Imaging Science Seminar: Shaping Visual Reasoning in LVLMs through Diverse Alignment

Imaging Science Seminar
Shaping Visual Reasoning in LVLMs through Diverse Alignment

Zoom Link here

Dr. Zhiqiang Tao
Assistant Professor, School of Information
Rochester Institute of Technology

Abstract:
Visual reasoning plays a vital role in applying Large Vision-Language Models (LVLMs) to broad applications, including embodied AI, healthcare, imaging, and search agents. It has evolved from perception and grounding to visual chain-of-thought (CoT), and more recently to latent reasoning paradigms. In this talk, we will investigate visual reasoning across diverse alignment strategies, spanning instruction tuning, preference optimization, and generative policy learning. First, we enable LVLMs to automatically discover visually relevant questions by designing a new question-driven SFT. Our key hypothesis is that active inquiring can better probe visual clues associated with the intended answers, reducing the need for manually prompt designs, and forming the foundation of visual reasoning. Second, we study a new structured preference optimization method to enhance visual preference with tree hierarchies that encompass instructions, responses, and dialogue interactions. We aim to explicitly prioritize conversations that exhibit strong visual dependency between questions and answers. Third, we propose a latent CoT approach to reformulate reasoning as a posterior inference problem using GFlowNets. By learning a generative policy that matches the reward distribution at the token level, we can sample diverse reasoning trajectories in the latent space, effectively mitigating reward hacking and mode collapse. Our proposed research has been implemented on various LVLMs (e.g., LLaVA, Qwen-VL, and InternVL) and across multiple sizes ranging from 0.5B to 13B, and could be applied to downstream tasks, such as autonomous driving, medical RAG, sign language translation, etc.

Speaker Bio:
Dr. Zhiqiang Tao is currently a tenure-track Assistant Professor in the School of Information at the Rochester Institute of Technology. He was an Assistant Professor in the Department of Computer Science at Santa Clara University from 2020 to 2022. He obtained his Ph.D. degree at Northeastern University in 2020, under the supervision of Prof. Yun Fu. He was a research intern at Adobe, CA, in 2018, and Alibaba’s DAMO Academy, WA, in 2019. He has published over 70 papers on peer-reviewed journals and conferences, including IEEE TPAMI, TNNLS, TIP, TKDE, NeurIPS, ICML, ICLR, CVPR, ICCV, ECCV, KDD, SIGIR, WWW, EMNLP, NAACL, IROS, etc. He serves as Associate Editor of IEEE TCSVT and Neurocomputing and has served as Area Chair, (senior) PC member, and reviewer for prestigious journals and international conferences. He won the 3rd place award in KDD Cup AutoML track in 2019. His research has been supported by the National Science Foundation (NSF), the National Institutes of Health (NIH), and the DEVCOM Army Research Laboratory (ARL). His recent research focuses on developing reliable and efficient AI systems using LVLMs, agentic AI, and policy optimization.

Intended Audience:
Beginners, undergraduates, graduates. Those with interest in the topic.

To request an interpreter, please visit myaccess.rit.edu