Yeongbin Seo

Hi! I am a first-year PhD student in Computer Science at the University of Toronto and Vector Institute. I am supervised by Prof. Rahul G. Krishnan. Previously, I did my Master at Yonsei University in South Korea, supervised by Prof. Jaehyung Kim and Prof. Jinyoung Yeo. I also collaborated closely with Prof. Dongha Lee. I received B.S. in Educational Science from Yonsei University. I also served as a journalist specializing in data analysis in Donga Ilbo.

Research Interest (ML, NLP)

My research is focused on building human-like self-evolving AI, and I believe this requires improving the architecture of LMs. For example, to accelerate RL in LMs, the LMs need goal-orientation (bidirectional attention) and faster generation speed, which motivated my research on diffusion LMs. Moreover, for RL of LMs, the token-level granularity of LMs should be compressed into concept-level units in order to avoid the long-horizon (i.e., sparse reward) problem, so I’m currently working on hierarchical LMs (i.e., latent reasoning).

Also, I conducted research on continual knowledge learning, which is essential in self-evolving AI. I also explored efficient learning via data pruning, which will be required for continual learning of a vast stream of text data. I am potentially open to research on embodied agent, to build the emotion and personality of LM.