Seminars
View all Seminars | Download ICal for this eventBuilding Trust in AI: Faithful Reasoning with Language Models
Series: Department Seminar
Speaker: Dr. Debjit Paul, EPFL, Switzerland
Date/Time: Nov 04 11:00:00
Location: CSA Auditorium, (Room No. 104, Ground Floor)
Abstract:
Language models (LMs) have demonstrated impressive capabilities in reasoning tasks, particularly through approaches like chain-of-thought prompting, which involve generating intermediate inference steps. However, these intermediate inference steps may be inappropriate deductions from the initial context and lead to incorrect final predictions. In this talk, I will present a causal method to measure how reliably LMs use intermediate reasoning steps when generating an answer. Building on this work, I will present methods to improve the reasoning faithfulness of LMs. First, I will discuss a method, REFINER, a framework designed to fine-tune LMs to generate coherent and correct reasoning chains. Next, I will describe FRODO, a method that improves the faithfulness of LMs in reasoning over these reasoning steps to arrive at the correct answer consistently. These advancements have significant implications for improving LLMs?? reasoning capabilities across diverse applications.
Speaker Bio:
Debjit Paul obtained his Masters and Ph.D degrees from Saarland and Heidelberg Universities, Germany, in 2017 and 2022 respectively. Currently he is a post-doctoral researcher at EPFL, Switzerland. His research interests are in Natural Language Processing, Artificial intelligence, and Applied Machine Learning.
Host Faculty: R Govindarajan