Research Scientist, NAVER Cloud (Image Vision)
I'm a Research Scientist at NAVER Cloud and a Ph.D. student at KAIST. My current research focuses on self-improving agents — agents for agents: meta-agents that diagnose a base agent's failures from its execution traces and optimize its prompts, harness, and tools from only a few labeled examples, with model weights frozen. I currently focus on visual reasoning agents, with the goal of developing methods that generalize across agent systems.
My background spans visual recognition — segmentation, detection, and matting — and grounded VLMs and visual reasoning agents, with 14+ papers at CVPR, ICCV, ECCV, NeurIPS, and AAAI (621 citations, h-index 10). I bring that experience in perception and diagnostic reasoning, along with a track record of taking research into products used by millions, to my current work on agents that improve through experience.
Agents for Agents
Building a meta-agent that improves a frozen base agent from only a handful of labeled examples. It audits the base agent's full execution traces against ground truth, localizes where they go wrong, then proposes and empirically verifies fix hypotheses to optimize the agent's prompts, harness, and tools — all while keeping model weights frozen.
Currently focused on few-shot self-improving visual agents, characterizing the challenges unique to visual reasoning to inform improvement methods that generalize across agent systems.
Studied grounded VLMs that connect language to pixel- and region-level evidence, and visual reasoning agents that compose specialist vision models into executable workflows.
Failure diagnosis & targeted program repair
Promptable, zero-shot segmentation & matting for anything; label-efficient human matting.
ZIM (ICCV'25 Highlight) · WSSHM
Weakly- and semi-supervised semantic & instance segmentation from cheap supervision.
PointWSSIS · BESTIE · DRS · WSSS-BED
Class-incremental and panoptic continual segmentation; lightweight, efficient models.
ECLIPSE · SSUL · EResFD · PfLayer
Visual reasoning agents. A diagnostic framework for 3D spatial reasoning that turns silent perception failures in visual program synthesis into typed diagnoses, driving targeted program repair to rival frontier VLMs without task-specific training.
"Beyond AI, Into Agents." A researcher's story from the AI agent era.
event →
CLOVA-X Image Editing: the world of pixel magic delivered by AI.
session →
How to land an AI research engineer role — career mentoring (NAVER CLOVA).
SSUL: Semantic Segmentation with Unknown Label for Class-Incremental Learning.
social →Weakly-Supervised Instance Segmentation (CVPR 2022), hosted by Prof. Chaeeun Rhee.