TextSeal2: Distortion-Free LLM Watermarking with Copula Mixing and Memoryless Races
In three-model tests, TextSeal2 detected 8–22 percentage points more 100-token answers than TextSeal at equal output diversity.
That result used a false-positive rate of 0.1%. The method merges multiple watermark keys into a single virtual key, avoiding the detection penalty of averaging scores across keys. Under speculative decoding, replacement tokens retain the watermark while the output follows the target model’s distribution and the acceptance rate stays unchanged. The authors report no measurable inference latency and detection gains of 18 percentage points over TextSeal and 22 over textGrain on 200-token answers under speculative decoding.
ArXiv · AI/CL/LG's note
That result used a false-positive rate of 0.1%. The method merges multiple watermark keys into a single virtual key, avoiding the detection penalty of averaging scores across keys. Under speculative decoding, replacement tokens retain the watermark while the output follows the target model’s distribution and the acceptance rate stays unchanged. The authors report no measurable inference latency and detection gains of 18 percentage points over TextSeal and 22 over textGrain on 200-token answers under speculative decoding.
ArXiv · AI/CL/LG's note
score 5