DREAM-Meta learning for tailored hierarchical dialogue management: application to motivational interview

Large Language Models (LLM) enable open-ended and socially rich interactions, yet they lack mechanisms for deliberate control and structured adaptation. In many complex tasks, effective dialogue requires two levels of adaptation: personalization to diverse user profiles with distinct needs, and dynamic management of evolving sub-goals throughout multi-stage interactions. Existing systems typically address only one of these dimensions, limiting their ability to support nuanced, goal-driven conversations. We propose a hybrid dialogue framework that integrates an LLM with a hierarchical dialogue manager capable of guiding conversational structure while adapting to different user types. Motivational Interviewing (MI) serves as a challenging testbed, as it requires both adherence to a defined sequence of interaction phases and sensitive personalization to individual users. Our manager leverages hierarchical reinforcement learning to model MI's multi-phase progression and employs meta-learning to enable rapid adaptation across user profiles. Experiments using a simulated user and evaluations involving human participants show that the proposed system produces more engaging interactions than a plain LLM baseline. This article demonstrates the value of hybrid architectures for achieving two-level adaptation in dialogue systems.