Ph.D. in Intelligence and Information
Seoul National University
Ph.D. DissertationProjection, Persistence, and Elicitation of Learned Internal Structure
Best Ph.D. Dissertation AwardGSCT representative awardee · 1 recipient among 27 Ph.D. graduates
Junhoo Lee
About Me
I view AI through the lens of self-reference. Many of its remarkable capabilities emerge from recursive structures in which a system's own outputs, representations, or experiences feed back into its subsequent operation. I am fascinated by why these structures appear across different levels of AI, how they give rise to new capabilities, and what makes them work.
My current research focuses on robot harnesses, looped models, and continual harnessing, while I remain open to exploring other directions. I am a postdoctoral researcher at KAIST, working with Prof. Jinwoo Shin, and received my Ph.D. from Seoul National University under Prof. Nojun Kwak.
Education
B.Sc. in Electrical and Computer Engineering
Seoul National University
B.S. ThesisImprovement of Elevator Control Algorithm Using Reinforcement Learning
News
Three papers accepted to NeurIPS 2026: Discriminative Score Function, Clean Data Can Still Carry Backdoors, and ManiFusion!
I received the Best Ph.D. Dissertation Award as the representative awardee of the SNU Graduate School of Convergence Science and Technology (1 recipient among 27 Ph.D. graduates).
I will be at ACL 2026 in San Diego, presenting our long paper "Unlocking the Potential of Diffusion Language Models through Template Infilling" as an oral presentation!
I will be at CVPR 2026 in Denver, presenting our "CSF: Black-box Fingerprinting via Compositional Semantics for Text-to-Image Models" paper!
Selected Publications
SelectedAll
RobotUse: Allocating Computation, Context, and Decisions
arXiv preprint arXiv:2610.04929, 2026.
ACRO: Actor–Critic Rollout Orchestration
Preprint, 2026.
Discriminative Score Function: Turning Pretrained Models into Functional Generative Priors
The Conference on Neural Information Processing Systems (NeurIPS), 2026.
Clean Data Can Still Carry Backdoors: Support-Persistent Backdoors for Model Reuse
The Conference on Neural Information Processing Systems (NeurIPS), 2026.
Unlocking the Potential of Diffusion Language Models through Template Infilling
Annual Meeting of the Association for Computational Linguistics (ACL), 2026. Long Paper · Oral Presentation
CSF: Black-box Fingerprinting via Compositional Semantics for Text-to-Image Models
The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026.
Deep Edge Filter: Return of the Human-Crafted Layer in Deep Learning
The Conference on Neural Information Processing Systems (NeurIPS), 2025.
Deep Support Vectors
The Conference on Neural Information Processing Systems (NeurIPS), 2024.
Any-Way Meta Learning
The AAAI Conference on Artificial Intelligence (AAAI), 2024.
SHOT: Suppressing the Hessian along the Optimization Trajectory for Gradient-Based Meta-Learning
The Conference on Neural Information Processing Systems (NeurIPS), 2023.
Do Not Think About Pink Elephant!
The IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPR Workshops, Responsible Generative AI), 2024.
Awards & Honors
NeurIPS Top Reviewer
Best Ph.D. Dissertation Award
GSCT, Seoul National University · 1 recipient among 27 Ph.D. graduates
InnoCore Fellowship
ICML Gold Reviewer
BK21 Future Innovation Talent Bronze Prize
USD 1,000
BK21 Outstanding Research Talent Fellowship
USD 3,500
Yulchon AI Star Scholarship
USD 8,000
3rd Place, SNU FastMRI Challenge (out of 107 teams)
USD 3,000
Kwanak Scholarship
full funded
National Science and Engineering Scholarship
full funded