Leibniz MMS Days 2026 - Keynote Lecture
Boleda, Gemma (Universitat Pompeu Fabra / ICREA, Barcelona, ES)
Why are Large Language Models so good at language?
The advent of deep learning has been a revolution in society, in general and also in particular regarding language, with millions
of people now using generative AI systems in their daily lives. It is also being a revolution in science, again in general and in
particular regarding language: For the first time in history, we have artificial models that reproduce linguistic behavior with a
high degree of accuracy. This raises the question, why are they so good? What do they have that was lacking before? In my talk, I
will provide theoretical arguments and empirical evidence for the view that deep learning models are successful because they
natively support two crucial properties of language: regularity and messiness. They are thus a synthesis between previous approaches
that focused on one at the expense of the other.