As AI models become increasingly powerful, it is an attractive proposition to use them in important decision making pipelines, in collaboration with human decision makers. But how should a human being and a machine learning model collaborate to reach decisions that are better than either of them could achieve on their own? If the human and the AI model were perfect Bayesians, operating in a setting with a commonly known and correctly specified prior, Aumann's classical agreement theorem would give us one answer: they could engage in conversation about the task at hand, and their conversation would be guaranteed to converge to (accuracy-improving) agreement. This classical result however would require making many implausible assumptions, both about the knowledge and computational power of both parties. We show how to recover similar (and more general) results using only computationally and statistically tractable assumptions, which substantially relax full Bayesian rationality. Joint work with Natalie Collina, Varun Gupta, and Surbhi Goel, based on a paper that will appear in STOC 2025.

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 method originally 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 single fixed vector in the LLM activation space.