Event Details

PhD Dissertation Defense - Nan Xu

Thu, Jun 26, 2025
11:00 AM - 1:00 PM
Location: RTH 306

Title: Building Trustworthy LLMs: Ensuring Inference-Time Coherence and Factuality


 
Date: Thursday, June 26th, 2025 | 11:00am-1:00pm
 

Venue: RTH 306 and Zoom <a href="https://usc.zoom.us/j/93200441032?pwd=7QqueUvIVf0WXl2LQ7AELW7ix31dNz.1 " target="_blank" class="vsoe-inline">https://usc.zoom.us/j/93200441032?pwd=7QqueUvIVf0WXl2LQ7AELW7ix31dNz.1 </a>
 
 

 


Committee Members: Xuezhe Ma (Chair), Muhao Chen, Jonathan May, Daniel E. O'Leary, Ram Nevatia 
 

 

Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in generating text, yet critical issues such as hallucinations and precision errors significantly impact their reliability in high-stakes tasks, warranting further in-depth research. In my thesis, I will first introduce my research aiming to tackle these challenges from the perspective of decoding, which is train-free and driven by models' own understanding of seen and generated texts. Specifically, I focus on 1) reducing undesired repetitions and off-topic generations by analyzing probability distribution of decoding steps for open-ended text generation and 2) mitigating hallucinations by studying models' uncertainty against user prompts for false-premise question answering. 


 


Motivated by the in-context learning (ICL) capabilities of Large Language Models (LLMs), multimodal LLMs incorporating an additional visual modality have demonstrated similar ICL abilities when provided with multiple image-text pairs as demonstrations. As the final part of this thesis, I will investigate whether multimodal LLMs can reliably perform a broad range of tasks without additional fine-tuning, including tasks that were not encountered during pretraining or that may even conflict with the pretraining data.