MultiTok: Variable-Length Tokenization for Efficient LLMs Adapted from LZW Compression

arXiv cs.CL / 4/27/2026

💬 OpinionDeveloper Stack & InfrastructureTools & Practical UsageModels & Research

共有:

Key Points

The paper introduces MultiTok, a variable-length tokenization method inspired by LZW universal compression to compress repetitive phrases into multi-word tokens for LLM training.
It argues that this approach can reduce training resource requirements—such as data volume and compute—while maintaining similar accuracy to established tokenizer and model baselines.
Experiments report that MultiTok achieves comparable performance to BERT and GPT standards both as a standalone tokenizer and as an add-on to existing tokenizers.
The authors claim roughly 2.5× faster training and over 30% less training data usage compared with conventional approaches.
Overall, MultiTok is positioned as a practical tokenization upgrade aimed at improving efficiency without sacrificing downstream language modeling quality.

Abstract

Large language models have drastically changed the prospects of AI by introducing technologies for more complex natural language processing. However, current methodologies to train such LLMs require extensive resources including but not limited to large amounts of data, expensive machinery, and lengthy training. To solve this problem, this paper proposes a new tokenization method inspired by universal Lempel-Ziv-Welch data compression that compresses repetitive phrases into multi-word tokens. With MultiTok as a new tokenizing tool, we show that language models are able to be trained notably more efficiently while offering a similar accuracy on more succinct and compressed training data. In fact, our results demonstrate that MultiTok achieves a comparable performance to the BERT and GPT standards as both a stand-alone tokenizer and an add-on to existing tokenizers while also providing close to 2.5x faster training with more than 30% less training data.