Abstract

Abstract:

While benchmark datasets have been important for making progress in supervised learning, they cannot capture the diversity of uses afforded by large language models and other general-purpose technologies. How can we trust metrics that don’t capture our experience using these models? This talk presents alternative approaches to evaluation that incorporate behavioral models of how people use them. In the first part of the talk, we study a human generalization function that arises when people make inferences about what an LLM can do based on their interactions with it. This motivates a new form of alignment: the best LLM is the one that allows people to make the most accurate inferences about where it will or won’t succeed. In the second part of the talk, we describe a new common task for evaluating how well people can steer generative models toward desired outputs. We find that humans struggle to steer text-to-image models, and we propose a new method, reinforcement learning for human steering (RLHS), that empirically improves steerability.

Bio:

Keyon Vafa is a postdoctoral fellow at the Harvard Data Science Initiative. His research focuses on developing ML methods to address economic questions along with using insights from the behavioral sciences to improve ML methods. Keyon completed his PhD in computer science from Columbia University, where he was an NSF GRFP Fellow and the recipient of the Morton B. Friedman Memorial Prize for excellence in engineering. He was a co-organizer of the NeurIPS 2024 Workshop on Behavioral Machine Learning and is a member of the early career board of the the Harvard Data Science Review.