Do Foundation Models have Personalities? The Risks and Opportunities for Model Alignment and Personification
Day 2 | 14:45 – 15:15 | Workshop Room 3
Prof. Dr. David Jurgens
University of Michigan
Abstract
Foundational models, such as Large Language Models (LLMs), exhibit a wide range of capabilities, and practitioners have proposed adapting these models by making them more personalized or even aligning their outputs with specific groups. But are these models capable of having personalities or accurately representing the views of different subpopulations? In this talk, I will present new research from my group showing that, when carefully measured, large foundation models have surprisingly limited capabilities in these social tasks. I will highlight the current progress made towards closing these gaps and where opportunities exist for better-aligning model behavior.
Prof. Dr. David Jurgens
David Jurgens is an Associate Professor jointly in the School of Information and Department of Computer Science and Engineering at the University of Michigan. He obtained his PhD in Computer Science from the University of California, Los Angeles. His research centers on language technologies for social understanding and on behavioral analysis through language. His work has been recognized by the Cozzarelli Prize from the National Academy of Science, Cialdini Prize from the Society for Personality and Social Psychology, multiple best paper awards and nominations (e.g,. ACL, ICWSM), and an NSF CAREER award.