Taylor Berg-Kirkpatrick

I’m an Associate Professor in the Department of Computer Science and Engineering at UC San Diego. My group studies generative models of structured data. We try to treat large generative models as objects of science rather than black boxes — asking how they learn, how to adapt and scale them, how to control what they produce, and how they fail — and we bring the structure and rigor of probabilistic modeling to each of those questions.

Recent work opens a new axis for scaling language models through recurrent depth (Parcae), gives controllable generation real guarantees in settings where the output has to be valid — runnable code, schema-conformant data, strings in a formal grammar — and makes pretrained generative models steerable and efficient at inference time. The same methods carry over to places standard tools can’t reach: historical print shops and undeciphered scripts, music and audio, and the security of models once they begin to act as agents.

Lately I’m most absorbed by language models that scale in new ways, generation that comes with guarantees, and turning generative models into instruments for reading archives the field can’t yet read.