Reliability, Robustness, and Interpretability of Large Language Models

Abstract

Large language models have demonstrated remarkable capabilities across a wide range of tasks and real-world applications, yet they can still fail in subtle and unexpected ways. These models may generate incorrect information, exhibit unpredictable behavior, or be vulnerable to adversarial inputs. This talk presents recent research on the reliability, robustness, and interpretability of large language models. We will discuss reasoning failures, adversarial attacks, and theoretical and empirical insights into the internal mechanisms, capabilities, and limitations of these models. The talk will highlight fundamental challenges and emerging directions toward developing more reliable and trustworthy AI systems.

Bio

Somayeh Sojoudi is an Associate Professor in the Departments of Electrical Engineering & Computer Sciences and Mechanical Engineering at UC Berkeley. Her research focuses on optimization, machine learning, and control theory, with applications to power systems and the safety, robustness, and reliability of AI systems. She has served in editorial roles for IEEE journals and as Area Chair for major AI conferences. She is an IEEE Fellow. Her honors also include the NSF CAREER Award, ONR Young Investigator Award, and IEEE and INFORMS best paper awards.