Minhak Song

About Me

Hi! I am a first-year PhD student in Computer Science at Stanford University, working with Percy Liang.

Previously, I completed my undergraduate studies in Mathematical Sciences at KAIST, where I worked with Chulhee Yun. I was also a visiting student at the University of Washington, working with Simon Shaolei Du and Sewoong Oh.

My research interests center on the science of deep learning, with a current focus on pretraining. My earlier work focused on understanding optimization dynamics in deep learning.

Research Interests

• Science of Deep Learning
• Pretraining

News

• [Sep. 2026] I am starting my PhD in Computer Science at Stanford University.
• [May. 2026] I was selected as a Gold Reviewer (top 25% of reviewers) at ICML 2026.
• [Apr. 2026] Two papers (Zeroth-Order Edge of Stability, Dichotomy of RLHF and DPO) are accepted to ICML 2026.
• [Apr. 2026] I gave a contributed talk on Zeroth-Order Edge of Stability at the ICLR 2026 Workshop on Scientific Methods for Understanding Deep Learning in Rio, Brazil.
• [Feb. 2026] Our paper won the Best Student Paper Award at ALT 2026.
• [Jan. 2026] Our paper on the implicit bias of per-sample Adam on separable data is accepted to ICLR 2026.
• [Dec. 2025] Our paper on the theory of the spurious alignment of SGD in ill-conditioned high-dimensional quadratics is accepted to ALT 2026.
• [Oct. 2025] I was selected as a Top Reviewer (top 8% of reviewers) at NeurIPS 2025.
• [Sep. 2025] Our paper on understanding the benefit of Schedule-Free Optimizer through the river-valley loss landscape is accepted to NeurIPS 2025.
• [Jun. 2025] I joined Sewoong Oh’s group as a visiting student researcher at the University of Washington.
• [May. 2025] Our paper on how the datasets, network architectures, and optimizers influence progressive sharpening is accepted to ICML 2025.
• [Jan. 2025] Our paper on identifying the spurious alignment of SGD in an ill-conditioned valley (a.k.a. river-valley) loss landscape is accepted to ICLR 2025.
• [Jan. 2025] I joined Simon Shaolei Du’s group as a visiting student researcher at the University of Washington.
• [Jan. 2024] Our paper on the optimization characteristics of linear Transformers is accepted to ICLR 2024.
• [Sep. 2023] Our paper on understanding the Edge of Stability in deep learning is accepted to NeurIPS 2023.

Publications

(* denotes equal contribution)

Services

Conference/Workshop Reviewer

• Neural Information Processing Systems (NeurIPS) 2024-2025 (Selected as a Top Reviewer at NeurIPS 2025)
• International Conference on Machine Learning (ICML) 2025-2026 (Selected as a Gold Reviewer at ICML 2026)
• International Conference on Learning Representations (ICLR) 2025-2026
• International Conference on Artificial Intelligence and Statistics (AISTATS) 2025
• ICML 2025-2026 Workshop on High-dimensional Learning Dynamics
• ICML 2026 Workshop on Reinforcement Learning from World Feedback
• ICLR 2026 Workshop on Scientific Methods for Understanding Deep Learning