Abstract: Large Language Models (LLMs) depend on massive datasets whose quality and influence remain largely opaque. Data valuation offers principled methods to quantify how training data contributes to model performance and behavior. Yet, scaling classical approaches such as influence functions to trillion-token corpora continues to be a major challenge. This talk introduces recent advances that address this gap, including the linearized influence kernel, a new and efficient metric that extends to LLMs with billion-scale parameters. We will also highlight system-level frameworks such as RapidIn and present empirical findings of LLM training, including the slowly change phenomenon, which enables forward-looking valuation of future training data. By combining principled algorithms, system optimizations, and case studies, the talk aims to bridge the gap between theory and practice.

 

Short Bio: Zhaozhuo Xu is an Assistant Professor in the Department of Computer Science at Stevens Institute of Technology. He received his Ph.D. from Rice University and an M.S. from Stanford University. His research develops randomized algorithms to enhance the efficiency of AI systems on commodity hardware. Dr. Xu’s work has appeared in leading venues such as NeurIPS, ICML, ICLR, OSDI, and ACL, as well as in journals including Nature NPJ AI. His innovations in scalable AI have been integrated into widely used libraries like Hugging Face. He serves as an Associate Editor for Neurocomputing and as an Area Chair for major conferences, including NeurIPS, ICLR, ICML, ACL, EMNLP, NAACL, and COLING. He is a recipient of the AAAI New Faculty Highlights (2025), the NSF CRII Award (2025), and the Stevens Bridging Award.