Theme: Geometrization of AI (TDAI Speaker Series)
Speaker: Misha Belkin, University of California, San Diego
Date & Time: Wednesday, February 18, 2026 · 3:00–4:00 PM (ET)
Location: Pomerene Hall Room 350 (Project Zone)
Food: Light refreshments provided
Host: Dena Asta, Subhadeep Paul (TDAI Geometrization of AI Theme Leads)
A trained Large Language Model (LLM) contains vast amounts of human knowledge. Yet it is often difficult to determine what these models truly “know,” how their knowledge is represented internally, and how reliably their behavior can be interpreted or controlled. LLMs may appear confident while being incorrect, incomplete, or even unintentionally misleading.
In this seminar, Dr. Misha Belkin will explore feature learning and the “linear representation hypothesis” as a framework for understanding internal representations in modern neural networks. He will introduce Recursive Feature Machines, a method originally developed for extracting relevant features from tabular data, and show how this approach can be adapted to probe and interpret the structure of representations learned by large language models.
The talk will demonstrate how fixed directions in the activation space of an LLM can correspond to meaningful concepts, and how manipulating a single such vector can enable precise monitoring and steering of model behavior toward or away from specific semantic attributes.
Through this lens, the seminar will connect geometric perspectives on representation learning with practical questions of interpretability, controllability, and reliability in large-scale AI systems.
About the Speaker:
Dr. Misha Belkin is a leading researcher in machine learning theory and representation learning, with foundational contributions to deep learning, optimization, and the geometry of high-dimensional models. His work focuses on understanding how modern AI systems learn, generalize, and organize information internally, with recent emphasis on feature learning and interpretability of large language models.
Learn more: http://misha.belkin-wang.org/