Minsu Kim

Minsu Kim is a CIFAR AI Safety Post-doc Fellow at KAIST and Mila, collaborating with Prof. Yoshua Bengio, Prof. Sungjin Ahn, and Prof. Sungsoo Ahn. His work bridges System 2 Deep Learning, Bayesian posterior inference, and combinatorial optimization.

• Ph.D., Industrial Engineering, KAIST (2025)
• M.S., Electrical Engineering, KAIST (2022)
• B.S., Mathematics and Computer Science (Dual Degree), KAIST (2020)

Kim's research focuses on enabling AI systems to measure uncertainty, represent causality, and perform sequential reasoning for safety-guaranteed planning. His methodology integrates GFlowNets and diffusion models with off-policy amortized inference, targeting applications in scientific discovery, hardware design optimization, and large language model alignment. Recent work explores Bayesian posterior inference through GFlowNets, aiming to unify deep learning with probabilistic reasoning.

His 15 most recent publications (2025–2024) reveal a trend of combining combinatorial optimization with generative models for tasks like molecular graph discovery, vehicle routing, and neural architecture search. He also investigates diffusion samplers for Bayesian inverse problems and symmetry-based neural methods (Sym-NCO) to enhance sample efficiency.

• Jang Yeong Sil Fellowship (2025)
• KAIST Presidential Best Ph.D. Thesis Award (2025)
• Qualcomm Innovation Fellowship (2023)
• DesignCon Best Paper Awards (2021–2022)

Kim's collaborations span KAIST's Industrial Engineering department and Mila's AI research groups. He actively contributes to academic peer review for top conferences (NeurIPS, ICML, ICLR) and journals (IEEE TNNLS, TPAMI), emphasizing the intersection of AI safety, uncertainty quantification, and systemic reasoning.