Hello! My name is Deniz, and I am a final-year PhD candidate in Computer Science advised by Prof. Antoine Bosselut at the NLP lab in EPFL, IC.

My research focuses on interpretability of large language models. I study how LLMs acquire, organize, and use internal representations, and how we can reliably interpret these processes. My work has explored how knowledge is localized within model parameters [1], how linguistic representations emerge and evolve throughout pretraining [2], and how the methods we use to probe models shape the conclusions we draw about their internal computation [3]. I also studied observable model behavior in real-world settings through a large-scale collaboration that compiled assessments from 50 EPFL courses [4].

More broadly, I am interested in understanding the mechanisms underlying model behavior, how these mechanisms develop during training, and how they ultimately influence model predictions and generalization.

I used to be a drone whisperer like you. Then I took an (interpretability-shaped) arrow in the knee. Back then, I worked on grounded language and language-guided robot instruction-following with Prof. Stefanie Tellex and Prof. Ellie Pavlick at the H2R and LUNAR laboratories at Brown CS.