Minsu Kim

I’m a Senior Research Scientist at Microsoft Frontier Tuning in Washington. I build an RL framework that automatically turns general-purpose frontier LLMs into enterprise-specific agents (e.g., legal agents). I train these agents to search unfamiliar corpora through long sequences of tool calls. The main technical challenges are expanding knowledge coverage for exploration, credit assignment, and policy staleness as the learner updates asynchronously during these long rollouts.

Outside my work, I’ve become interested in scalable oversight. I see poor judge calibration and reward hacking as major obstacles to progress in both RL and long-term AI safety.

Before joining Microsoft, I worked on off-policy RL and structured exploration as a postdoctoral fellow with Yoshua Bengio at Mila and Sungjin Ahn and Sungsoo Ahn at KAIST.

I received my Ph.D. from KAIST under Jinkyoo Park, studying RL for combinatorial optimization. I completed my M.S. at KAIST under Joungho Kim’s supervision, studying RL for electrical design automation with signal/power integrity considerations. I received my B.S. in Mathematics and Computer Science from KAIST.

For professional inquiries and speaking invitations, email minsukim at microsoft dot com. For independent research or non-commercial academic collaboration in a personal capacity, email minsukim dot ai at gmail dot com.