New preprint "Improving Confidence Estimation in Language Models with Linear Representations of Ambiguity" is now on arXiv.
Hi, I’m Themistoklis (Themis) Nikas.
I’m a CS PhD student at Boston University, where I work with Professor Aaron Mueller. My research focuses on mechanistic interpretability of large language models (LLMs). In particular, I study how LLMs internally represent uncertainty. Recently, I’ve been using linear representations of ambiguity to make models better calibrated through fine-tuning.
Before Boston University, I earned an integrated Master’s degree in Computer Science and Engineering from the University of Ioannina, where I worked with Professor Panayiotis Tsaparas on link analysis algorithms.
News
Attended the Athens NLP Summer School (AthNLP) 2026.
Co-organized the 3rd New England Mechanistic Interpretability Workshop (NEMI) at BU.
Presented our work "Minimizing the Cost of PageRank Fairness" at WebGraph @ WSDM 2026.