Benjamin Minixhofer
b{lastname}@gmail.comI am a fourth-year PhD student in the Language Technology Lab at the University of Cambridge, supervised by Anna Korhonen and Edoardo M. Ponti. During my PhD, I spent time at Ai2, working with Luca Soldaini and Valentin Hofmann, at Google DeepMind in Zürich, and at the University of Edinburgh.
Before my PhD, I did my undergrad in AI at JKU Linz; during this time, I was fortunate to gain research experience at Cohere, Huawei Noah’s Ark Lab (2x) and H2O.ai, as well as by competing on Kaggle.
My research aims to make language models cheaper to build and easier to change; this centers on retrofitting: reusing existing models instead of training new ones from scratch. Examples are moving or connecting a model to a new subword tokenizer (Zero-Shot Tokenizer Transfer, Cross-Tokenizer Distillation) or converting subword-level LLMs to hierarchical byte-level models (Retrofitting Language Models to Operate over Bytes). I also like building small, efficient task-specialized models, such as Segment Any Text for sentence segmentation.
If you’d like to chat about any of these topics (or about 🐱⛵🥾), drop me an email!
Speaker bio
Benjamin Minixhofer is a PhD candidate in the Language Technology Lab at the University of Cambridge, supervised by Anna Korhonen and Edoardo M. Ponti. His research focuses on reusing language models, with work on tokenizer transfer, cross-tokenizer distillation and retrofitting language models to operate over bytes published at venues such as ACL, NeurIPS and Nature. Beyond Cambridge, he has done research at the Allen Institute for AI, Google DeepMind and Cohere, among others.
News
| Oct 10, 2026 | Attended COLM 2026 in San Francisco and had a great time at the panel in the tokenization workshop! |
|---|---|
| Oct 7, 2026 | Bolmo is published in Nature as Retrofitting Language Models to Operate over Bytes! We also released new byteified Qwen and Llama checkpoints in the Bolmo collection on Hugging Face. |
| Jan 21, 2026 | I wrote a (my first!) blog post on Four Ingredients for Successful Language Model Retrofitting. |
| Dec 19, 2025 | Our new preprint Bolmo: Byteifying the Next Generation of Language Models is up on arXiv! |
| Sep 18, 2025 | Universal Cross-Tokenizer Distillation via Approximate Likelihood Matching is accepted at NeurIPS! See you (probably) at EurIPS! |
| Jul 7, 2025 | Started an internship at Ai2 in Seattle over the Summer to work with Luca Soldaini and Valentin Hofmann! |
| Nov 29, 2024 | Talk about «The Past, Present and Future of Tokenization» at the NLIP Seminar in Cambridge. This talk was based on an Invited Lecture at the University of Göttingen in early November. Slides. |
| Nov 27, 2024 | Attended the ELLIS NLP Workshop at Dagstuhl. Some nice photos. |
| Sep 25, 2024 |
Zero-Shot Tokenizer Transfer is accepted at NeurIPS 2024. See you in Vancouver! 
|
| Jul 24, 2024 | I presented Zero-Shot Tokenizer Transfer at Google DeepMind and Mozilla. Slides. |
