Donghyun Son

prof_pic.png

I’m an undergraduate student majoring Computer Science and Engineering at Seoul National University, and currently an exchange student at UT Austin. Here, I’m fortunate to work with Saurabh Agarwal for an efficient agentic system, advised by Aditya Akella. In SNU, I worked on quantization algorithms for KV cache, supervised by Sungjoo Yoo. Prior to that, I worked as a machine learning engineer at Hyperconnect (acquired by Match Group), where I built an ML-based content moderation system for Match Group brands.

I’m broadly interested in efficient algorithms for model training and inference. My previous works focus on data efficient methods including multiple subtasks approach, domain generalization, and few-shot personalization.

Recently, I’m interested in building an efficient LLM inference system. Particularly, I’m interested in the following topics.

  • KV cache compression: quantization, eviction
  • agentic systems: optimizing multi-agent systems
  • compressing the reasoning path

Aside from research, I enjoy algorithmic problem solving and have competed in competitive programming contests such as ICPC, Google Hashcode, and SCPC. You can find me on codeforces and BOJ.

I’m currently looking for job or Ph.D. opportunities that are aligned with my interests. Feel free to reach me!

Links: Github / CV / Google Scholar / X

 

selected publications

  1. Preprint
    NSNQuant.png
    NSNQuant: A Double Normalization Approach for Calibration-Free Low-Bit Vector Quantization of KV Cache
    Donghyun Son, Euntae Choi, and Sungjoo Yoo
    2025
    arXiv
  2. WSDM23 (Oral)
    wsdm23.png
    Reliable decision from multiple subtasks through threshold optimization: Content moderation in the wild
    Donghyun Son*, Byounggyu Lew*, Kwanghee Choi, and 5 more authors
    In Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, 2023
    arXiv
  3. OOD-CV@ICCV23
    ood-cv23.png
    Gradient estimation for unseen domain risk minimization with pre-trained models
    Byounggyu Lew*, Donghyun Son*, and Buru Chang
    In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023
    arXiv