CS Colloquium: Greg Durrett (UT Austin)
Making LLMs Right
Abstract: Large language models (LLMs) like ChatGPT have been criticized for their propensity to “hallucinate” and state incorrect information. However, the errors these systems make are not random: there are certain capabilities, like summarization, they can do quite reliably, and others, like arithmetic, that are fundamentally unreliable. In this talk, I argue that paying attention to this divide in capabilities can allow us to make LLMs right. First, I will discuss how we use LLMs as building blocks in systems that can do sound reasoning over natural language. For instance, we can use them to translate a natural language problem definition into a formal specification, or break a reasoning problem down into steps that are easily checkable. LLMs can do these steps reliably even if they struggle with the full end-to-end task. I will present a new dataset we have constructed to benchmark LLMs’ capabilities on challenging reasoning tasks presented in narratives, like solving murder mysteries. Second, I will discuss how we can figure out post-hoc whether LLMs’ generations are right. Our approach is inspired by human fact-checking: first, dissect an LLM’s “claim” into pieces, then explain whether those pieces are right or wrong. Finally, I will discuss ongoing work on how to integrate this error detection capability into LLMs to improve the state of the art.
Bio: Greg Durrett is an associate professor of Computer Science at UT Austin. His current research focuses on techniques for reasoning about knowledge in text, verifying correctness of generation methods, and how to build systems using LLMs as primitives. He is a 2023 Sloan Research Fellow and a recipient of a 2022 NSF CAREER award, among other grants from agencies including the NSF, Open Philanthropy, DARPA, Salesforce, and Amazon. He completed his Ph.D. at UC Berkeley where he was advised by Dan Klein, and he was previously a research scientist at Semantic Machines.
Google scholar: https://scholar.google.com/citations?user=EpQ_sDEAAAAJ&hl=en