MrT5: Dynamic Token Merging for Efficient Byte-level Language Models

Date Presented: 3/27/25
Speaker: Julie Kallini, Stanford University

Visit links below to subscribe and for details on upcoming seminars:
https://www.isi.edu/isi-seminar-series/
https://www.isi.edu/events/

Abstract: Models that rely on subword tokenization have significant drawbacks, such as sensitivity to character-level noise like spelling errors and inconsistent compression rates across different languages and scripts. While character- or byte-level models like ByT5 attempt to address these concerns, they have not gained widespread adoption—processing raw byte streams without tokenization results in significantly longer sequence lengths, making training and inference inefficient. This work introduces MrT5 (MergeT5), a more efficient variant of ByT5 that integrates a token deletion mechanism in its encoder to dynamically shorten the input sequence length. After processing through a fixed number of encoder layers, a learned delete gate determines which tokens are to be removed and which are to be retained for subsequent layers. MrT5 effectively “merges” critical information from deleted tokens into a more compact sequence, leveraging contextual information from the remaining tokens. In continued pre-training experiments, we find that MrT5 can achieve significant gains in inference runtime with minimal effect on performance, as measured by bits-per-byte. Additionally, with multilingual training, MrT5 adapts to the orthographic characteristics of each language, learning language-specific compression rates. Furthermore, MrT5 shows comparable accuracy to ByT5 on downstream evaluations such as XNLI, TyDi QA, and character-level tasks while reducing sequence lengths by up to 75%. Our approach presents a solution to the practical limitations of existing byte-level models.

Speaker's Bio: Julie Kallini is a second-year Ph.D. student in Computer Science at Stanford University, advised by Christopher Potts and Dan Jurafsky. Her research focuses on natural language processing (NLP), with an emphasis on computational linguistics/cognitive science, tokenization, and model architecture. Her paper, "Mission: Impossible Language Models," won Best Paper Award at ACL 2024. Her work is supported by the NSF Graduate Research Fellowship, the Stanford School of Engineering Graduate Fellowship, and the Stanford EDGE Fellowship.
Before starting her Ph.D., Julie was a software engineer at Meta, where she worked on machine learning for advertisements. Julie graduated summa cum laude from Princeton University with a B.S.E. in Computer Science and a minor in Linguistics.

Видео MrT5: Dynamic Token Merging for Efficient Byte-level Language Models канала USC Information Sciences Institute