Abstract

Recent progress in language models and artificial agents has renewed foundational questions about how intelligence is learned, structured and represented. In this omnibus talk, I present three complementary perspectives on these questions. First, I discuss ontology creation using large language models, grounded in the notion of competence: what it means for a model to meaningfully possess and organise knowledge beyond surface-level performance. Second, I explore how inspiration from human pain and aversive signals can inform more efficient reinforcement learning, viewing pain as a mechanism for shaping learning dynamics rather than merely optimising reward. Third, I examine the geometry and dimensionality of representation spaces in large language models and what these structures reveal about overfitting, convergence, and ultimately grokking. Though very different in focus, these perspectives jointly contribute to a deeper understanding of the nature of unsupervised learning.

Prof. Dr. Milica Gašić is a professor in Dialog Systems and Machine Learning at Heinrich Heine University, Düsseldorf, Germany. Prior to her current appointment she was a group leader in Saarland University and, before that, a lecturer at the University of Cambridge. Her research focuses on fundamental questions of human-computer dialogue modelling and lie in the intersection of Natural Language Processing and Machine Learning. She is a recipient of the European Research Council Starting Grant, the Alexander von Humboldt Sofja Kovalevskaja Award and the Lamarr Fellowship. Prof. Gašić served as the vice-president of SIGDIAL 2021 -2024. She is a member of the International Scientific Advisory Board of DFKI , a member of ACL , a member of ELLIS and a senior member of IEEE .