Tianyi Xu

I’m Tianyi Xu, graduated from the University of Wisconsin–Madison (B.S. in Computer Science, Data Science, and Mathematics). I am advised by Prof. Junjie Hu at UW–Madison. I work closely with Shaobo Wang and Prof. Linfeng Zhang at Shanghai Jiao Tong University, and Prof. Claudia Solís-Lemus at UW–Madison.

I’m interested in building modern AI systems that are efficient, generalizable, and capable of understanding and acting across multiple modalities, especially under limited supervision and real-world constraints. Concretely, my work focuses on:

1. Data-centric & label-efficient learning — Designing methods for data selection, mixing, and self/weak supervision so that large-scale models can learn from noisy, heterogeneous data instead of only clean benchmarks.

2. Foundation models & agents— Building and steering pretrained models (LLMs, VLMs, etc.) for specific tasks, with an emphasis on efficiency, reliability, and agentic behaviors such as planning, reasoning, and safe decision-making.

3. Multimodal learning — Representation learning and building systems that perceive, act, and reason across different modalities.

4. AI for science and society — Applying these ideas to real-world problems where data is scarce or noisy: biodiversity monitoring from soundscapes, tone-aware speech modeling for accessibility, and multimodal toolkits/benchmarks for biomedical and clinical AI.

I’m seeking Ph.D. opportunities for Fall 2026.