APS
Speaker: Sayantan Sinha, IIT Bombay
Host: Debraj Das
Title: Large Language Models, Statistics, and WatermarkingÂ
Time, day and date: 4:00:00 PM – 5:00:00 PM, Wednesday, September 30
Venue: Conference Room
Abstract: In this seminar presentation we examine the mathematical and statistical principles underlying modern Large Language Models (LLMs), tracing their evolution from foundational probabilistic language modeling to transformer architectures. We will also see details how classical statistical concepts including maximum likelihood estimation, scaling laws, calibration, and preference optimization empower modern training pipelines. As a primary case study, we go through rigorous mathematical analysis of statistical text watermarking, focusing on green/red list generation, logit biasing, z-statistic hypothesis testing, detection power, and robustness against editing.