Jeongjae Lee

I am a M.S. student at KAIST AI’s BISPL lab, advised by Dr. Jong Chul Ye. I received my B.S. in Electrical and Computer Engineering from Seoul National University (Summa Cum Laude).

Research Interests

My research interests lie in denoising generative models, reinforcement learning, and controllable generation. Recently, I have focused on RL-based post-training of diffusion and flow-based models, with an emphasis on improving reward alignment. I also maintain broader interests in applications to science and medicine.

Recent Works

• [ICML 2026 SPIGM Oral] RSM: Reward Score Matching, a unifying framework for reward-based fine-tuning of diffusion and flow-matching models. RSM shows that many existing methods can be understood through a common objective, leading to a smaller and more interpretable design space.

• [ICLR 2026] PCPO: Proportionate Credit Policy Optimization, a policy-gradient objective for preference alignment of diffusion and flow-matching models. PCPO improves training stability by replacing importance sampling weights with a log-ratio surrogate and correcting disproportionate credit assignment across timesteps.

Past Experience

• Postbacc Intern @ BISPL Worked on industrial co-op projects on medical AI.

• Undergrad Intern @ DeepMetrics Worked on automation of mechanical ventilator control in ICUs.

• Undergrad Intern @ Steinegger Lab Contributed to Foldseek, an open-source biotechnology software for protein structure comparison (published at Nature Biotechnology). Optimized LDDT alignment score computation, achieving a 13x speedup.