Juho Jung

I am a Research Scientist (Alternative Military Service) at VUNO Inc. My research focuses on adapting multimodal LLMs to medical images with uncertainty awareness, and on improving uncertainty quantification and reducing hallucinations of LLMs in real-world settings.

Before joining VUNO, I received my Master's degree from the Graduate School of Applied Artificial Intelligence at SKKU (Full Academic Excellence Scholarship), advised by Prof. Jinyoung Han. There I worked on uncertainty-aware multi-task learning for medical images and multimodal feature alignment under complex and unknown modalities. I received my bachelor's degree (Summa Cum Laude) in Applied Artificial Intelligence from SKKU.

I was also a Visiting Scholar in the School of Computer Science Intensive Program in AI at Carnegie Mellon University, where I had the privilege of working with Zhi-Qi Cheng, Alexander Hauptmann, and David Mortensen.

Email  |  CV  |  Google Scholar  |  GitHub  |  LinkedIn

Research Interests

I am interested in advancing uncertainty-aware multimodal learning for medical imaging. My work explores how to adapt large vision-language models to clinical data, quantify and calibrate their uncertainty, and mitigate hallucinations in real-world medical settings. My current research emphasizes:

• Vision-Language Models for medical imaging — adapting large vision-language models to clinical data
• Uncertainty Quantification in multimodal learning — quantifying and calibrating what models do not know
• Hallucination detection and mitigation — making model outputs reliable in real-world clinical settings
• Representation Learning for medical images — learning robust features from limited and heterogeneous medical data

News

• [Oct 2026] New preprint on arXiv! SPARCL (Spatially Adaptive Regional Coverage for Active Learning) is an entropy-gated acquisition strategy that balances predictive uncertainty with spatial coverage, adapting to both focal and diffuse lesion patterns in chest X-rays. See the code.
• [Sep 2026] One paper accepted to NeurIPS 2026 as a poster! ODS deconstructs multi-task active learning: gradient conflict is moved out of the acquisition score and resolved by the optimizer, and selection becomes two ordered stages of informativeness filtering and coverage maximization. See the project page.
• [Sep 2026] I was recognized as a Top Reviewer at NeurIPS 2026.
• [Jul 2026] I served as a Program Committee member for AAAI 2027.
• [May 2026] I was recognized as a Top Reviewer at ICML 2026.
• [Mar 2026] I served as a reviewer for MICCAI 2026.

Selected Publications

(Equal contributions are denoted by *)

Selected Projects

Academic Services

Template by Jon Barron.