HiMark adds controllable multi-bit watermarks to LLM text
Original title: HiMark: Hierarchical and Controllable Multi-bit Watermarking for Large Language Models
With 32 bits in 300 tokens, authors report 90.01% extraction accuracy after paraphrasing while preserving text quality.
Researchers propose HiMark, a hierarchical multi-bit watermark for text generated by large language models. Embedding uses candidate sampling, tournament elimination, support-based refinement, and threshold filtering. Detection aggregates model-agnostic token evidence into votes for extraction. The authors say the method improves extraction accuracy without sacrificing text quality, including under paraphrasing.
Unconfirmed: Accuracy and quality figures are the authors’ experimental results, not an independent evaluation.
- The reported paraphrasing result embeds 32 bits in text of 300 tokens.
- Detection accumulates multi-layer support evidence from each observed token.
- The stated goal is provenance, copyright protection, and accountability for generated text.
From the original article
1 min read at arxiv.orgWith the rapid adoption of large language models (LLMs) in text generation applications, it has become increasingly difficult to trace generated content and assign responsibility for its use, raising concerns about provenance, copyright protection, and accountability.
By Ke Zhu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.