TextSeal2: Distortion-Free LLM Watermarking with Copula Mixing and Memoryless Races
TextSeal2 keeps LLM watermarks intact during speculative decoding without changing output diversity or acceptance rates. Its reported detection gains make it relevant to builders evaluating provenance for generated text.
TextSeal2 combines multiple watermark keys with Gaussian copula noise, then uses an extra race over leftover draft races during speculative decoding. The authors report **no measurable inference latency** and unchanged acceptance rates across **three LLMs**.
For agent builders, this offers a watermarking approach designed to preserve the target model’s output distribution while supporting speculative decoding. Consider it when evaluating provenance for generated text or serving stacks that use draft models.
TextSeal2 combines multiple watermark keys with Gaussian copula noise, then uses an extra race over leftover draft races during speculative decoding. The authors report **no measurable inference latency** and unchanged acceptance rates across **three LLMs**. For agent builders, this offers a watermarking approach designed to preserve the target model’s output distribution while supporting speculative decoding. Consider it when evaluating provenance for generated text or serving stacks that use draft models. At a false-positive rate of **10⁻³**, it detects **8–22 percentage points** more 100-token answers than TextSeal at equal diversity. Under speculative decoding, it gains **18 and 22 points** over TextSeal and textGrain on 200-token answers; the material gives no deployment code or independent replication details.
TextSeal2 addresses provenance during speculative decoding, reporting watermark detection gains without measurable latency or changed acceptance rates in its tested models. Its evidence narrows to those comparisons; it does not establish deployment performance or independent replication.