Stat Cafe - Patricia Ning
Scalable Bayesian Inference for Large Language Model Analysis
- Time: Monday 3/24/2025 from 11:30 AM to 12:30 PM
- Location: BLOC 457
Description
Understanding and controlling large language models (LLMs) remains a fundamental challenge in AI research. Bayesian inference provides a principled approach to modeling uncertainty and structure in LLMs, enabling improved interpretability and adaptability. In this talk, we explore scalable Bayesian inference techniques, including sequential Monte Carlo (SMC) and probabilistic programming, to enhance LLM analysis. We discuss how these methods help uncover latent structures, improve in-context learning, and enforce constraints during text generation. By bridging theoretical insights with practical applications, we demonstrate how Bayesian methods can provide better control, robustness, and efficiency in LLMs. Finally, we highlight open challenges and future research directions in probabilistic modeling for large-scale deep learning systems.
Gallery (Photos by Samantha Williams)
