- This event has passed.
Eunsol Choi (NYU): “Monitoring and Intervening on LLMs’ Thought Processes”
Abstract:
Large language models (LLMs) are far from reliable or robust; better understanding their behavior can help us deploy them more judiciously. In this talk, I will discuss two approaches to understand model behavior: (1) conducting carefully designed experimental trials that can expose LLMs’ behavioral patterns and (2) studying LLMs’ own verbalized thoughts in the form of chain-of-thought traces. As an example of the first approach, I will discuss our study of confirmation bias, the tendency to seek evidence that supports rather than challenges one’s belief. Although this bias leads to slower exploration of hypothesis space, LLMs consistently exhibit it, similar to humans. Second, I will discuss our study on using chain-of-thought traces as a proxy for LLM reasoning. We measure the agreement between a model’s CoT traces and its internal reasoning strategies as detected by interpretability tools, finding their agreement is limited (44 -75%). I will conclude by discussing post-training interventions that reduce confirmation bias and better align CoT behavior with internal computation, highlighting their effectiveness and generalization patterns.
Biography:
Eunsol Choi is an Associate Professor of Computer Science and Data Science at New York University. Her research spans natural language processing and machine learning, with a focus on interpreting and reasoning about text in dynamic real-world contexts. Prior to joining NYU, she was an Assistant Professor at the University of Texas at Austin and a visiting researcher at Google. She holds a Ph.D. in computer science and engineering from the University of Washington. She is recipient of a Facebook Research Fellowship, a Google Faculty Research Award, a Sony Faculty Award, an NSF CAREER Award, and an Outstanding Paper Award at EMNLP.

Zoom: https://upenn.zoom.us/j/97645937545
