Shanghua Gao
I am a Research Fellow at Harvard University working with Prof. Marinka Zitnik, following my Ph.D. at Nankai University advised by Prof. Ming-Ming Cheng. My research spans foundational AI models — including representation learning, generative modeling, and efficient architectures — and AI for Science, where I develop autonomous agentic AI systems for biomedical reasoning and scientific discovery. I have published in top-tier venues in AI-for-science, including Cell (1), TPAMI (4), and major machine learning conferences such as CVPR and NeurIPS (6). My work has received over 7,000 citations and 8,000 GitHub stars and has successfully transitioned into real-world products.
On the foundational AI models side, I have introduced several influential models for large-scale representation and generative intelligence, including Res2Net, a widely adopted multi-scale backbone; LUSS, the first fully unsupervised large-scale semantic segmentation framework; MDT, the first mask diffusion transformer enabling state-of-the-art image synthesis and efficient training; and UniTS, the first unified multi-task time-series foundation model. These contributions establish general-purpose modeling capabilities across vision, multimodal learning, and time-series analysis.
On the AI for Science side, I develop autonomous, agentic AI systems that integrate scientific knowledge, tool use, and multi-step reasoning. My work published in Cell introduces the first AI agent for biomedical discovery. Building on this direction, ToolUniverse provides a unified ecosystem of AI scientists that enables large-scale, cross-domain scientific interactions, forming the infrastructure for general-purpose scientific agents. TxAgent is an agentic "AI scientist" for medicine that leverages multi-step reasoning and extensive interactions with ToolUniverse to perform therapeutic decision-making with high accuracy and interpretability. Together, these contributions outline a new paradigm for agentic AI in science — systems that autonomously reason across heterogeneous data, tools, and scientific domains.