VocabX rebuilds a small model's vocabulary to cut inference cost by 77%
A Hugging Face blog post previews the VocabX paper, which rebuilds a model's tokenizer around the cost of each token rather than how often it appears. The authors say that on an automated Portuguese essay grader built on Qwen3-4B, the change cut measured inference cost by 77% and raised the bench score from 0.703 to 0.810, with the base model and training recipe unchanged. The method picks tokens by the price of the field they appear in, and builds both sub-word pieces and multi-word spans into a single token. The numbers come from a paper still under review and may change before publication.