Xinyan Velocity Yu 余忻妍
I am a Research Engineer at Microsoft Office of Applied Research , where I work on collaborative agents and privacy-preserving synthetic data generation.
I am also a Computer Science Ph.D. student on leave at the University of Southern California, advised by Jesse Thomason and Dani Yogatama. At USC, my research has explored how language models retrieve, organize, and transfer knowledge across tasks, languages, and modalities through retrieval augmentation, representation analysis, and rigorous evaluation. I completed my B.S. and M.S. at the University of Washington, where I worked with Hannaneh Hajishirzi, Luke Zettlemoyer, Akari Asai, Sewon Min, and Jungo Kasai.
I was born and raised in Chongqing, China.
News
- Released Dr. DocBench, a benchmark for expert-level and difficult document parsing.
- Joined Microsoft Office of Applied Research as a Research Engineer.
- Cancer-Myth was accepted to ICLR 2026. I won’t make it to Brazil this time—please send photos!
- I’m going to NeurIPS! See everyone in San Diego!
- CodeRAG-Bench appeared in Findings of NAACL 2025. I’ll miss Albuquerque this time.
- The Semantic Hub Hypothesis appeared at ICLR 2025. Singapore—and its food—will have to wait for another trip!
Research
My current work focuses on training and evaluating collaborative agents that interact with people, other agents, and their environments. I also build agentic pipelines for synthetic data generation, using reinforcement learning to optimize their behavior while incorporating differential privacy.
More broadly, I study how language models retrieve, organize, and transfer knowledge across tasks, languages, and modalities, with an emphasis on retrieval augmentation, representation analysis, rigorous evaluation, false-premise detection, and document intelligence.
Publications
* denotes equal contribution.
Peer-reviewed
Transactions on Machine Learning Research (TMLR), 2026
Findings of NAACL 2025
NAACL 2024 Oral Presentation
NeurIPS Datasets and Benchmarks Track, 2023
ACL 2023 Oral Presentation
Findings of ACL 2023
Preprints
Preprint, 2026