Hsun-Yu Kuo

I am an M.Sc. student in Computer Science at EPFL, specialising in computer science theory. My research interests include data-centric AI, reinforcement learning, robustness, extrapolation, and causality. More broadly, I am interested in how diversity can help shape better AI systems and a better society.

I am currently a project student at the MLO Lab at EPFL, where I work with Martin Jaggi and El Mahdi Chayti. Previously, I interned with the Robust Machine Learning Group at the Max Planck Institute for Intelligent Systems and the CKIP Lab at Academia Sinica.

Feel free to reach out if you would like to discuss research or potential collaboration.

News

Sep 2026	NeurIPS 2026: Our paper, “Stabilizing Extrapolation in Looped Transformers via Learned Stochastic Stopping”, has been accepted to the main track!
Aug-Sep 2026	I attended MLSS in Tübingen, Germany.
Jul 2026	I presented our work on looped transformers as a poster at the ICML 2026 AdaptFM workshop.
Aug 2025	Gave a talk on synthetic data weighting at the Guided Generation Group (GGG).
Jan 2025	ICLR 2025: Our paper, “Not All LLM-Generated Data Are Equal”, was accepted and selected as a Spotlight!

Selected Publications

NeurIPS ’26

Stabilizing Extrapolation in Looped Transformers via Learned Stochastic Stopping

Hsun-Yu Kuo, El Mahdi Chayti, Patrik Reizinger, and 2 more authors

Advances in Neural Information Processing Systems, 2026

Main track. Also presented as a poster at the ICML 2026 AdaptFM workshop.

Bib Website

@misc{kuo2026stabilizingextrapolationloopedtransformers,
  title = {Stabilizing Extrapolation in Looped Transformers via Learned Stochastic Stopping},
  author = {Kuo, Hsun-Yu and Chayti, El Mahdi and Reizinger, Patrik and Brendel, Wieland and Jaggi, Martin},
  year = {2026},
  eprint = {2606.29983},
  archiveprefix = {arXiv},
  primaryclass = {cs.LG},
  url = {https://arxiv.org/abs/2606.29983},
}

ICLR ’25

Spotlight

Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification

Hsun-Yu Kuo, Yin-Hsiang Liao, Yu-Chieh Chao, and 2 more authors

In International Conference on Learning Representations, 2025

Spotlight (top 5.1%)

Bib Website

@inproceedings{ICLR2025_03dfa2a7,
  author = {Kuo, Hsun-Yu and Liao, Yin-Hsiang and Chao, Yu-Chieh and Ma, Wei-Yun and Cheng, Pu-Jen},
  booktitle = {International Conference on Learning Representations},
  editor = {Yue, Y. and Garg, A. and Peng, N. and Sha, F. and Yu, R.},
  pages = {1161--1185},
  title = {Not All {LLM}-Generated Data Are Equal: Rethinking Data Weighting in Text Classification},
  url = {https://proceedings.iclr.cc/paper_files/paper/2025/file/03dfa2a7755635f756b160e9f4c6b789-Paper-Conference.pdf},
  volume = {2025},
  year = {2025},
}