<2026 Spring Colloquium #3: How Do Knowledge and Belief Emerge in Language Models: Anatomy and Surgical Techniques (Prof. Yohan Jo)>

ㅇ Speaker: Prof. Yohan Jo (Department of Data Science)

ㅇ Date & Time: Thursday, July 2, 2026, 5:00 PM

ㅇ Location: Room 105, Building 302

ㅇ Lecture Overview
    Today, AI systems that have become part of our everyday lives — such as ChatGPT and Gemini — are built on language models.

While language models appear on the surface to possess extensive knowledge and particular beliefs, the internal mechanisms by which such knowledge and beliefs are expressed remain extremely difficult to interpret.

As a result, problems such as a model failing to answer using knowledge it already possesses, or expressing a user-specified persona in an unnatural way, often occur even though it remains difficult to diagnose and correct their underlying causes.

This lecture focuses on geometric techniques for interpreting these internal mechanisms, covering how to correct inconsistencies between a model's knowledge and its answers, how to fix incorrect knowledge,

psychometric methods for accurately analyzing a model's beliefs, and two distinct mechanisms by which a model's beliefs are expressed.

Building on these approaches, the lecture discusses how to make language models more interpretable, trustworthy, and controllable tools.

ㅇ Speaker Bio

    Professor Yohan Jo conducts research at the SNU Graduate School of Data Science on AI agents, reasoning techniques, value alignment, and interpretability methods.

ㅇ Watch the lecture: https://www.youtube.com/watch?v=dwRcciKYavs