Looking at the fact desk note, I need to fix the signal-to-noise ratio claim: the paper says the signal itself is halved, while the standard deviation falls by only sqrt(2), making the SNR reduction a factor of sqrt(2), not a halving. The prior sentence conflated these. I'll preserve everything else verbatim.
TextSeal2, a watermarking scheme for large language model output built by researchers at Meta Superintelligence Labs, closes a gap that speculative decoding had opened in every prior distortion-free watermark: when a draft model's tokens get rejected and replaced, the replacement text used to lose its signal or split it across keys. According to the paper, posted to arXiv, TextSeal2 detects 18 and 22 percentage points more 200-token watermarked answers than its predecessor TextSeal and a rival method called textGrain, respectively, when compared at the same acceptance rate under speculative decoding.
The mechanics turn on two changes to how the watermark's keys touch the sampler. TextSeal, the team's earlier scheme, routed each token to one of two keys with a coin flip, which meant the detector didn't know which key generated a given token and had to average both keys' scores — the paper states this halves the signal while the standard deviation falls only by a factor of the square root of 2, so the signal-to-noise ratio is reduced by that same square-root-of-2 factor, not halved outright. TextSeal2 instead merges the two keys' values into what the paper calls a single "virtual key" via a Gaussian copula borrowed from a separate method called AutoMark, and couples that virtual key to fresh noise. Every token is then scored against this merged key rather than split between two, and the paper states the construction extends to any number of additional keys — including publicly shared ones — "without diluting detection."
The second change targets speculative decoding specifically. Gumbel-max sampling, the scheme both TextSeal and TextSeal2 build on, is framed in the paper as an "exponential race" between candidate tokens, each running a lane at a speed set by its probability. Because exponential distributions are memoryless, the leftover race time among tokens that lost a drafting round is itself a fresh exponential draw. TextSeal2 exploits that property by running what the paper calls a "residual race" on those leftovers to decide both where the drafted block stops and what replacement token gets emitted — reproducing block verification's own stopping rule exactly, so the paper's Proposition 1 claims the output distribution matches the target model precisely and acceptance is unchanged.
On the operational numbers, the paper reports its decoding step lands "within 0.6%" of TextSeal's latency once key-only computation overlaps the model's forward pass, and states plainly that the scheme "adds no measurable inference latency." Acceptance held steady too: with block verification and 8 drafts per round on Qwen3.5-9B paired with a Qwen3.5-0.8B drafter, unwatermarked decoding emits 2.74 tokens per round, and the paper reports TextSeal2's residual race differs by only +0.02 tokens per round (95% interval -0.01 to +0.05), matching the baseline schemes' acceptance within 0.04 tokens. The one scheme that didn't hold acceptance was textGrain's own speculative-decoding recipe, which verifies between watermarked distributions directly and the paper says loses 5% of tokens per round as a result.
The trade-off shows up outside speculative decoding, in the base case. At matched output diversity on Qwen3.5-27B, the paper reports TextSeal2 trails textGrain by 5 percentage points (interval -10 to 0) at 100 tokens — the only setting among the three models tested where a baseline beats it. The paper also notes that paraphrasing by an LLM "removes almost every watermark," a limit it says applies to token-level schemes generally, not a weakness specific to TextSeal2. And the detection numbers throughout rest on a false-positive rate of 10⁻³ calibrated empirically against 100,000 null p-values, rather than a closed-form statistical guarantee, which the paper states explicitly is because "no closed-form null is exact in practice."
For teams running speculative decoding purely to cut inference cost, the practical read is that watermarking and acceptance-rate optimization no longer have to trade against each other: TextSeal2's residual race lets a provenance signal ride along on drafted tokens without retuning the detector or giving up the tokens-per-round gains speculative decoding exists to deliver — but the gain over simpler schemes narrows or reverses at larger models, so it's worth checking the diversity-detection curve at your specific model size before swapping it in.