Seil Kang
seil at stanford dot edu
seil at yonsei dot ac dot kr
I am a Ph.D. student in Computer Science at Yonsei University, advised by Seong Jae Hwang. Since Fall 2026, I have also been a Visiting Student Researcher at the Stanford AI Lab (SAIL), working with Azalia Mirhoseini. My visit to Stanford is supported by a research scholarship from HM Group.
I work on Machine Learning and Systems, focusing on making models train and run more efficiently, with a particular interest in hardware-centric approaches. Previously, I worked on the interpretability, reliability, and monitorability of vision and language models.
CV / Google Scholar / X / LinkedIn / GitHub
Experience
Stanford UniversityJul 2026 - Present
Visiting Student Researcher · Stanford AI Lab (SAIL)
Advisor: Azalia Mirhoseini
Education
Yonsei University
Ph.D. Student · CS
Advisor: Seong Jae Hwang- Yonsei University
B.S. · Chemical and Biomolecular Engineering
Open Source Contributions
News
- Jun. 2026I will join Stanford University, CS as a Visiting Student Researcher in Fall 2026.
- Jun. 2026Our paper Real-Time Visual Attribution Streaming in Thinking Model received the Best Paper Award at SCALE @ ICML.
- Jun. 2026Our paper Real-Time Visual Attribution Streaming in Thinking Model was selected as an ICML Main Track Spotlight.
- Jun. 2026Two papers of ours were accepted to ICML 2026.
- Feb. 2026Our two papers are accepted by CVPR 2026!
Older news
- Sep. 2025The paper Interpreting Attention Heads for Image-to-Text Information Flow in Large Vision-Language Models is accepted by NeurIPS 2025 Mechanistic Interpretability Workshop as a spotlight paper!
- Sep. 2025The paper Rare Text Semantics Were Always There in Your Diffusion Transformer is accepted by NeurIPS 2025!
- Aug. 2025Honored with the 2nd prize 2025 at AI Graduate School Symposium, hosted by South Korea's Ministry of Science and ICT!
- Feb. 2025The paper Your Large Vision-Language Model Only Needs A Few Attention Heads For Visual Grounding was selected as a highlight at CVPR 2025! (acceptance rate = 3.7%)
- Feb. 2025The paper Your Large Vision-Language Model Only Needs A Few Attention Heads For Visual Grounding is accepted by CVPR 2025!
- Feb. 2025The paper See What You Are Told: Visual Attention Sink in Large Multimodal Models is accepted by ICLR 2025!
Publications
- Real-Time Visual Attribution Streaming in Thinking Model
ICML 2026 (Spotlight, <2.2%) Best paper at SCALE Workshop in ICML
- Interpreting Attention Heads for Image-to-Text Information Flow in Large Vision-Language Models
NeurIPS 2025 Mechanistic Interpretability Workshop (Spotlight, <13%)
- Neuron-Level Approach for Multi-Hop Reasoning in Large Vision-Language Models
Technical Report
- Your Large Vision Language Model Only Needs A Few Attention Heads for Visual Grounding
CVPR 2025 (Highlight, <3%)
