Jeonghun Park

I am an integrated M.S./Ph.D. student in Computer Science and Engineering at Seoul National University. My current research sits at the intersection of multimodal perception, commonsense reasoning, and agentic systems, with a longer-term goal of building reliable self-improving agents that can reason, act, and learn in physical environments.

Most recently, I worked as a research intern at the Human-Centered Computer System Lab at SNU, where I proposed a Code-Text Action style agent that combined textual reasoning with syntactic code execution and improved performance over a ReAct baseline. I also interned at FADU, building retrieval and tool-integration systems for production document understanding.

Before starting graduate study, I completed my B.S. at SNU with summa cum laude honors and served in the R.O.K. Air Force during my undergraduate years. I enjoy projects that force models to connect reasoning with execution, whether that means memory-efficient training systems, ARC-style problem solving, or embodied and game-like interactive environments.

Research Interests

• Multimodal perception
• Commonsense reasoning
• Computer-use agents
• Self-improving agents
• Physical AI

Recent Highlights

• Built an ARC-AGI solver with LoRA fine-tuning, augmentation, and test-time scaling that ranked 1st out of 35 teams on the hidden evaluation in a deep learning course project.
• Contributed memory-efficient training features to Samsung NNTrainer, including gradient checkpointing and lightweight optimizers.
• Built a minimal deep learning framework with define-by-run autograd and GPU support via CuPy.

A downloadable version of my latest CV is available here, and selected project write-ups are collected on the projects page.