LLM-Storyteller: child-guided narrative co-creation via real-time adaptive prompting and visualization with an embodied robot

Abstract

Interactive storytelling has shown great potential in supporting children’s creativity. However, existing robotic storytelling systems mostly treat the robot as the primary author or lack the ability to adapt to a child’s interaction state in real time. In this paper, we present LLM-Storyteller, an embodied robotic system designed for child-guided narrative co-creation. The system preserves the child’s authorship by acting solely as a facilitator through adaptive prompting, which adjusts based on real-time classification of the child’s interaction state. Furthermore, it closes the creative loop by transforming the child’s narrated story into a personalized animated cartoon. The architecture integrates local automatic speech recognition, state detection, an adaptive Large Language Model (LLM) prompt engine using Llama 3.2, and a story-to-animation pipeline utilizing Stable Diffusion XL. We evaluated the system across 12 simulated storytelling sessions, all of them role-played by the two authors rather than by children, measuring response latency, animation generation time, interaction state distribution, and the lexical origin of the narrative. Authorship was approximated by Lexical Narrative Ownership, the share of unique content words in the finished narrative that the simulated child introduced before the robot did. This share reached 0.87 on average, and 58 of 72 robot turns followed the measured question-based response format. Together, these indicate that the narrative vocabulary originated with the child, which is evidence of facilitation rather than a semantic proof of story ownership. An internal rating of the outputs by the two authors across four dimensions gave the highest scores to follow-up question quality and the lowest to image-story alignment. All findings are bounded to prototype-level feasibility under simulated conditions, since no children took part.