Gyorgy Turan : On the brittleness of large language models: A journey around set membership

Posted by Vishesh Jain , part of the Mathematical Computer Science Seminar.

At

Dec. 1, 2025, 3 p.m.

In

1227 SEO

Abstract

Large language models (LLM) have impressive performance on hard tasks, but also exhibit brittleness in simple tasks. We describe an experiment on a basic ``sub-reasoning'' task: deciding if an element belongs to a set. The results give a comprehensive picture of the various types of errors that can occur. In the second part of the talk we give a brief overview of the mathematical challenges posed by the goal of understanding how a neural network works, including understanding what an LLM ``knows''. Joint work with Gabor Berend, Lea Hergert, Mark Jelasity and Mario Szegedy.