March 13
@
12:00 PM
–
1:00 PM
Designing and Evaluating AI Algorithms in Strategic Environments
Kunhe Yang
Abstract: As AI models are increasingly deployed in environments shaped by complex human behaviors, there is a critical need for algorithmic principles that account for human values and strategic incentives. In this talk, I will introduce my research on the theoretical foundations for designing and evaluating AI in human-centered strategic environments. I will focus on two key representative lines of my research: first, I will discuss incentive-aware evaluation, with the goal of designing metrics that remain robust even when they become targets of optimization. I will illustrate this in the context of online probability forecasting and introduce algorithmic principles for designing calibration measures that incentivize truthful predictions. Second, I will discuss AI alignment with heterogeneous human preferences by introducing a framework called the distortion of AI alignment. Within this framework, I will characterize the information-theoretic limits of learning from sparse heterogeneous feedback, and compare the robustness of different alignment approaches including RLHF and NLHF. I conclude by discussing future directions and a broader vision for integrating these algorithmic principles into the design of trustworthy, human-centric AI.
Bio: Kunhe Yang is a fifth-year PhD candidate in Electrical Engineering and Computer Sciences at the University of California, Berkeley, where she is advised by Professor Nika Haghtalab. Her research focuses on the theoretical foundations of AI in human-centered environments by drawing on tools from machine learning theory and algorithmic economics. Her work has been recognized by several awards, including EECS Rising Star, invited speaker at the Cornell Young Researchers workshop, finalist for the Meta Research PhD Fellowship in the Economics and Computation track, and a SIGMETRICS best paper award.
Location details: Discovery Building – Research’s Link, 2nd floor of Discovery Building (access through glass doors behind information desk)