| Kang Minkyu I'm a 4th-year undergraduate at UNIST interested in Vision-Language-Action (VLA), Scalable Reinforcement Learning, and Diffusion Models. My main research question is how RL can be scalable? — what training methods, architectures, and supervision signals make reinforcement learning work at large, real-world scales. Recommended reads on this website are highlighted. |  |
Blog Posts |
 | Project Weasel A research initiative on making Vision-Language-Action (VLA) models genuinely useful in the real world. The main project, the W series, tackles the core problems VLAs face today and aims toward agents with real autonomy; sub-projects explore important VLA-adjacent topics. Sub-projects so far: W0 (an open VLA for long-horizon tasks), MineMate (multi-agent emergence with humans), and Mean Collapse in VLA (memory-based steering). |
 | Ouroboros: Reinforcement-Guided Flow Matching A reinforcement-guided flow matching framework that unifies critic optimization and generative diffusion in latent space for improved text-to-image alignment. |
Publications |