About me

I’m a fourth-year Ph.D. student at the Efficient Computing Lab in the Department of Computer Science & Engineering at POSTECH, advised by Prof. Eunhyeok Park. Before joining POSTECH, I completed my B.S. in the Department of Computer Science & Engineering at Kyung Hee University.

I’m currently focusing on Efficient AI, particularly in enhancing Memory Efficiency and Computation Acceleration during training and inference of various models (Vision, LLM, Physical AI, etc.) via Quantization and Low-rank Approximation.

Research Keywords

• Low-Precision Training (A01, A02, C03, P01, P05)
• Low-Precision Inference (C01, C02, P02)
• KV Cache Compression via Low-rank Approximation (P04, PT01)
• Parameter Efficient Fine-tuning of LLMs (P03)
• HW–SW Co-design via Kernel Optimization (P01, P02, P04, P05)

News

• [Nov. 03, 2025] I’ve joined AMD as a Research Associate.
• [Oct. 21, 2025] I’ve been selected as a Winner of Qualcomm Innovation Fellowship Korea.
• [Oct. 01, 2025] I’ve been selected for the BK21 Outstanding Graduate Student International Training Scholarship by the Ministry of Science and ICT, South Korea.
• [Mar. 02, 2025] 1 paper has been accepted to ICML 2025.
• [Feb. 25, 2025] 1 paper has been accepted to CVPR 2025.
• [Oct. 28, 2024] 1 paper has been accepted to WACV 2025.

Publications

• [A02] AdaHOP: Fast and Accurate Low-Precision Training via Outlier-Pattern-Aware Rotation
  Seonggon Kim, Alireza Khodamoradi, Pranathi Vasireddy, Kristof Denolf, Eunhyeok Park
  arXiv 2604.

• [C03] HOT: Hadamard-based Optimized Training
  Seonggon Kim, Juncheol Shin, Seung-taek Woo, Eunhyeok Park
  Computer Vision and Pattern Recognition (CVPR 2025), Nashville.

• [C02] Merge-Friendly Post-Training Quantization for Multi-Target Domain Adaptation
  Juncheol Shin, Minsang Seok, Seonggon Kim, Eunhyeok Park
  International Conference on Machine Learning (ICML 2025), Vancouver.

• [C01] PTQ4VM: Post-training Quantization for Visual Mamba
  Younghyun Cho*, Changhun Lee*, Seonggon Kim, Eunhyeok Park
  Winter Conference on Applications of Computer Vision (WACV 2025 Oral), Tucson.
  *Equal contribution.

• [A01] HLQ: Fast and Efficient Backpropagation via Hadamard Low-rank Quantization
  Seonggon Kim, Eunhyeok Park
  arXiv 2406.

Patents

• [PT01] Power Iteration Method Based on Hadamard PCA and KV Cache Compression Methodology Using the Same
  Seonggon Kim, Taehyeon Kim, Eunhyeok Park
  Korean Patent Application No. 10-2025-0189902, filed Dec. 03, 2025.

Projects

• [P05] Low-Precision Training on AMD GPU architecture, Nov. 2025 - May 2026
  Advanced Micro Devices
  • Conducted research on Low-Precision (MXFP4) Training on the AMD CDNA4 architecture.
  • Implemented ROCm kernel for training acceleration.
  • Achieved 3.6X memory compression, 1.46X acceleration over BF16.
  • Published in A02.

• [P04] Solutions for slow SVD approximation problem of KV Cache compression, Feb. 2025 - Present
  POSTECH
  • Conducted research on the slow SVD approximation problem during KV Cache compression.
  • Implemented CUDA kernel for SVD approximation acceleration.
  • Achieved 5.1X acceleration and 95% similarity with exact SVD operation.
  • Filed as a Korean patent application PT01.
  • This work is currently under review.

• [P03] Solutions for misaligned weight initialization problem of LoRA finetuning, Jun. 2025 - Present
  POSTECH
  • Conducted research on Effective LoRA weight initialization.
  • Achieved 99% accuracy of full fine-tuning with only rank 16.
  • This work is currently under review.

• [P02] GEMV Accelerator for LLM inference on Intel Gaudi-2, Jun. 2024 - Jun. 2025
  Naver & Intel Joint Research Center
  • Conducted research on fast LLM inference on the Intel Gaudi-2 architecture.
  • Implemented custom GEMV kernel for Gaudi with TPC-C language.
  • Transplanted LUT Quantization from CUDA to Gaudi TPC.

• [P01] Solutions for efficient training in limited GPU environments, Jun. 2023 - Present
  Ministry of Science and ICT of Korea
  • Conducted research on Low-Precision (INT4) Training in limited GPU environments.
  • Implemented CUDA kernel for training acceleration.
  • Achieved up to 75% memory savings and 2.6X acceleration over FP32.
  • Published in A01, C03.

Experience

• Research Associate, Nov. 2025 - May 2026
  Advanced Micro Devices, Longmont, CO, USA

• Software Engineer Intern, Jul. 2022 - Feb. 2023
  Spirent Communications, San Jose, CA, USA

• Software Engineer Intern, Feb. 2022 - Jun. 2022
  Common Computer, Seoul, Korea

• Research Intern, Mar. 2021 - Dec. 2021
  SI Analytics, Daejeon, Korea

Awards & Honors

• Qualcomm Innovation Fellowship Korea, Winner, Oct. 2025

• BK21 Outstanding Graduate Student International Training Scholarship, Recipient, Oct. 2025

• CVPR 2021 EarthVision Workshop, Land Cover Classification Challenge, 5th Prize, Jun. 2021

Education

• M.S./Ph.D. in Computer Science and Engineering, POSTECH
  Sep. 2023 - Present

• B.S. in Computer Science and Engineering, Kyung Hee University
  Mar. 2017 - Aug. 2023

Teaching Experience

• Teaching Assistant, Sep. 2026 - Dec. 2026
  CSED342: Artificial Intelligence, POSTECH

• Teaching Assistant, Mar. 2025 - Jun. 2025
  CSED311: Computer Architecture, POSTECH

Academic Service

• Reviewer, 2026
  Conference on Neural Information Processing Systems (NeurIPS 2026)