【系综合学术报告】2026年第25期
标题：What Can Statistics Offer to Language Models: Watermarking and Evaluation
报告人：李翔（University of Pennsylvania）
时间：2026年6月3日（周三）上午9:00-10:00
地点：文北102 腾讯会议：954-265-005
摘要：Large language models (LLMs) have transformed how we generate and process information, yet two foundational challenges remain: ensuring the authenticity of their outputs and accurately evaluating their true capabilities. In this talk, I argue that both challenges are, at their core, statistical problems, and that statistical thinking can play an important role in advancing reliable and principled research on large language models. I will present two lines of work that approach these problems from a statistical perspective.
The first part introduces a statistical framework for language watermarks, which embed imperceptible signals into model-generated text for provenance verification. By formulating watermark detection as a hypothesis testing problem, this framework identifies pivotal statistics, provides rigorous Type I error control, and derives optimal detection rules that are both theoretically grounded and computationally efficient. It clarifies the theoretical limits of existing methods, such as the Gumbel-max watermark, and guides the design of more robust and powerful detectors. The second part focuses on language model evaluation, where I study how to quantify the unseen knowledge that models possess but may not reveal through limited queries. To that end, I introduce a statistical pipeline, based on the smoothed Good–Turing estimator, to estimate the total amount of a model’s knowledge beyond what is observed in benchmark datasets. The findings reveal that even advanced LLMs often articulate only a fraction of their internal knowledge, suggesting a new perspective on evaluation and model competence. Together, these projects represent an ongoing effort to develop statistical foundations for trustworthy and reliable language models, with applications ranging from watermark detection to model evaluation.
This talk is based on the following works:
https://arxiv.org/abs/2404.01245
https://arxiv.org/abs/2506.02058
and will briefly mention follow-up studies:
https://arxiv.org/abs/2411.13868
https://arxiv.org/abs/2510.22007
报告人简介：Xiang Li is a postdoctoral researcher at the University of Pennsylvania and an incoming Assistant Professor in the Department of Statistics at Rutgers University. His research focuses on statistical machine learning and trustworthy artificial intelligence, with a particular focus on the statistical foundations of large language models, including watermarking, uncertainty quantification, and model evaluation. His work has appeared in leading journals and conferences in statistics and machine learning. More broadly, his research aims to develop principled statistical methods for understanding, evaluating, and improving modern AI systems.
【系综合学术报告】2026年第26期
标题:Model-Based Derivative-Free Optimization: From Robust and Scalable Models to AI and Quantum Applications
报告人：谢鹏程（University of California）
时间：2026年6月3日（周三）上午10:00-11:00
地点：文北102 腾讯会议：954-265-005
摘要:
This lecture will present recent advances in model-based derivative-free optimization for black-box scientific and engineering problems. In many modern applications, such as materials discovery, climate and fluid simulation, machine learning, and quantum computing, accurate gradient information is unavailable because objective evaluations arise from complex simulators, experiments, or noisy black boxes with unknown analytic form. Consequently, progress often relies on repeated trial evaluations, which can be extremely expensive and time-consuming. Model-based derivative-free optimization methods, pioneered by the late Professor Michael J. D. Powell at Cambridge, were developed for such settings. However, existing methods can become unreliable for high-dimensional or large-scale problems, under noise, or when computational budgets are limited, as surrogate models may lose accuracy and algorithms may require many unnecessary trial steps before making progress. In this talk, I will discuss new model-based derivative-free optimization methods with improved robustness and scalability, including more reliable local approximation models, subspace-based algorithms for very high-dimensional problems, and mechanisms that enhance reliability under noise, uncertainty, and transformed objectives. I will also highlight connections with AI-enhanced optimization, applications in AI and quantum computing, and scientific computing problems involving numerical differential equations.
报告人简介:
Dr. Pengcheng Xie is a postdoctoral scholar at Lawrence Berkeley National Laboratory and the University of California in the United States, working with Division Director Stefan M. Wild since July 2024. His research focuses on computational mathematics, mathematical optimization, machine learning, and numerical analysis. He earned his PhD from the Chinese Academy of Sciences, where he was advised by Prof. Ya-xiang Yuan. Pengcheng has published in leading journals including SIAM Journal on Optimization, IMA Journal of Numerical Analysis, Optimization Methods and Software, Journal of Computational and Applied Mathematics, Journal of Computational Mathematics, Journal of the Operations Research Society of China, Chinese Annals of Mathematics, and American Control Conference. He has also organized or presented at major international conferences such as SIAM MDS 2026, IOS 2026, SIAM CSE 2025, INFORMS 2025, SIAM NCC 2024, ISMP 2024, ICIAM 2023, and CSIAM 2022. His academic honors and recognitions include the SIAM-NCC Early Career Travel Award (2024), Oxford Hooke Fellowship Finalist (2024), the Hua Loo-keng Prize/Scholarship (2019–2024), the CAS AMSS President Prize/Scholarship (2023), financial support from ICIAM (2023), National Outstanding Student Award of the National Program for Cultivating Top Students in Fundamental Sciences (2021), and the Gold Award at the China Southeast Mathematical Olympiad. Beyond his research, Pengcheng has been active in professional service. He is a member of SIAM and MOS (since 2019), organized the minisymposiums at SIAM conferences, and serves as editor, reviewer, and referee for multiple journals, including Mathematical Reviews/MathSciNet of the American Mathematical Society, SIAM Journal on Optimization, Mathematical Programming Computation, IMA Journal of Numerical Analysis, Optimization, Journal of the Operations Research Society of China, etc. He served as President of the CAS SIAM Student Chapter (2021–2022), and is the founding Editor-in-Chief of the Journal of Zhufeng (Fundamental Science), supported by the Ministry of Education of China since 2018. In addition, he has led multiple mathematical modeling and machine learning teams since 2016. He has been invited to deliver 1-hour lectures in Oxford, Argonne National Lab, Berkeley National Lab, and Harvard.
