I am a third-year PhD student at the Language Technology Lab of the University of Cambridge. My research focuses on:

• Maximally resource-efficient synthesis of insights and models: I believe retrofitting is a highly promising, underrated area of research that may allow gaining architectural insights and creating models at a small fraction of the cost of current common practices. I also believe in the utility of creating effective ultra-efficient task-specialized models.
• Improving modularity and malleability of LLMs: As ecosystems emerge around LLMs, there is a risk of these ecosystems remaining mutually incompatible. I aim to improve the modularity (degree of compartmentalization) and malleability (ease of adaptation) of LLMs to remove barriers across LLM ecosystems, aiming to enable new ways of information interchange and collaboration across LLMs.
• Making units of computation (a.k.a. tokens) adaptive: A fundamental restriction of current LLMs is that the tokenization remains largely fixed after the one-time decision how to tokenize. I work on ways to make tokenization adaptive, such as Zero-Shot Tokenizer Transfer and Cross-Tokenizer Distillation, with the primary goal of improving inference efficiency and equity across languages.

I’ve been fortunate to gain research experience at Ai2, Google DeepMind, Cohere, H2O.ai, Huawei Noah’s Ark Lab (2x), Johannes Kepler University Linz (where I also did my undergrad) and through competing on Kaggle.

If you’d like to chat about any of these topics (or about 🐱⛵🥾), drop me an email!