From Tokens to Concepts: Designing LLMs to Think More Like Humans

Abstract:

Large language models (LLMs) learn to predict tokens, but are tokens the right units for understanding and designing their cognition? This talk explores three questions:

1. Do LLMs represent knowledge through human-interpretable concepts?
2. Can explicitly modeling concepts improve robustness, generalization, downstream performance, and alignment with human judgments?
3. How can we mathematically formulate concepts in humans and LLMs to compare them at a common level of analysis?

The talk positions concepts as a bridge between cognitive science and AI, offering new ways to understand how LLMs organize knowledge and to design models that think more like humans.

Bio:

Chen Shani is a Senior Lecturer at Tel Aviv University, where she leads the CoAI Lab, studying concepts, knowledge, and cognition in AI systems. Her research focuses on understanding what language models represent, how their knowledge is organized, and how these representations can be deliberately shaped. Before joining Tel Aviv University, she was a postdoctoral researcher at Stanford University and a researcher at Amazon. She holds a PhD from the Hebrew University of Jerusalem.