ByteSpan: Information-Driven Subword Tokenisation

ICML 2025 Workshop TokShop Submission45 Authors

Archiving Submission: No (non-archival)

Previous Venue If Non Archival: It has not been accepted or published at another venue but will be released on Arxiv.

Keywords: morphological segmentation, subword representations, information

TL;DR: We propose a new information-driven subword tokeniser which uses an external byte-level LM to identify predictable contiguous byte sequences.

Abstract: Recent dynamic tokenisation methods operate directly on bytes and pool their latent representations into patches. This bears similarities to computational models of word segmentation that determine lexical boundaries using spikes in an autoregressive model's prediction error. Inspired by this connection, we explore whether grouping predictable bytes — rather than pooling their representations — can yield a useful fixed subword vocabulary. We propose a new information-driven subword tokeniser, ByteSpan, that uses an external byte-level LM during training to identify contiguous predictable byte sequences and group them into subwords. Experiments show that ByteSpan yields efficient vocabularies with higher morphological alignment scores than BPE for English. Multilingual experiments show similar compression and rényi efficiency for 25 languages.

Submission Number: 45