Byeonghyun Pak

Research Officer
Agency for Defense Development
Daejeon, Republic of Korea

byeonghyun.pak@gmail.com
github.com/ByeongHyunPak
@Byeonghyun_Pak
Google Scholar

I am a Research Officer at the Agency for Defense Development, Korea's national defense R&D agency (akin to DARPA) and a First Lieutenant in the Republic of Korea Army. I received my B.S. from DGIST's Interdisciplinary Program, concentrating in Computer Science & Engineering.

Research Interests

My research goal is to develop robots with human-level versatility that can (1) generalize across diverse tasks and environments, (2) adapt to unseen domains via physics-aware reasoning, and (3) deploy in dynamic, real-world settings.

To this end, my research focuses on three core areas:

• Vision–Language Representations (ECCV 2024)

for open-world semantics and domain generalization.

• Neural Scene Dynamics (ongoing)

for learning physics-aware policies.

• Implicit Neural Representations (CVPR 2023)

for continuous scene representations from sensor observations.

Publications

NeurIPS 2025

Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration

Yunghee Lee, Byeonghyun Pak, Junwha Hong, Hoseong Kim

Paper

ECCV 2024

Textual Query-Driven Mask Transformer for Domain Generalized Segmentation

Byeonghyun Pak*, Byeongju Woo*, Sunghwan Kim*, Dae-hwan Kim, Hoseong Kim

Paper Page Code

B-spline Texture Coefficients Estimator for Screen Content Image Super-Resolution

CVPR 2023

🏆 Highlight (Top 2.5%)

B-spline Texture Coefficients Estimator for Screen Content Image Super-Resolution

Byeonghyun Pak*, Jaewon Lee*, Kyong Hwan Jin

Paper Page Code

Ongoing Projects

Active

Dynamics-Aware Policy Learning

We learn policies that model the underlying dynamics of an environment to forecast short-horizon. This look-ahead enables anticipatory behavior such as intercepting moving targets.

Submitted to ICLR 2026

Anonymous Submission

Byeongju Woo, Zilin Wang, Byeonghyun Pak, Sangwoo Mo, Stella X. Yu

We propose a image-text representation learning method that enables fine-grained multimodal understanding without extra annotations.

Experience

First Lieutenant

Mar 2023 – Present

Republic of Korea Army (ROKA)

Selected as one of 20 research officers nationwide dedicated to science-and-technology R&D for national defense

Research Officer for National Defense

Mar 2023 – Present

Agency for Defense Development (ADD)

Manager: Dr. Eunjin Koh/Advisor: Dr. Hoseong Kim

Investigated and improved domain generalization for reliable infrared imagery object detection in data-scarce settings (1 publication in ECCV 2024).
Constructed synthetic datasets for rare/low-visibility targets via diffusion models and accelerated the generation process (1 publication in NeurIPS 2025).

Undergraduate Research Intern

Dec 2021 – Feb 2023

Image Processing Laboratory @ DGIST

Advisor: Prof. Kyong Hwan Jin

Researched implicit neural representations (INRs) for solving inverse problems and proposed a B-Spline INR super-resolution algorithm (1 publication in CVPR 2023).