[Abstract]

Recent vision-language models (VLMs) have demonstrated remarkable capabilities across a wide range of multimodal tasks, but adapting them to new tasks or improving their reliability often requires additional training and computational resources. In this talk, I will introduce recent training-free approaches that enhance the capabilities of pretrained VLMs without updating model parameters. The talk will cover methods for improving visual grounding, reasoning, and interpretability by leveraging internal representations, attention patterns, and inference-time interventions. I will also discuss our recent work and broader opportunities for developing more efficient, adaptable, and trustworthy multimodal AI systems.

[Biography]

Seong Jae Hwang is an Associate Professor in the Department of Artificial Intelligence at Yonsei University, where he leads the Medical Imaging and Computer Vision Laboratory. He received his Ph.D. in Computer Sciences from the University of Wisconsin–Madison and previously served as an Assistant Professor at the University of Pittsburgh. His research focuses on computer vision, medical imaging, multimodal learning, and trustworthy AI, with recent interests in vision-language models, model interpretability, and training-free adaptation. His work has been published in major AI and medical imaging venues including CVPR, NeurIPS, ICML, MICCAI, and Medical Image Analysis.