-
PhD Thesis Proposal - James Huang
Fri, May 09, 2025 @ 03:00 PM - 04:30 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Collaborative Decision-Making of Language Models
Date and Time: Friday, May 9th, 2025 | 3:00p - 4:30p
Location: GCS 502C
Committee Members: Muhao Chen, Fred Morstatter, Laurent Itti, Robbin Jia, Dan O'Leary
Abstract: While general-purpose language models have demonstrated strong performance on a wide range of tasks, they still have their own weaknesses such as biases, misalignment, lack of task-specific knowledge, etc. One promising way of addressing these challenges is to combine the strengths of different language models. In this proposal, I will outline my research exploring various strategies to facilitate collaborative decision making of language models. Specifically, I will present 1) a shortcut mitigation method via ensemble-based attention debiasing, 2) a decoding-time alignment framework that uses model-based reward functions to guide model generation, and 3) an unlearning method that removes sensitive knowledge by learning a logit offset. Finally, I will discuss future directions for language model collaboration.Location: Ginsburg Hall (GCS) - 502C
Audiences: Everyone Is Invited
Contact: James Huang
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor.