BIMSA >
BIMSA Computational Math Seminar
A Journey of Large Foundation Models in Science: Language, Multi-Modality and Agent
A Journey of Large Foundation Models in Science: Language, Multi-Modality and Agent
Organizers
Speaker
Renqian Luo
Time
Thursday, October 23, 2025 1:00 PM - 2:00 PM
Venue
Online
Online
Zoom 518 868 7656
(BIMSA)
Abstract
Large foundation models are transforming the landscape of scientific discovery by bridging natural language understanding, multimodal perception, and autonomous reasoning. This talk presents Dr. Luo's work of large pretrained language models in science, with a focus on their applications across biomedicine, materials science, and molecular design. Beginning with BioGPT, an early domain-specific generative transformer for biomedical text mining, we further introduce the developement of NatureLM, a generalist foundation model across diverrse domains in science, and analyze how scaling laws and prompt-based adaptation enable cross-domain generalization. We then move forward to multimodality in science and introduce UniGenX, a unified multimodal framework that represents both symbolic and numerical scientific data as sequences, supporting high-precision generation of 3D molecular and material structures. Finally, we discuss the Science Copilot Agent, a multi-agent system that integrates large models, scientific computation, and human feedback to assist end-to-end scientific workflows—from hypothesis generation to experimental design. Together, these advances illustrate the emerging paradigm of AI-accelerated science, where foundation models act not only as language models, but as universal scientific engines.
Speaker Intro
Dr. Renqian Luo was a Senior Researcher at Microsoft Research. He obtained his PhD from University of Science and Technology of China in 2021, in a joint program with Microsoft Research Asia. He has devoted himself into artificial intelligence research and development in machine learning, deep learning, natural language processing, large language model and foundation model. He has published many papers on top conferences or journals like NeurIPS, ICML, EMNLP, SIDKDD, ICASSP, Briefings in bioinformatics with 4000+ citations. His representative works include BioGPT, MedPrompt, NAO, NAS-BERT, LightSpeech, etc. His BioGPT and MedPrompt are widely reported and bring high impact to the community.