I model how people disagree, then use that to personalize and align LLMs.

Hi! I’m a fourth-year PhD student at the University of Michigan School of Information (UMSI), advised by David Jurgens in the Blablablab. I received a BA in Statistics and Data Science from UC Berkeley in 2023.

My research is on pluralistic alignment: modeling how individuals and groups differ in their judgments, and using those representations to personalize and align LLMs. Standard practice aggregates labels into a single “gold standard,” treating disagreement as noise. On subjective tasks, though, disagreement is often meaningful, rooted in who the annotator is and how the task is framed. I build models that preserve it instead of averaging it away.

• DEM-MoE models annotator disagreement with demographic-aware experts that capture structured, group-level variation in judgments (ACL 2026, oral).
• VISTA, built during my R&D internship at Kitware, infers a user’s values from a few survey answers and steers a frozen LLM toward them, beating prompt-based personalization with no added prompt tokens (under review, ARR).

Currently

I’m building B-HAP, a hierarchical Bayesian framework that generates per-annotator LoRA adapters from task, data, and annotator features, pooling behavior across population, subgroup, and individual levels — extending the same pluralistic-alignment approach behind VISTA to richer, adapter-based personalization.