JuHeon's DevStories only I can tell

하주헌 JuHeon Ha

GitHubLinkedin

I am an M.S. student in Artificial Intelligence at Ajou University.

Profile picture

My research focuses on LLM interpretability, human-centered AI, and AI safety and alignment, with particular interests in persona, emotion, and social bias in LLMs. I explore how AI can model human psychology and behavior, and the underlying model mechanisms that give rise to human-like behavior.

Before beginning my graduate studies, I worked as a frontend developer.

  • LLM Interpretability
  • LLM Safety & Alignment
  • Representation Engineering
  • Activation Steering
  • Human-Centered AI

Publications

EMNLP 2026 Main Conference

Accepted

, , ,

  • LLM Empathy
  • Activation Steering
  • Mechanistic Interpretability
Abstract

This work investigates the activation-level structure of supportive empathy in instruction-tuned LLMs. Using the EPITOME framework and contrastive activation steering, we examine the controllability and representational geometry of emotional reactions, interpretations, and explorations, as well as their interaction with persona-induced activation shifts.

Project PagePaper (arXiv)GitHubRepository

Under Review

2 papers

Nudge and Merge: Simple Post-Hoc Sparsification of Dense Adapters for Parameter-Efficient Fine-Tuning on Sparse LLMs

AAAI 2027Second AuthorUnder review

  • Sparse LLMs
  • Parameter-Efficient Fine-Tuning
  • Adapter Merging

Do LLMs Empathize with You, or with Your Words? Revealing How Stated Distress Makes a Friend's Loss Seem Yours

ACL Rolling Review (ARR)First AuthorUnder review

  • LLM Empathy
  • Person-Event Attribution
  • Emotional Cues

Articles