Most watermarking schemes for large language models rely on token IDs, which are tied to a specific tokenizer's vocabulary. That can make detection brittle if text is re-tokenized or processed by a different tokenizer. A new approach, Tokenization-Invariant Watermarking (TI-WM), sidesteps this by moving the detection layer to bytes.

TI-WM embeds a watermark during token generation, but it keys and scores the watermark on canonicalized UTF-8 bytes rather than on token IDs. The detector never invokes the tokenizer, so it does not depend on how the original text was tokenized.

Because the method is byte-local and tokenization-invariant, it is designed to remain robust across different tokenizers and text-processing pipelines. The work is described in a recent arXiv paper and offers a potential path toward more portable watermarking for LLM-generated text.