PhD Dissertation Defense - Zihao He

Title: Aligning Large Language Models with Human Perspectives
 
Date & Time: Wednesday, February 5th - 12:00p - 2:00p
 
Location: RTH 306  
 
Committee: Kristina Lerman (Chair, CS), Emilio Ferrara (CS), Marlon Twyman (Communication)  
 
Abstract: Large Language Models (LLMs) are increasingly deployed in real-world applications. However, their ability to accurately represent diverse human perspectives remains a critical challenge. This thesis investigates LLM alignment, which refers to how closely these models reflect the ideologies, values, and communication styles of specific communities. First, I develop methods for aligning LLMs to online communities and introduce Community-Cross-Instruct, a framework that generates structured instruction-answer pairs to enhance fidelity and scalability. Second, I propose comprehensive evaluation frameworks to assess alignment beyond positional stances, including affective alignment (how well LLMs capture emotional and moral tones) and multidimensional evaluations across authenticity, toxicity, and harm. Finally, I explore ethical risks in alignment, demonstrating how minimal biased data during instruction tuning can shift an LLM’s behavior, raising concerns about ideological manipulation. These findings highlight the technical, evaluation, and ethical complexities of LLM alignment, providing a foundation for ensuring that LLMs reflect diverse human perspectives and stay robust to ideological manipulation.  
 
Zoom Link: https://usc.zoom.us/j/97020518118?pwd=mZeDv2WhswDGTouNvvWFI9NFqhO5KR.1

Location: Ronald Tutor Hall of Engineering (RTH) - 306

Audiences: Everyone Is Invited

Contact: Zihao He