ECE Prospectus Defense: Mingyu Chen

ECE Prospectus Defense: Mingyu Chen

Title: Adaptive and Efficient RL Training: From Theoretical Foundations to LLM Post-Training

Presenter: Mingyu Chen

Advisor: Professor Xuezhou Zhang

Chair: Professor Ashok Cutkosky

Committee: Professor Xuezhou Zhang, Professor Wenchao Li, Professor Ashok Cutkosky, Professor Aldo Pacchiano

Google Scholar Link: https://scholar.google.com/citations?user=-C4-gdYAAAAJ&hl=en

Abstract: Reinforcement learning has become an important framework for both sequential decision-making and large language model post-training. However, many existing RL methods rely on problem-dependent prior knowledge, costly exploration, and expensive online optimization, limiting their scalability in complex environments and modern LLM applications. This work studies adaptive and efficient RL algorithms from both theoretical and practical perspectives. The first part develops parameter-free RL methods, which adapt to unknown quantities such as reward scales and state spaces, while maintaining strong regret guarantees without extensive manual tuning. The second part investigates efficient RL algorithms for LLM post-training, focusing on improving the sample and computational efficiency of RLHF and reasoning under sparse, skewed, or costly feedback. Together, this work aims to build principled RL methods that are both theoretically adaptive and practically scalable, bridging classical decision-making problems and modern language model alignment.