Structured Representation Learning

Abstract. As LLMs move from research labs to real-world applications, understanding and controlling their behavior has become critical, given their rapid evolution and opaque internal mechanisms. This tutorial explores principled representation learning as a foundation for controllability, interpretability, and transferability. Participants will learn how to build interpretable, modular representations that guide model behavior, improve reasoning efficiency, and extend capabilities to new tasks through recombination. Grounded in human-centered measures and careful data design, this tutorial offers a roadmap for more robust, transparent, and trustworthy LLM-assisted systems.