DIKU Bits - Language Models, by Maria Vlachou and Mike Zhang

This DIKU Bits will be a little different than usual, as there will be two speakers, each presenting a project.

Maria will talk about how credibility assessment shapes asylum decisions and how large language models can help evaluate these judgments. She will outline key indicators that influence credibility and show how LLMs and semantic search can extract and synthesize signals relevant to credibility and risk. Furthermore, she will focus on practical methods for turning complex narratives into structured insights.

Mike will be talking about how prompting a language model with phrases like “a pineapple in its own juice” reveals whether it truly grasps the cultural quirks that native speakers take for granted. He will argue that standardized tests can’t capture this deeper, native-level nuance, making native Danes the ideal judges of a model’s cultural awareness. He will present a study in which Danes challenge large language models with culturally specific prompts to see how well these systems understand Danish culture.

DIKU Bits is a lecture series for students (and (Ph.D.) employees) about current research at DIKU, held every other week . The lectures aim to inspire choices of elective courses, project writing, etc., but also to provide greater insight into colleagues’ research fields and create a common meeting space for both staff and students.