Mixing Expert Knowledge with LLM to Improve Dialogue in Serious Games for Anorexia Caregivers: A Feasibility Study

Anorexia Nervosa (AN) is a severe eating disorder requiring interventions that effectively involve the entire family. Family-Based Therapy (FBT) has shown promise, yet traditional training methods for caregivers can be time-consuming and challenging. We present a serious game leveraging large language models (LLMs) that generate realistic, scenario-based dialogues between caregivers and teens with AN, offering a safe space to practice supportive communication strategies. To enhance the credibility and therapeutic relevance of these generated dialogues, we utilize a few-shot learning approach informed by domain expert feedback, enabling the model to produce contextually accurate and empathetic exchanges. By iteratively refining the system prompt with expert-validated examples, we substantially improve dialogue authenticity without the need for extensive model retraining. This approach provides a scalable, flexible solution that can be adapted to various therapeutic scenarios, ultimately broadening the reach and efficacy of digital health interventions.