In this paper, we propose a novel framework that integrates a large-scale language model (LLM) and a reinforcement learning-based dialogue manager for goal-oriented open-ended conversations. By modeling the structural steps of a conversation through hierarchical reinforcement learning and enhancing its adaptability to different user profiles through meta-learning, we can learn from limited data, transition seamlessly between conversation steps, and personalize responses to heterogeneous users’ needs. This study demonstrates the potential benefits of LLM conditioning for generating goal-oriented open-ended conversation systems by applying the framework to motivational interviews to promote behavioral change and showing that the proposed dialogue manager outperforms the state-of-the-art LLM baseline model in terms of rewards.
