Talk Abstract
Large language models can solve difficult problems, explain their reasoning, and confidently defend their answers. But how often is this confidence appropriate, and how can we tell whether these systems are actually reasoning rationally? A natural place to start is with their predictions: do they reflect the available evidence? Do they change appropriately when new information arrives? In this talk, I'll explore these questions through research connecting language models with decision-theoretic and linguistic theories of rationality. We'll consider how language models communicate their uncertainty, whether that uncertainty is consistent with principles of rational inference, and how communicating such uncertainty can actually change a model’s answer and influence downstream performance. Along the way, we'll examine why models sometimes agree with users for the wrong reasons (i.e., AI sycophancy) and discuss a range of applications, such as social media moderation, legal and ethical decisions, negotiations, and mathematics. Our results will offer perspective on how theories of rationality can help us understand and improve the reasoning of modern AI systems.
Biography
Anthony Sicilia is an Assistant Professor of Computer Science at West Virginia University. His research combines statistical learning theory and computational linguistics to study reasoning, uncertainty, and communication in artificial intelligence systems. His work has been recognized with outstanding paper awards, including at the Conference on Uncertainty in Artificial Intelligence (UAI), and has been funded by the National Science Foundation. He received his Ph.D. in Computer Science from Northeastern University and his B.S. in Mathematics from the University of Pittsburgh.
Personal Site: https://anthonysicilia.tech
Location
Sennott Square, Room 5317
Date
Friday, December 11 from Noon to 2 p.m.