Compute Optimal Tokenization

arXiv cs.CL / 5/5/2026

📰 NewsIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper studies how token “information granularity,” controlled by compression rate (average bytes per token), changes observed language-model scaling trends.
By training 988 latent tokenized models (BLT) from 50M to 7B parameters with configurable compression rates, the authors probe tokenization effects far beyond the typical ~4.57 bytes/token of common BPE.
The results suggest that in compute-optimal settings, the best model size scales with data size measured in bytes rather than with token count, challenging the common “scale by tokens” intuition.
The optimal compression rate is different from BPE-derived values and tends to decrease as compute increases, and the conclusions extend to both latent and subword tokenization as well as non-English languages.

Abstract

Scaling laws enable the optimal selection of data amount and language model size, yet the impact of the data unit, the token, on this relationship remains underexplored. In this work, we systematically investigate how the information granularity of tokens, controlled by the compression rate (i.e., average bytes of text per token), affects scaling trends. We train 988 latent tokenized models (BLT) ranging from 50M to 7B parameters that enable setting the desired compression rate. This flexibility allows us to study the role of compression rate well beyond 4.57 bytes per token obtained with a popular BPE tokenizer. Our experiments reveal that in compute-optimal configurations, model parameter counts scale proportionally to data size measured in bytes, not in tokens as commonly perceived (Kaplan et al., 2020; Hoffmann et al., 2022). Furthermore, we discover that the optimal compression rate differs from the one obtained with BPE and decreases with compute. These findings generalize to both latent and subword tokenization, as well as to languages other than English, guiding language model developers on tokenization scheme selection for maximal compute efficiency.