Minbeom Kim

Hi! I am a Ph.D student at Seoul National University. I am fortunate to be advised by Prof. Kyomin Jung. Previously, I received B.E. degrees from KAIST. I have also interned at NAVER AI LAB for controlled text generation, advised by Hwaran Lee.

In my Ph.D. course, my research question is as follows: “How can we guarantee generations that are aligned with personalized preferences and controlled under hard constraints while still preserving pre-aligned general preferences from generative foundation models?” While novel training methods might offer a potential solution, I believe that controlling the generation at the decoding-time is a more suitable approach for practical, real-world applications. This path presents significant challenges. We need to understand both the theoretical issues and the empirical effects of adjusting pre-aligned models to accommodate new personal values or hard constraints. Based on this fundamental understanding, I must investigate how to manipulate the distribution at the decoding-time to obtain the desired generations. Although it looks deeply challenging, I am enjoying this journey and am looking forward to the new insights it will bring.😎