Hans Hao-Hsun Hsu

me.jpg

Hi, my name is Hans Hao-Hsun Hsu. I am Machine Learning PhD student at Georgia Tech working with Pan Li.

My current research interests lie in (1) large language models as agents and multi-agent systems (2) geometric deep learning and its application to scientific discovery/engineering (3) trustworthy AI on graphs and agents (reliability, robustness and explainability).

Prior to Georgia Tech, I was a machine learning researcher at Celeris Therapeutics, an Austria- and UK-based startup, working on PROTACs-related drug discovery projects. I obtained my M.S. degree (with high distinction) from the Technical University of Munich advised by Daniel Cremers, where I worked on uncertainty estimation in geometric deep learning. I also hold a B.S. degree from National Taiwan University.

Feel free to reach out to me for any collaboration or research opportunities.

selected publications

  1. llm_forgot.png
    Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness
    Rongzhe Wei, Peizhi Niu, Hans Hao-Hsun Hsu, Ruihan Wu, Haoteng Yin, Mohsen Ghassemi, Yifan Li, Vamsi K Potluru, Eli Chien, Kamalika Chaudhuri, Olgica Milenkovic, and Pan Li
    In Preprint , 2025
    Paper Code
  2. tsa.png
    Structural Alignment Improves Graph Test-Time Adaptation
    Hans Hao-Hsun Hsu, Shikun Liu, Han Zhao, and Pan Li
    In Preprint , 2025
    Paper Code
  3. GATS_new.png
    What Makes Graph Neural Networks Miscalibrated?
    Hans Hao-Hsun Hsu, Yuesong Shen, Christian Tomani, and Daniel Cremers
    In Neural Information Processing Systems (NeurIPS) , 2022
    Paper Code
  4. graph_ece.png
    A Graph Is More Than Its Nodes: Towards Structured Uncertainty-Aware Learning on Graphs
    Hans Hao-Hsun Hsu, Yuesong Shen, and Daniel Cremers
    In NeurIPS GLFrontiers , 2022
    Paper Code
  5. vae_ad_framework.png
    Automated Antenna Testing Using Encoder-Decoder-based Anomaly Detection
    Hans Hao-Hsun Hsu, Jiawen Xu, Ravi Sama, and Matthias Kovatsch
    In International Conference on Machine Learning and Applications (ICMLA) , 2021
    Paper

other research projects

  1. protac.png
    GraphPROTACs: Out Of Distribution DC50 Prediction
    Celeris Therapeutics, 3.2023 - 1.2024

    Curated a PROTAC dataset and proposed GraphPROTACs, achieving state-of-the-art generalizability (AUROC increased from 55% to 75%) and enhanced model interpretability, with further improvements from training on synthetic data.

    Paper
  2. pong.gif
    Adversarial Attacks on Reinforcement Learning Agents Trained with Self-Play
    Technical University of Munich, 11.2020 - 3.2021

    Researched training RL agents in multi-agent environments and examined their robustness through observation-based and adversarial policy attacks.

    Paper Code
  3. IMG_6989.png
    Neural Network Interpretability
    Technical University of Munich, 4.2020 - 8.2020

    Researched deep learning interpretability methods and explored their ability to interpret neural network prediction uncertainty.

    Paper Code