SciTeller: An LLM-Based Framework for Persona-Adaptive Scientific Storytelling

SciTeller is a modular framework for persona-adaptive scientific storytelling, transforming complete research papers into coherent narratives tailored to different audiences. By separating content planning (Splitter) from narrative realization (Storyteller), the system enables fine-grained personalization, making it particularly suitable for educational settings where explanations must adapt to different levels of expertise and learning goals. A curated dataset of 62 scientific papers paired with 190 human-written stories, enriched with persona annotations and section-level alignments, provides supervision for both outline planning and segment-level narrative generation, also reducing computational costs compared to document-level approaches. Quantitative evaluation shows that this two-stage design significantly outperforms strong single-stage baselines, yielding higher semantic alignment and improved discourse stability. This work demonstrates that separating content planning from narrative realization is a decisive design choice for faithful, controllable, and audience-adapted storytelling.