Jaehui Hwang

Research Scientist, Backbone Research, NAVER AI Lab

I am a Research Scientist in Backbone Research at NAVER AI Lab. I received my Ph.D. from Yonsei University in 2025, advised by Prof. Jong-Seok Lee at the MCML (Multimedia Computing and Machine Learning) Group, and my B.S. from Yonsei University in 2019.

My research focuses on understanding and improving the behavior of foundation models, including large language models. My recent work examines how LLMs learn to reason and how post-training shapes their behavior, including studies of on-policy distillation, reasoning traces, and the evaluation of reasoning and judgment. My broader interests include out-of-distribution generalization, robustness, and safety. During my Ph.D., I analyzed neural network vulnerabilities and how model architectures and input data shape model behavior.

Email Google Scholar GitHub LinkedIn CV (last updated: Dec 2024)

jaehui.hwang@navercorp.com · j.jaehui.hwang@gmail.com

News

Oct 2026One paper accepted to WACV 2027.
Oct 2026NP-OPD: On-Policy Distillation with Negative-Policy Rollouts is out. It introduces rollouts from a lower-capability negative policy while preserving the original distillation reward formulation.
Aug 2026Two papers accepted to EMNLP 2026 (Main and Findings).
Jul 2026On-Policy Delta Distillation (OPD²) is on arXiv. It uses a delta signal between the teacher and its base model to capture the changes induced by reasoning tuning.
Feb 2026One paper accepted to CVPR 2026.
2025Received my Ph.D. from Yonsei University and joined NAVER AI Lab as a Research Scientist.

Publications

Newest first · ^* equal contribution · ^† corresponding author

Detecting and Mitigating Memorization in Diffusion Models Using Special Tokens

J. Kim, J. Park, J. Hwang, J.-H. Kim, S. Kim, J.-S. Lee

ConferenceWACV 2027 (accepted)

On-Policy Distillation with Negative-Policy Rollouts

Jaehui Hwang^†, Dongyoon Han, Sangdoo Yun, Byeongho Heo^†

Preprint2026Project page arXiv Code

Revisiting Complete Reasoning Traces for Post-Training

Jaehui Hwang^†, Sangdoo Yun, Byeongho Heo, Dongyoon Han^†

ConferenceEMNLP 2026 FindingsarXiv Code

Beneath the Surface of Chains-of-Thought: A Mechanistic Interpretation of Reasoning Operations in LLMs

Seogyeong Jeong, Jaehui Hwang, Dongyoon Han, Geonmo Gu, Alice Oh^†, Taekyung Kim^†

ConferenceEMNLP 2026 MainarXiv Code

Beyond Detection: Analyzing and Classifying Global and Local Memorization in Diffusion Models

J. Kim, J. Park, J. Hwang, J.-S. Lee

ConferenceIEEE ICIP 2026

On-Policy Delta Distillation for Multilingual Math Reasoning

Byeongho Heo, Jaehui Hwang, Sangdoo Yun, Dongyoon Han

Preprint2026arXiv Code

On-Policy Delta Distillation

Byeongho Heo, Jaehui Hwang, Sangdoo Yun, Dongyoon Han

Preprint2026arXiv Code

MuCo: Multi-turn Contrastive Learning for Multimodal Embedding Model

Geonmo Gu, Byeongho Heo, Jaemyung Yu, Jaehui Hwang, Taekyung Kim, Sangmin Lee, HeeJae Jun, Yoohoon Kang, Sangdoo Yun^†, Dongyoon Han^†

ConferenceCVPR 2026arXiv Code

Oops, Wait: Discourse Tokens Matter in Reasoning Model

Jaehui Hwang, Byeongho Heo, Sangdoo Yun, Dongyoon Han

Preprint2026arXiv

What Defines Good Reasoning in LLMs? Dissecting Reasoning Steps with Multi-Aspect Evaluation

Heejin Do, Jaehui Hwang, Dongyoon Han, Seong Joon Oh, Sangdoo Yun^†

Preprint2025arXiv

Training-Free Framework for Defending Unsafe Image Synthesis Attack

Junha Park, Jaehui Hwang, Ian Ryu, Hyungkeun Park, Jiyoon Kim, Jong-Seok Lee

ConferenceIEEE ICASSP 2026arXiv

Anomaly Score: Evaluating Generative Models and Individual Generated Images based on Complexity and Vulnerability

Jaehui Hwang^*, Junghyuk Lee^*, Jong-Seok Lee

ConferenceCVPR 2024Paper arXiv

Jaehui Hwang, Dongyoon Han, Byeongho Heo, Song Park, Sanghyuk Chun^†, Jong-Seok Lee^†

ConferenceECCV 2024arXiv Code

Classifying Gas Data Measured Under Multiple Conditions Using Deep Learning

Hojung Lee, Jaehui Hwang, Hwin Dol Park, Jaehun Choi, Jong-Seok Lee

JournalIEEE Access 2022Paper

Temporal Shuffling for Defending Deep Action Recognition Models against Adversarial Attacks

Jaehui Hwang, Huan Zhang, Jun-Ho Choi, Cho-Jui Hsieh, Jong-Seok Lee^†

JournalNeural Networks 2024Paper arXiv Code

Just One Moment: Structural Vulnerability of Deep Action Recognition against One Frame Attack

Jaehui Hwang, Jun-Hyuk Kim, Jun-Ho Choi, Jong-Seok Lee

ConferenceICCV 2021Paper arXiv

Universal Kriging for Loran ASF Map Generation

Pyo-Woong Son, Jun Hyo Rhee, Jaehui Hwang, Jiwon Seo

JournalIEEE Transactions on Aerospace and Electronic Systems 2018Paper

On the Repeatability of EEG-based Image Quality Assessment

Jaehui Hwang, Seong-Eun Moon, Jong-Seok Lee

ConferenceIEEE SMC 2018Paper

TDOA-based ASF Map Generation to Increase Loran Positioning Accuracy in Korea

Jaehui Hwang, Pyo-Woong Son, Jiwon Seo

ConferenceIEEE ICCE-Asia 2018Paper

Experience

Research ScientistBackbone Research, NAVER AI Lab2025 – present
Research InternML Research, NAVER AI Lab
Advisors: Sanghyuk Chun, Dongyoon HanMar 2022 – Sep 2022

Education

Ph.D.School of Integrated Technology, Yonsei University
Advisor: Prof. Jong-Seok Lee2019 – 2025
B.S.School of Integrated Technology, Yonsei University
Full scholarship (IITP)2016 – 2019

Academic activities

Award: Merit Academic Paper Award, 2022-2 Yonsei Superior Paper Awards, Dec 2022
Reviewer: ICLR2025–2026NeurIPS2025–2026ICML2025–2026ACL ARR2025–2026CVPR2022–2024ICCV2023, 2025ECCV2022, 2024TMLR2023–2025
Talks: “Just One Moment: Structural Vulnerability of Deep Action Recognition against One Frame Attack,” Korean Conference on Computer Vision (KCCV), 2022