Dissertation Title:  Fostering Pluralism with Language Models

Abstract: 

Social polarization stands as one of the most pressing challenges of our time. The widespread use of social media has intensified extreme polarization. Pluralism offers a response to this growing polarization by creating meaningful engagement across differences. Unlike mere tolerance, pluralism acknowledges real differences while seeking common ground through constructive dialogue. This thesis explores how Large Language Models (LLMs) can foster pluralism through a progressive framework of detecting polarization, modeling diversity, and cultivating engagement. First, we demonstrate LLMs' ability to detect polarization through CommunityLM, which probes partisan worldviews, and Bridging Dictionary, which identifies polarized language use in public discourse. Second, we develop the PersonaLLM framework to understand LLMs' capabilities in modeling diverse social perspectives, especially in role-playing and storytelling. Finally, we leverage these insights to cultivate engagement in the form of constructive dialogue, a key element of pluralism that reveals both common understandings and meaningful differences. Our Legal Storytelling work transforms complex laws into accessible narratives, creating common ground for diverse audiences to engage with specialized knowledge. In addition, FaciliTrain builds on LLMs' role-playing capabilities to create a virtual environment where facilitators can practice guiding conversations with real-time feedback that bring diverse perspectives into mutual engagement. Together, these approaches show how LLMs can cultivate engagement that transforms diversity into pluralism, establishing a framework for bridging societal divides and fostering the mutual understanding essential for a pluralistic society.

Committee members:

Deb Roy (Advisor), Professor of Media Arts and Sciences, MIT
Yoon Kim, Assistant Professor of EECS, MIT
Diyi Yang, Assistant Professor of Computer Science, Stanford University