TextSeal2 maintains watermark signals during speculative decoding

2026-10-30

Meta Superintelligence Labs introduces TextSeal2, a distortion-free LLM watermarking method that preserves its signal under speculative decoding without adding measurable inference latency.

On 29 Oct 2026, Meta Superintelligence Labs released TextSeal2. It solves a specific problem with the previous TextSeal version, which used a Gumbel-max scheme to route tokens between two keys via a coin flip. That approach was suboptimal because the detector had to average scores from both keys without knowing which one the model actually used.

TextSeal2 fixes this by merging multiple keys into a single virtual key. It then couples this to fresh noise using AutoMark's Gaussian copula. I think this is a cleaner way to handle key management because it allows for the inclusion of publicly shared keys without diluting the detection signal.

The real win here is how it handles speculative decoding. Most watermarks break when a draft token is rejected and replaced. TextSeal2 uses an additional exponential race, run on the leftovers of the draft races, to verify keyed drafts. Because of memorylessness, this process reproduces block verification.

The result is a system where tokens that replace rejected drafts remain watermarked with the same keys. The acceptance rate of the speculative decoding remains unchanged. I'd argue that any watermark that slows down inference is dead on arrival, so the claim that this adds no measurable latency is the only metric that actually matters for production.

I work in AI and I built brendon.bot to track these shifts. I build agentic AI systems, but the underlying plumbing—like how we track provenance without killing performance—is where the actual friction lives.

I don't know the exact false positive rate of TextSeal2 compared to the original TextSeal, as the preview doesn't list the specific detection accuracy percentages. Without those numbers, we can't be sure if the "virtual key" merge introduces any noise into the detection confidence.

Still, moving from a coin-flip routing system to a copula-based merge is a smart move. It avoids the blunt averaging that plagued the first version. If this actually holds up under heavy speculative drafting, it makes watermarking a viable safety tool rather than a research curiosity that ruins the user experience.