I am a Ph.D. student at the University of Toronto and the Vector Institute, advised by Professor Zhijing Jin and supported by the Connaught International Scholarship. Previously, I completed my M.S. at KAIST AI with Se-Young Yun and Kimin Lee, and my B.S. at Seoul National University. I work on post-training and self-improvement for language models, including exploration, adaptive curricula, knowledge distillation, reinforcement learning, memory-augmented agents, and continual learning.
My given name is 용진 (Yong-jin) in Korean.
My research asks how language models learn across the entire learning pipeline: what experiences they should learn from; where information should move among context, external memory, and model parameters; how they should explore; and how they can seek and use feedback. I am also interested in proactiveness and curiosity—how a system can recognize what it does not know and decide what to learn next. Rather than treating these as separate stages, I study how their interactions can help models continue improving after initial training.
Transferability for General Reasoning: An Automated Curriculum for Multi-Domain RLVR ⚡
Preprint · ICML 2026 RLxF Workshop (Oral, Top 4 Best Papers out of 222 submissions)
Making Complex Reasoning Student-Friendly: A Hybrid Distillation Framework for Small Language Models
Scaling Post-training for LLMs Workshop @ ICLR 2026
Self-Distillation Zero: Self-Revision Turns Binary Rewards into Dense Supervision ⚡
Preprint · ICML 2026 RLxF Workshop (Oral, Top 4 Best Papers)
Learning is, at its core, fitting data, so the data often matters most. I am interested in which data should be used for which tasks, how to organize it into curricula, and how to construct effective data for knowledge distillation.
What should a model absorb into its parameters, retain in external memory, or receive through context? I am interested in how stability, reuse, and forgetting should affect that choice.
A system can only learn from the experiences it produces. I am curious about agents that guide their own exploration and about whether useful exploration should happen in text, latent, or task-specific spaces.
Notes in preparation:
A look at exploration in classical reinforcement learning, why it becomes a bottleneck in reinforcement learning for language models, and which ideas may be worth trying next.
An analogy to how humans learn, followed by a recap of research on updating a model through its parameters or through external memory.
A researcher's tour of the Claude ecosystem: models, Claude Code, the Agent SDK, MCP, and skills, and what it signals about where AI systems are heading.
I am always happy to connect and very open to collaborations on post-training, exploration, memory, agents, or anything adjacent. Reach out at dyyjkd@gmail.com, or find me on LinkedIn and Google Scholar.