LLM Reasoning & Alignment

Preference learning, personalization, and long-horizon reasoning in large language models.

Overview

How can an LLM understand what a user wants when their preferences are incomplete, noisy, or scattered across conversations?

My research explores personality as a latent signal behind user preferences, using it to guide personalized responses and generalize to unseen choices.

PACIFIC: Personality-Driven Preference Alignment in LLMs

Status: Research at UC Irvine Co-first author
→ Key insight: Personality traits can help organize preference evidence and identify which context is useful for personalized decisions.

What we’re building:

  • PACIFIC benchmark: 1,200 synthetic preference–query pairs across 20 domains, annotated with Big Five personality traits.

  • Persona-aware retrieval: A contrastively fine-tuned dual-encoder retriever that selects personality-consistent preferences for LLM responses.

  • Controlled evaluations: Experiments examining trait alignment, noisy context, and personality guidance.

Key result: On PACIFIC, Gemma-3-4B-IT achieved 76% answer-choice accuracy with trait-aligned, personality-labeled preferences, compared with 29.25% using mixed-trait preferences

Why it matters: Personalization depends on retrieving useful evidence and applying it appropriately. This work connects user modeling, representation learning, and retrieval-augmented generation.

Future Research Directions

  1. Preference generalization — How can past interactions inform recommendations in unfamiliar domains?
  2. Robust user modeling — How should models handle uncertain or contradictory preference evidence?
  3. Adaptive personalization — How can user representations evolve as preferences change?

Paper & Resources

← Back to home