Paper Overview
- Field: NLP
- Authors: Tom Sander, Pierre Fernandez, Itai Gat
- Date: 2025-10-28
- arXiv: 2510.25907
- At the same output-token diversity, TextSeal2 detects 8 to 22 percentage points more 100-token answers than TextSeal at a false positive rate of 10⁻³, performing on par with textGrain but without requiring post-hoc logit optimization.
- Under speculative decoding, at equal acceptance rates, it detects 18 and 22 percentage points more 200-token answers than TextSeal and textGrain, respectively.
Summary
We 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.
Its predecessor, TextSeal, restores diversity in the Gumbel-max scheme by routing each token to one of two keys via a coin flip. This is suboptimal because the detector averages the scores from both keys without knowing which key was used.
TextSeal2 instead merges two or more keys into a single virtual key and couples it to fresh noise through AutoMark's Gaussian copula. Any number of keys, including publicly shared keys, can be merged this way without diluting detection.
Under speculative decoding, TextSeal2 verifies key-carrying drafts through an additional exponential race run over the remainder of the draft contest. By memorylessness, this race reproduces block verification, so the output exactly follows the target model.
Results
TextSeal2 was evaluated on three LLMs: