Hi! I am a Ph.D. student at KAIST School of Computing, advised by Prof. Alice Oh.
I study how reasoning works internally in language models and how that understanding can enable better intervention and control. My current research focuses on reasoning representations, mechanistic interpretability, and latent reasoning.
Broadly, I am interested in three connected questions:
My long-term goal is to build language models capable of increasingly sophisticated reasoning while remaining reliable, interpretable, and controllable, enabling people to use AI safely and confidently.
Much of the work on LLM reasoning focuses on the reasoning that models produce as text. I am interested in what happens beneath the output: how reasoning is represented internally, what distinguishes different reasoning processes, and what separates successful from failed computation.
In Beneath the Surface of Chains-of-Thought,[1] we study whether different reasoning operations are distinguishable inside language models beyond what can be observed from their generated text. We find that reasoning operations exhibit distinct representation-level signatures, particularly in intermediate layers, suggesting that internal representations can reveal aspects of reasoning that are not apparent from surface-level outputs.
This motivates the direction I am pursuing now: if useful information about reasoning exists in latent representations, can we use that information not only to understand reasoning, but also to guide and improve it? I am particularly interested in soft-token-based latent reasoning and, more broadly, in model-level interventions that make reasoning more reliable and controllable.
My path toward these questions began with behavior-level evaluation across multilingual, cross-cultural, and localization settings.[2][3][5] These projects gradually led me toward questions that behavioral evaluation alone cannot answer: why behavioral differences emerge, how they are represented internally, and how manipulating internal representations can alter downstream behavior.[4][6] This progression—from behavior to internal mechanisms and then to intervention—now shapes how I think about reasoning in language models.
[1] Seogyeong Jeong, Dongyoon Han, Jaehui Hwang, Geonmo Gu, Alice Oh, Taekyung Kim. 2026.
Beneath the Surface of Chains-of-Thought: A Mechanistic Interpretation of Reasoning Operations in LLMs.
EMNLP 2026 Main Conference (Long). [Demo]
[2] *Seogyeong Jeong, *Kiwoong Park, Seyoung Song, Eunsu Kim, Ken E. Friedl, Jaeho Kim, Alice Oh. 2026.
LoCar: Localization-Aware Evaluation of In-Vehicle Assistants through Fine-Grained Sociolinguistic Control.
ACL 2026 Industry Track.
[3] *Seyoung Song, *Seogyeong Jeong, Eunsu Kim, Jiho Jin, Dongkwan Kim, Jay Shin, Alice Oh. 2025.
MUG-Eval: A Proxy Evaluation Framework for Multilingual Generation Capabilities in Any Language.
Findings of EMNLP 2025. Also presented at the 5th Workshop on Multilingual Representation Learning (MRL 2025).
[4] Soyeon Park, *Seogyeong Jeong, *Sunwoo Kim, Alice Oh. 2026.
The Role of Fine-grained Harm Signals in LLM Safety.
NeurIPS 2026 Workshop – Interpretability for Science.
[5] *Junyeong Park, *Seogyeong Jeong, *Seyoung Song, Yohan Lee, Alice Oh. 2025.
LLM-C3MOD: A Human-LLM Collaborative System for Cross-Cultural Hate Speech Moderation.
NAACL 2025 Workshop – C3NLP. [Slides] [Poster]
[6] Haeun Yu, Arnav Arora, Seogyeong Jeong, Nadav Borenstein, Siddhesh Pawar, Jisu Shin, Jiho Jin, Junho Myung, Alice Oh, Isabelle Augenstein. 2025.
Entangled in Representations: Mechanistic Investigation of Cultural Biases in Large Language Models.
arXiv preprint.
[7] Paul Röttger, Giuseppe Attanasio, Felix Friedrich, Janis Goldzycher, Alicia Parrish, Rishabh Bhardwaj, Chiara Di Bonaventura, Roman Eng, Gaia El Khoury Geagea, Sujata Goswami, Jieun Han, Dirk Hovy, Seogyeong Jeong, Paloma Jeretič, Flor Miriam Plaza-del-Arco, Donya Rooein, Patrick Schramowski, Anastassia Shaitarova, Xudong Shen, Richard Willats, Andrea Zugarini, Bertie Vidgen. 2025.
MSTS: A Multimodal Safety Test Suite for Vision-Language Models.
arXiv preprint.
* Indicates equal contribution.