Byeonghu Na (나병후)

Biography

I am a Postdoctoral Researcher at KAIST, working with Prof. Il-Chul Moon at the KAIST Institute for Disruptive Robotics. I received my Ph.D. in Industrial & Systems Engineering from KAIST in 2025 under the supervision of Prof. Il-Chul Moon, following an M.S. degree in Industrial & Systems Engineering and B.S. degrees in Mathematical Sciences and Industrial & Systems Engineering from KAIST. I was a Research Intern at NAVER Clova AI Research. Starting in September 2026, I will join Stanford University as a Visiting Postdoctoral Scholar, working with Prof. Stefano Ermon, funded by the Sejong Science Fellowship. 

My research interests include generative models and AI alignment, with a focus on developing probabilistic models to capture diverse human perspectives. I aim to establish principled frameworks to characterize target distributions, optimization objectives, and divergence measures in complex real-world systems.

News

• [2026.05] Recognized as a Outstanding Reviewer at CVPR 2026. (Top 5%)

• [2026.05] Recognized as a Gold Reviewer at ICML 2026. (Top 25%)

• [2026.04] One paper (LiDAR) accepted at ICML 2026.  (Spotlight presentation, Top 2.2%)

• [2026.01] Four papers (WPR, AMiD, CSD, AC-Sampler) accepted at ICLR 2026.

• [2025.12] Received the 2025 KAIST Institutes Outstanding Researcher Award from KAIST Institutes.

• [2025.08] Three papers (DATE, STG, BPO) accepted at NeurIPS 2025.

• [2025.08] One paper (ImpConceptEmb) accepted at NeurIPS 2025 GenProCC Workshop.

• [2025.06] Recognized as an Outstanding Reviewer at CVPR 2025.

• [2025.03] Joined the KAIST Institute for Robotics as a Postdoctoral Researcher.

• [2025.02] Received the Best PhD Dissertation Award from KAIST College of Engineering.

• [2025.01] One paper (DBAE) accepted at ICLR 2025. (Spotlight presentation, Top 5%)

• [2024.12] DiffRS was awarded as one of the Winners of the Qualcomm Innovation Fellowship Korea (QIFK) 2024.

• [2024.12] Successfully defended my Ph.D. dissertation! (Title: Improving Conditional Information Estimation for Diffusion-based Generative Models)

Selected Publications

Semantic-aware Wasserstein Policy Regularization for Large Language Model Alignment (WPR)

• Byeonghu Na, Hyungho Na, Yeongmin Kim, Suhyeon Jo, HeeSun Bae, Mina Kang, and Il-Chul Moon

• ICLR 2026 (The Fourteenth International Conference on Learning Representations)

• [paper] [code]

Training-free Safe Text Embedding Guidance for Text-to-Image Diffusion Models (STG)

• Byeonghu Na, Mina Kang, Jiseok Kwak, Minsang Park, Jiwoo Shin, SeJoon Jun, Gayoung Lee, Jin-Hwa Kim, and Il-Chul Moon

• NeurIPS 2025 (The Thirty-ninth Annual Conference on Neural Information Processing Systems)

• [paper] [code]

Diffusion Adaptive Text Embedding for Text-to-Image Diffusion Models (DATE)

• Byeonghu Na, Minsang Park, Gyuwon Sim, Donghyeok Shin, HeeSun Bae, Mina Kang, Se Jung Kwon, Wanmo Kang, and Il-Chul Moon

• NeurIPS 2025 (The Thirty-ninth Annual Conference on Neural Information Processing Systems)

• [paper] [code]

Reward-based Input Construction for Cross-document Relation Extraction (REIC)

• Byeonghu Na*, Suhyeon Jo*, Yeongmin Kim, and Il-Chul Moon

• ACL 2024  (The 62nd Annual Meeting of the Association for Computational Linguistics) [Oral presentation, Top 4%]

• [paper] [code]

Diffusion Rejection Sampling (DiffRS)

• Byeonghu Na, Yeongmin Kim, Minsang Park, Donghyeok Shin, Wanmo Kang, and Il-Chul Moon

• ICML 2024 (The Forty-first International Conference on Machine Learning)

• Winner of the Qualcomm Innovation Fellowship Korea (QIFK) 2024

• Best Poster at the Samsung AI Forum 2024

• [paper] [code]

Label-Noise Robust Diffusion Models (TDSM)

• Byeonghu Na, Yeongmin Kim, HeeSun Bae, Jung Hyun Lee, Se Jung Kwon, Wanmo Kang, and Il-Chul Moon

• ICLR 2024 (The Twelfth International Conference on Learning Representations)

• [paper] [code]

Multi-modal Text Recognition Networks: Interactive Enhancements between Visual and Semantic Features (MATRN)

• Byeonghu Na, Yoonsik Kim, and Sungrae Park

• ECCV 2022 (European Conference on Computer Vision) 

• [paper] [code]