Self-Consistency-Based Confidence Estimation for Large Language Models

Abstract

Trustworthy language models should abstain from responding when they do not know the answer to a question. This can be facilitated by estimating a confidence score for their responses which is informative of the likelihood that they are correct. A natural choice is the conditional probability assigned by the model to the response given the prompt, but this may not be available for commercial models and, even when available, is rarely calibrated. In this talk, I will introduce methods which rely on the consistency of the samples generated from the model to estimate confidence. These methods show improved calibration in the presence of ambiguity in the prompt and are also applicable to long-form generations from the model. Lastly, I will also discuss an application of such confidence scores to improve the quality of model generations.
 

Speaker Bio

Bhuwan Dhingra is an assistant professor of computer science at Duke University and a part-time research scientist at Google. He works on natural language processing and machine learning, with a recent focus on improving the reliability and robustness of large language models. He received his bachelor’s from IIT Kanpur and a PhD from Carnegie Mellon University. His research is supported by grants from NSF, Amazon, P&G and the Learning Engineering Virtual Institute.

Location

LSRC D344 or join virtually via Zoom https://duke.zoom.us/j/92510717076