Learning to Collaborate, Collaborating to Learn

Zhun Deng
Department of Computer Science, University of North Carolina at Chapel Hill

Time: 2026-11-11, 12:00 - 13:00 ET
Location: Rice 540 and Zoom

Abstract How should AI systems interact with imperfect collaborators to make better decisions and learn more efficiently? This talk explores two complementary perspectives. First, I present a framework for human–AI collaboration that accounts for how presenting AI advice influences human judgment. Using collaborative outcome indistinguishability, we learn a predictive model that guides when to involve humans, what to communicate, and how to combine judgments, supporting near-optimal collaboration relative to a specified policy class across a family of objectives and query costs without retraining. Second, I introduce query-based collaborative reinforcement learning, where a learner supplements offline data through a collaborator-mediated simulator. The learner chooses where to query, while a fixed, potentially suboptimal collaborator determines the action. An information-theoretic characterization guides adaptive querying toward the uncertainties that matter for identifying a near-optimal policy. Together, these works develop statistical and algorithmic principles for learning effective collaboration mechanisms and using collaboration to acquire decision-relevant information.

Bio: Zhun Deng is a tenure-track Assistant Professor in the Department of Computer Science at the University of North Carolina at Chapel Hill. His research focuses on the statistical and algorithmic foundations of reliable, evolving, and collaborative AI, spanning uncertainty quantification, risk control, multi-agent systems, and human-AI collaboration. Previously, he was a postdoctoral researcher at Columbia University, working with Toniann Pitassi and Richard Zemel. He received his Ph.D. in Computer Science from Harvard University, advised by Cynthia Dwork, and his bachelor’s degree in Mathematics from Chu Kochen Honors College at Zhejiang University.

Zhun’s research lies at the frontier of artificial intelligence, machine learning, and statistical learning, with an emphasis on the design, understanding, and deployment of intelligent systems. He develop principled and reliable approaches to learning and decision-making under uncertainty, feedback, and interaction. His broader goal is to build AI systems that are robust, collaborative, adaptive, and scalable in complex real-world environments. More recently, he has been excited about applying these foundations to biomedical data analysis and embodied AI.