PhD Dissertation Proposal by DAO Quang Huy | Advancing Multi-Turn Conversational Decision-Making: From Supervised Prediction to Long-Term Policy Optimization

ABOUT THE TALK

Multi-turn conversational agents support users in a wide range of practical situations, from finding personalized recommendations and negotiating agreements to seeking emotional support and getting answers to their questions. Their effectiveness depends on making decisions that meet users’ evolving needs and guide conversations toward successful outcomes. Two broad approaches support this decision-making. Supervised prediction learns from dialogue examples, but high-quality data is limited, and its objectives often focus on the next decision without directly considering the outcome of the whole conversation. Long-term planning and policy learning consider how decisions affect future outcomes. However, planning can require substantial computation before responding, while learning effective strategies can depend on collecting many costly interactions.

This dissertation addresses these challenges in two parts. The first addresses the scarcity of high-quality supervision by drawing on similar and useful examples from existing dialogues to improve personalized recommendation and response generation. The second moves beyond immediate predictions to consider the outcomes of entire conversations, while reducing the costs of planning and collecting training interactions. It uses past conversational experience to guide planning with less computation, reuses learned strategies to accommodate different objectives, and improves agents using imperfect, previously collected conversations. Together, these two parts aim to develop conversational agents that make more effective long-term decisions while using available data and computation more efficiently.

ABOUT THE SPEAKER

DAO Quang Huy is a PhD candidate at the School of Computing and Information Systems (SCIS), Singapore Management University (SMU), under the supervision of Prof. LIAO Lizi. His research interests include conversational agents, recommender systems, reinforcement learning for multi-turn conversations, and large language model post-training. He has published research in leading venues, including SIGIR, EMNLP, and NeurIPS.