Ph.D. in Artificial Intelligence
2024 — PresentYonsei University
Visual Intelligence Lab · Advisor: Prof. Youngjung Uh
Diffusion and flow-based generative models, identity-consistent generation, and predictive states for video world models.
CURRICULUM VITAE / 2026
Shinseong Kim.
Ph.D. Student in Artificial Intelligence
RESEARCH PROFILE
I study generative models through the lens of mathematics, with a focus on diffusion, flow matching, and predictive representations for physically plausible video world models.
Diffusion & flow matching / Controllable generation / Consistent generation / Video world models / Physical understanding / Robot Planning (VLA, WAM)
Education
M.S. in Mathematics
2021 — 2023Yonsei University
Advisor: Prof. Sungchul Lee
Thesis: Neural Tangent Matrix Factorization Recommendation System. Studied harmonic analysis and oscillatory integrals; explored neural tangent kernels and graph-based machine learning.
B.S. in Mathematics
2012 — 2019Chonnam National University
Gwangju, Republic of Korea
Research experience
Microsoft Research Asia
2026Research Intern · Multimodal Reasoning Group · Beijing, China
Worked with Jianlong Fu on video world models and future-predictive representations. Developed an internal predictive state for video diffusion transformers to improve the physical plausibility of generated futures and enable early candidate selection.
Visual Intelligence Lab
Aug — Dec 2023Research Intern · Yonsei University · Seoul, Republic of Korea
Honors & awards
Outstanding Reviewer
IEEE/CVF Conference on Computer Vision and Pattern Recognition
Published papers
TCFG: Tangential Damping Classifier-free Guidance
CVPR 2025
A geometric view of classifier-free guidance: damp misaligned tangential components to keep generation closer to its intended condition.
Preprints & manuscripts
World Models Need a Physical Registry: Learning Internal Predictive States for Physically Plausible World Models
arXiv 2026 · Submitted to ICLR 2027
An internal predictive state that evolves alongside video representations, bringing physically plausible futures into focus before denoising is complete.
When a Trigger Is Not a Key: Capability Leakage and Isolation in Personalized Diffusion Models
Manuscript · Submitted to NDSS 2027
Research on trigger conditioning, capability leakage, and isolation in personalized diffusion models.
ASemConsist: Adaptive Semantic Feature Control for Training-Free Identity-Consistent Generation
arXiv preprint · 2025
Training-free semantic feature control that balances a consistent identity with the freedom of each individual scene.
* Equal contribution. Submission does not imply acceptance.
Academic service
Conference reviewer
CVPR · NeurIPS · AAAI · ICLR · ICML
Languages
Korean · Native
English · Fluent
Japanese · Beginner
Last updated October 29, 2026