Jaemin Kim

About Me

I am a Ph.D. student at Graduate School of AI, KAIST, where I also completed my master’s degree under the supervision of Prof. Jong Chul Ye.

My research focuses on reward-maximizing generative AI, exploring ways to steer the outputs of advanced models such as diffusion models and large language models (LLMs) to optimize certain rewards. I work at the intersection of generative modeling, trajectory-level optimization, and black-box guidance, with the goal of making generative AI more controllable.

News

• [2026.06] Our paper (AgentPSO) is accecpted to 3rd ICML Workshop (AI4Math) 2026.
• [2026.05] I started internship at Sony!
• [2026.05] Our paper (Universal Reasoner) is accecpted to ICML 2026.
• [2026.01] Our paper (Training-free Reward Guiding) is accecpted to ICLR 2026.
• [2026.01] Our paper (Universal Reasoner) received a Silver prize at 32nd Samsung Humantech Paper Award.
• [2025.06] Our paper (Free²Guide) is accecpted to ICCV 2025.

My work trajectory

The central idea is to view generation and reasoning as trajectories, and to develop guidance signals that can steer these trajectories toward desired outcomes. This work has evolved along three connected directions:

• Trajectory Modeling. I study how generative trajectories can connect arbitrary distributions, providing a flexible foundation for controlled generation and manipulation ([C1]).
• Guidance from Generative Signals. I investigate how diffusion samples, motion cues, and vision-language feedback can reveal or construct useful guidance directions for generation, including gradient-free, plug-and-play, and training-free approaches ([C2–C4]).
• Reward-Directed Control and Reasoning. Building on these guidance mechanisms, I develop methods that optimize image generation and language-model reasoning toward explicit objectives through trajectory-level control and composable reasoning modules ([C5–C6]).

Publications

Services

Conference Reviewers

• ICLR 2026, ICML 2026

Powered by Jekyll and Minimal Light theme.