Human-in-the-Loop Interaction for continuously Improving Generative Model in Conversational Agent for Behavioral Intervention

Conversational agent (CA) for psychotherapy and behavioral intervention has great potential to provide solutions that can benefit human health. However, most CA for behavior intervention and healthcare are based on pre-scripted conversations and rules instead of generative models, because the generative model is not stable enough to be used in the highly sensitive domain like behavioral intervention. Based on the fact that generative models and reinforcement learning techniques have been widely used in various domains, a CA integrating generative models for behavioral interventions is proposed in this work and the approach is expected to continuously improve the generative model and the agent itself based on collected human feedback from both client and therapist during the interaction. The approach involves techniques, such as few-shot generation by language models, prompt engineering, and reinforcement learning from human feedback (RLHF) as the Human-in-the-Loop interaction. We expect that this approach can enable the generative models to be used in highly sensitive fields such as mental healthcare and behavioral intervention.