TextSeal2 watermarks LLM output under speculative decoding without distortion
Original title: TextSeal2: Distortion-Free LLM Watermarking with Copula Mixing and Memoryless Races
At equal diversity and a 10^-3 false-positive rate, it detects 8 to 22 points more 100-token answers than TextSeal.
Researchers present TextSeal2, a distortion-free large language model watermark that keeps rejected speculative drafts watermarked with the same keys and leaves the acceptance rate unchanged. It merges two or more keys, including public keys, into one virtual key via AutoMark's Gaussian copula and uses a memoryless exponential race so output matches the target model. Evaluated on three LLMs, it adds no measurable inference latency and needs no post-logits optimization.
- At equal token diversity, detection of 100-token answers rises 8 to 22 points over TextSeal.
- Under speculative decoding it beats TextSeal by 18 points and textGrain by 22 on 200-token answers.
- Any number of keys can be merged without diluting detection.
- Output diversity is preserved and inference latency is not measurably increased.
From the original article
1 min read at arxiv.orgWe present TextSeal2, a distortion-free watermark for large language models that keeps its signal under speculative decoding: tokens that replace rejected drafts remain watermarked with the same keys, and the acceptance rate is unchanged. It also preserves output diversity and adds no measurable inference latency.
By Tom Sander · 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.