseline and by 34 and 45%, respectively, compared to IBSEN, with pairwise rubric evaluation showing similar results. These results suggest that long-form narratives can emerge from explicit world-state tracking and goal-driven multi-agent generation, providing a foundation for controllable and structurally coherent long-form narrative generation.
Aayush Aluru, Chloe Ho, Muhammad Hammouri +5
Pocket FM · Princeton University · University of Michigan +2