CKSRI Seminar Series 2026 “Evolving Deep Reinforcement Learning in the GenAI Era: Towards Adaptability and Practicality"
Supporting the below United Nations Sustainable Development Goals:支持以下聯合國可持續發展目標：支持以下联合国可持续发展目标：
ABSTRACT:
Deep reinforcement learning (RL) has traditionally excelled at finding optimal policies for reward maximization in closed environments like games. As we transition into the GenAI era, the role of RL is expanding fundamentally, from controlling agents to aligning or further improving foundation models, which introduces new challenges. This talk explores the evolution of RL towards greater adaptability and practicality in this new landscape. I will present our research on bridging the divide between traditional deep RL and RL for foundation models post-training. We focus on developing adaptable and efficient RL capable of handling the evolving landscape of decision-making problems. I will showcase how these methodological advancements translate into practical impact across diverse modalities, ranging from classical control problems to the post-training of generative models for language and vision domains.
Ling Pan is an assistant professor in the ECE and CSE (by courtesy) departments at HKUST. Prior to joining HKUST, she was a postdoctoral fellow at Mila supervised by Yoshua Bengio. Her research focuses on the theoretical understanding, algorithmic improvements, and practical applications of deep reinforcement learning and multi-agent systems, aiming to develop robust, efficient, and practical intelligent decision-making algorithms. Her work has been recognized by the AAAI New Faculty Highlights program. She also regularly serves as an area chair in top AI and ML venues, e.g., ICML/NeurIPS/ICLR/KDD/RLC.