BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//132.216.98.100//NONSGML kigkonsult.se iCalcreator 2.20.4//
BEGIN:VEVENT
UID:20251120T205059EST-9347RcMAuE@132.216.98.100
DTSTAMP:20251121T015059Z
DESCRIPTION:Title: What Can Statistics Offer to Language Models: Watermarki
 ng and Evaluation\n\nAbstract: `Large language models (LLMs) have transfor
 med how we generate and process information\, yet two foundational challen
 ges remain: ensuring the authenticity of their outputs and accurately eval
 uating their true capabilities. In this talk\, I argue that both challenge
 s are\, at their core\, statistical problems\, and that statistical thinki
 ng can play an important role in advancing reliable and principled researc
 h on large language models. I will present two lines of work that approach
  these problems from a statistical perspective.\n\nThe first part introduc
 es a statistical framework for language watermarks\, which embed impercept
 ible signals into model-generated text for provenance verification. By for
 mulating watermark detection as a hypothesis testing problem\, this framew
 ork identifies pivotal statistics\, provides rigorous Type I error control
 \, and derives optimal detection rules that are both theoretically grounde
 d and computationally efficient. It clarifies the theoretical limits of ex
 isting methods\, such as the Gumbel-max and inverse-transform watermarks\,
  and guides the design of more robust and powerful detectors. The second p
 art focuses on language model evaluation\, where I study how to quantify t
 he 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 rev
 eal that even advanced LLMs often articulate only a fraction of their inte
 rnal knowledge\, suggesting a new perspective on evaluation and model comp
 etence. Together\, these projects represent an ongoing effort to develop s
 tatistical foundations for trustworthy and reliable language models\, with
  applications ranging from watermark detection to model evaluation.\n\n🔗 Z
 oom: https://mcgill.zoom.us/j/85469273736\n	Meeting ID: 854 6927 3736\n
DTSTART:20251124T163000Z
DTEND:20251124T173000Z
LOCATION:Room 1104\, Burnside Hall\, CA\, QC\, Montreal\, H3A 0B9\, 805 rue
  Sherbrooke Ouest
SUMMARY:Xiang Li (University of Pennsylvania)
URL:/mathstat/channels/event/xiang-li-university-penns
 ylvania-369123
END:VEVENT
END:VCALENDAR