Test-time adaptation · Medical imaging

Hyeongyu Kim

Compiler Engineer · Hyundai Motor Company

My research focuses on test-time adaptation: how models can adapt when the data they encounter differs from their training data. I also work on domain adaptation and generalization, and their applications to medical image segmentation and reconstruction.

I am a Compiler Engineer in the Compiler Team at Hyundai Motor Company, where I work on TVM-based compilation for custom NPUs. I received my Ph.D. in Electrical and Electronic Engineering from Yonsei University in August 2026, advised by Prof. Dosik Hwang at the Medical Artificial Intelligence Lab.

Portrait of Hyeongyu Kim
Ph.D., Yonsei University

Selected publications

All publications
  1. NeurIPS 2026 Accepted

    MIRA: Mutual Information guided calibration for Reliable Test-Time Adaptation

    Hyeongyu Kim* , Mingyeong Jang* , Youngjun Song , GeonHui Han , Dosik Hwang

    * Equal contribution

    Paper
  2. CVPR 2026

    AcTTA: Rethinking Test-Time Adaptation via Dynamic Activation

    Hyeongyu Kim* , Geonhui Han* , Dosik Hwang

    * Equal contribution

    Adapt activation thresholds and gradient sensitivity at test time, while keeping the pretrained network weights fixed.

    AcTTA activation and derivative curves for ReLU and GELU parameterizations.
    Paper Project
  3. NeurIPS 2025

    Buffer layers for Test-Time Adaptation

    Hyeongyu Kim* , Geonhui Han* , Dosik Hwang

    * Equal contribution

    Insert small, trainable Buffer layers into a frozen backbone to adapt to new domains without updating its original parameters.

    Buffer layers inserted into a frozen backbone, compared with normalization updates and full fine-tuning.
    Paper Code

Recent updates

Archive

  1. Our paper MIRA has been accepted at NeurIPS 2026!

  2. Our paper SEG.A challenge overview is available online in Medical Image Analysis.

  3. I received my Ph.D. from Yonsei University, advised by Prof. Dosik Hwang.

  4. In August 2026, I joined Hyundai Motor Company as a Compiler Engineer in the Compiler Team, working on NPU compilation.

  5. I was recognized as an Outstanding Reviewer at CVPR 2026.

Explore

Interactive lab (Work in progress)
Compare Tent, Buffer, and AcTTA on the same inputs.
Research notes (Work in progress)
Short explanations of adaptation, tensor programs, and memory reuse.
Now
Current work, interests, and reading.
Get in touch
Email, profile links, and a QR contact card.

Serving as a reviewer

Conferences
CVPR 2026, 2025 · Outstanding Reviewer (2026)
NeurIPS 2026
MICCAI 2025
ICPR 2024
Journals
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)
Medical Image Analysis
Biomedical Engineering Letters (BMEL)