Reproduce “No Tokenizer, Still a Gap: Cross-Lingual Fairness of Homoglyph and Invisible-Character Watermarking” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alexander Nemecek, Yuqiao Xu, Joseph Falkenburg, Erman Ayday
The authors measure Unicode watermark capacity across 40 language varieties and 24 script groups. The method substitutes lookalike characters or inserts invisible characters without tokenization. Capacity ranges from 7.09 bits per character for English and Swahili to 2.14 for Khmer; script accounts for 96% of variation. After modeled copy-paste, a cover-informed detector needs 92.96 code points for Lao, an upper-bound measurement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.