Xenia Ohmer

Xenia Ohmer

Postdoc, NLP researcher, University of Osnabrück

About me

Hi, I am Xenia, a Postdoctoral Researcher at the Institute of Cognitive Science, University of Osnabrück. Currently, I’m part of the Natural Language Processing Group, where I am fortunate to be working under the supervision of Elia Bruni.

Research interests

My research revolves around Natural Language Processing, Computational Linguistics, and Cognitive Science, with a focus on the following areas:

  • Evaluation and interpretation of large language models
  • Grounding and intuitive physics understanding in large language models
  • Language emergence and multi-agent communication
  • Using deep neural networks to model cognitive phenomena

Short bio

I pursued my Bachelor’s in Cognitive Science at both the University of Osnabrück and the University of Zürich (during an ERASMUS program). For my Master’s, I studied Mathematics in Medicine and Life Sciences at the University of Lübeck. After a brief stint in Computational Molecular Biology research in Berlin, I went back to Osnabrück to do a Ph.D. in the Research Training Group Computational Cognition. My Ph.D. was jointly supervised by Peter König, Professor for “Neurobiopsychology” and Michael Franke, who now serves as a Professor at the Department of Linguistics at the University of Tübingen. During my Ph.D., I worked on various projects related to neural network models of word learning, multi-agent communication, language emergence, and pragmatics (see my thesis). Recently, I started my Postdoc at the Natural Language Processing Group, University of Osnabrück, where I am supervised by Elia Bruni. While I continue to research the topics of my Ph.D., I’m also deeply fascinated by the latest advancements in NLP, leading me to explore projects involving the evaluation and interpretation of large language models.

News

  • 22.04.2024 - New preprint with Dieuwke Hupkes and Elia Bruni: Equipped with our multisense consistency method, we dive deep into an exploration of the semantic understanding of LLMs. Our experiments show that LLMs “understanding” is (still?) linked to form.
  • 16.04.2024 - Our GRASP benchmark for evaluating the language grounding and physical understanding capabilities of video-based multimodal LLMs has been accepted to IJCAI 2024.
  • 20.02.24 - Two papers accepted at LREC-COLING! One with Leon Ackermann investigating the relationship between skill neurons and robustness in prompt tuning and one with Kristina Kobrock, Elia Bruni, and Nicole Gotzner on the role of context in emergent communication.