Soo Min Kwon

SooMin.jpg

I am a final year Ph.D. student in the Department of Electrical and Computer Engineering at the University of Michigan, Ann Arbor, advised by Prof. Laura Balzano and Prof. Qing Qu. Previously, I received my M.S. and B.S. degrees from Rutgers University, where I worked with Prof. Anand D. Sarwate.

I am interested in a wide range of problems, from theoretical deep learning to practical and efficient algorithms for generative models. I have worked on post-training algorithms for LLM reasoning tasks, theoretical analyses of in-context learning, and diffusion models for solving inverse problems.

[CV] [Google Scholar]

news

May 07, 2026 Our paper, Out-of-Distribution Generalization of In-Context Learning: A Low-Dimensional Subspace Perspective, won the best paper award at the AISTATS 2026 Workshop for Causality in the Age of AI Scaling!
Apr 01, 2026 My internship work at Google Research, CoDistill-GRPO: A Co-Distillation Recipe for Efficient Group Relative Policy Optimization, is now released!
Jan 22, 2026 A paper, Out-of-Distribution Generalization of In-Context Learning: A Low-Dimensional Subspace Perspective, is accepted to AISTATS 2026!

selected publications

  1. AISTATS 2026
    Out-of-Distribution Generalization of In-Context Learning: A Low-Dimensional Subspace Perspective
    Soo Min Kwon*, Alec S. Xu*, Can Yaras, Laura Balzano, and Qing Qu
    International Conference on Artificial Intelligence and Statistics (AISTATS), 2026
    arXiv
  2. ICLR 2025
    Learning Dynamics of Deep Matrix Factorization Beyond the Edge of Stability
    Avrajit Ghosh*, Soo Min Kwon*, Rongrong Wang, Saiprasad Ravishankar, and Qing Qu
    The Thirteenth International Conference on Learning Representations (ICLR), 2025
    arXiv
  3. NeurIPS 2024
    BLAST: Block-Level Adaptive Structured Matrices for Efficient Deep Neural Network Inference
    Changwoo Lee, Soo Min Kwon, Qing Qu, and Hun-Seok Kim
    The Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS), 2024
    arXiv
  4. AISTATS 2024
    Efficient Compression of Overparameterized Deep Models through Low-Dimensional Learning Dynamics
    Soo Min Kwon, Zekai Zhang, Dogyoon Song, Laura Balzano, and Qing Qu
    International Conference on Artificial Intelligence and Statistics (AISTATS), 2024
    arXiv
  5. ICLR 2024
    Solving Inverse Problems with Latent Diffusion Models via Hard Data Consistency
    Bowen Song*, Soo Min Kwon*, Zecheng Zhang, Xinyu Hu, Qing Qu, and Liyue Shen
    The Twelfth International Conference on Learning Representations (ICLR Spotlight, Top 5%), 2024
    arXiv