Yoav Gelberg

About Me

I'm a graduate student at the University of Oxford, working with Haggai Maron, Michael Bronstein, and Yarin Gal. My research is focused on making language models more adaptable and reliable, closing the gap between what LLMs can do in principle and what they consistently do in deployment. To that end, I work on post-training adaptation, long-context language modeling, misbehavior detection/monitoring, and data efficient architecture design. A recurring theme in my work is exploiting structure in neural parameter spaces and decoding model behavior from weights and activations.

I spent the summer of 2025 at Sakana AI, studying positional embeddings and length generalization in LLMs. Before Oxford, I completed my undergraduate studies at the Technion as a Rothschild scholar, studied robustness of expander graphs at the Weizmann Institute, developed fair learning algorithms at Fairgen, and taught math and programming at the ARDC. When I can find the time, I love baking 🍞, kitesurfing 🪁, and snowboarding 🏂️.