TextSeal
A distortion-free LLM watermarking method based on Gumbel-max sampling that distributes tokens among multiple keys via key routing, but detection requires averaging scores across multiple keys.
Timeline (last 90 days)
TextSeal routes each token to one of two keys by a coin flip, restoring output diversity in the Gumbel-max sampling scheme.
TextSeal's detector needs to compute the scores of both keys and average them, yet has no way of knowing which key was actually used during generation, causing the detection signal to be diluted.
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.
All Facts (3)
TextSeal routes each token to one of two keys by a coin flip, restoring output diversity in the Gumbel-max sampling scheme.
50%UnverifiedTextSeal's detector needs to compute the scores of both keys and average them, yet has no way of knowing which key was actually used during generation, causing the detection signal to be diluted.
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%