Date: February 14, 2025
Time: 12:00 PM - 1:00 PM
Location: 1889 Museum Road, Gainesville, FL, 32611
Host: Department of CISE; Faculty Host: Dr. Zhe Jiang
Admission: Free

Zoom Talk: https://ufl.zoom.us/j/97967468472

Biography: Mingyu Derek Ma is a PhD candidate in Computer Science at UCLA working with Prof. Wei Wang. He is also a Machine Learning Scientist at Genentech Prescient Design, working on drug discovery agents with Dr. Stephen Ra and Prof. Kyunghyun Cho. His work focuses on the architecture, training, and agentic use of generative language models inspired by and applied to clinical, medical, and scientific scenarios. He is currently working on equipping the language models with the intuition and knowledge of domain experts, such as clinicians or scientists, and utilizing them as assistants for scientific discovery. His work has been recognized as one of the top 15 most-cited papers at NAACL 2024 and has been published at leading AI conferences like ACL, AAAI, and NeurIPS. He received the J.P. Morgan Chase AI PhD Fellowship and Amazon Fellowship. He has research experience at Amazon AGI, UC Santa Cruz, CUHK and MIT. He earned his BS from The Hong Kong Polytechnic University, where he was the commencement speaker.

Title of the Talk: Elevating Large Language Models
to Expert Intelligence

Abstract: Large Language Models (LLMs) have been applied to expert domains and scientific contexts, such as clinical diagnosis and drug discovery. However, the generalizability that characterizes LLMs in the general domain does not readily translate to scientific and expert tasks. Unlike general natural language tasks, scientific data is densely packed, homogeneous, and less self-explanatory. Moreover, expert-level tasks, such as those performed by physicians, engineers, or scientists, require deep domain knowledge, intuitive reasoning, and multi-step planning, refined through years of specialized training. In this talk, I will first discuss capturing implicit expert intuition for individual decision-making, using clinical diagnosis prediction as a case study. Next, I will extend the focus to compositional, project-level reasoning and automation, highlighting the development of LLM agents for scientific discovery. Finally, I will address critical concerns around fairness and safety in generative LLMs, proposing novel methods for unsupervised bias mitigation.