Jin Kim

I'm Jin, a Ph.D. candidate in Physics and Biology in Medicine at UCLA. I research multimodal foundation models: vision-language alignment, visual grounding, supervised and RL-based post-training (SFT, GRPO), longitudinal image analysis, and explainable evaluation, with medical imaging as my main application area.

At UCLA's Computer Vision and Imaging Biomarkers Lab, my recent work covers reliable structured outputs from medical VLMs (Med-R), phrase grounding versus detection for chest X-ray localization, longitudinal change extraction (Q-LACE), and auditable report evaluation. In 2026 I was an Applied Scientist Intern at Amazon (Computer Vision - Devices), working on multimodal post-training. Before graduate school I led robotics automation projects at ABB Korea. A short research portfolio summarizes selected projects.

Recent News

• [Oct 2026] Paper submitted to Journal of Imaging Informatics in Medicine - "Q-LACE: Separating Benchmark Performance from Conditional Reliability in Longitudinal Chest Radiograph Assessment"
• [Oct 2026] Paper accepted in Journal of Imaging Informatics in Medicine - "Comparing Phrase Grounding and Detection for Conditional Chest-Radiograph Abnormality Localization"
• [Oct 2026] Built DanceCraft with M. Wasil Wahi-Anwar at the Claude Creator Summit Hackathon, LA Tech Week - a webcam dance challenge where Claude turns pose-tracking scores into a focused practice plan.
• [Oct 2026] Completed my Applied Scientist internship at Amazon (Computer Vision - Devices).
• [Sep 2026] MIRASOL 2026 Workshop held at MICCAI 2026 in Strasbourg - as Publication Committee Co-Chair, coordinated camera-ready proceedings for 70 accepted papers.
• [Sep 2026] Talk at the Changbal AI/DS Group Seminar, Bellevue, WA - "Can We Trust Research Conducted by AI Agents?" [Video]
• [Jul 2026] Built "CD4+ T Cell Regulator Atlas" for Built with Claude: Life Sciences, a hackathon by Anthropic with Gladstone Institutes (Research Track) - master-regulator prioritization from genome-scale Perturb-seq data.
• [Jun 2026] Paper published in Medical Image Analysis - "Beyond the LUMIR challenge: The pathway to foundational registration models"

Research Interests

Vision-Language Models & Post-Training

Post-training vision-language models so their outputs are reliable and checkable. Med-R uses hard-gated SFT and GRPO so that only schema-valid outputs earn task rewards, and Q-LACE studies query-conditioned longitudinal reasoning on chest X-rays. At Amazon I worked on vision-language model post-training and image generation. Earlier work includes dual-path radiology report generation (MICCAI 2025 Workshop). Research direction outlined in PhD dissertation proposal.

Visual Grounding & Spatial Reasoning

Connecting language to locations in images. A first-author study accepted in JIIM shows that direct phrase grounding localizes a specified chest X-ray finding far more often than detection (40.5% vs. 3.3% success at IoU ≥ 0.2 on 489 external cases). TRACE and SPAR (ICCV 2025 workshops) use textual reasoning to predict where objects should be placed, connecting VLMs to robotic manipulation.

Medical Image Analysis

Self-supervised learning and foundation models addressing data scarcity challenges in computer vision. Developing methods to handle limited annotated data through innovative pretraining approaches and domain adaptation techniques. Core expertise demonstrated in PhD qualifying examination on foundation models and knowledge distillation. Recent presentations include foundation models with deep layer adapters at RSNA 2024 and self-supervised learning for chest X-ray segmentation at SPIE 2024.

Agentic AI

AI agents that turn natural-language tasks into executable computer vision training and inference workflows, with external verification and reproducible experiments. Recent work includes autonomous computer vision development with SimpleMind, published on arXiv with open-source code.

Earlier Work: AR Surgical Navigation

Previous research in AR-enhanced navigation systems for precise surgical procedures, particularly in ENT applications. This work includes AI-Enhanced AR Navigation System for Precise Mastoidectomy from my Master's thesis, which established core competencies in medical imaging that inform current Vision-Language model research.

Education

Experience