ASSET Seminar featuring Misha Belkin (University of California, San Diego)

Talk Title: From kernel machines to  the linear representation hypothesis for monitoring and steering LLMs

Abstract: A trained Large Language Model (LLM) contains much of human knowledge. Yet, it is difficult to gauge the extent or accuracy of that knowledge, as LLMs do not always "know what they know'' and may even be unintentionally or actively misleading. In this talk I will discuss feature learning  introducing Recursive Feature Machines — a powerful generalization of the classical kernel methods designed for extracting relevant features from tabular data. I will demonstrate how this technique enables us to detect and precisely guide LLM behaviors toward almost any desired concept by manipulating a fixed vector in the LLM activation space. I will also discuss how the same method allows for probing for whether LLM  exhibits motivated reasoning.