TextSeal2
A distortion-free large language model watermarking technique that merges multiple keys into a virtual key and couples noise using a Gaussian copula. It improves detection capability while keeping the output distribution unchanged, and it is compatible with speculative decoding.
Timeline (last 90 days)
TextSeal2 fuses two or more keys into one virtual key and couples it with fresh noise via AutoMark's Gaussian copula.
Any number of keys (including publicly shared keys) can be merged in the TextSeal2 way without weakening detection strength.
Under speculative decoding, the replacement tokens for rejected drafts in TextSeal2 still use the same key to maintain the watermark, and the acceptance rate is unchanged.
TextSeal2 runs another keyed exponential race on the residual of the draft competition, exploiting the memorylessness of the exponential distribution to reproduce block verification behavior so that the output strictly follows the target model distribution.
The paper evaluates TextSeal2 on three LLMs; at a false positive rate of 10^-3 and equal output token diversity, the detection rate for 100-token answers is 8 to 22 percentage points higher than TextSeal.
Under the same conditions, TextSeal2 performs comparably to textGrain detection and requires no post-processing optimization of the logits.
The paper claims that TextSeal2 adds no measurable inference latency and preserves output diversity.
The above experimental data all come from the paper's self-reported experimental setup, and actual results may vary with model, task and data distribution.
The paper currently does not disclose details on TextSeal2's cross-model generalization, robustness under adversarial attacks, or stability of long-text detection.
All Facts (7)
TextSeal2 fuses two or more keys into one virtual key and couples it with fresh noise via AutoMark's Gaussian copula.
50%UnverifiedAny number of keys (including publicly shared keys) can be merged in the TextSeal2 way without weakening detection strength.
50%UnverifiedUnder speculative decoding, the replacement tokens for rejected drafts in TextSeal2 still use the same key to maintain the watermark, and the acceptance rate is unchanged.
50%UnverifiedTextSeal2 runs another keyed exponential race on the residual of the draft competition, exploiting the memorylessness of the exponential distribution to reproduce block verification behavior so that the output strictly follows the target model distribution.
50%UnverifiedThe paper evaluates TextSeal2 on three LLMs; at a false positive rate of 10^-3 and equal output token diversity, the detection rate for 100-token answers is 8 to 22 percentage points higher than TextSeal.
50%UnverifiedUnder the same conditions, TextSeal2 performs comparably to textGrain detection and requires no post-processing optimization of the logits.
50%UnverifiedThe paper claims that TextSeal2 adds no measurable inference latency and preserves output diversity.
50%