Meta's TextSeal2 watermark survives speculative decoding, detects 89% of answers

rohanpaul_ai · x · 2026-11-02

Meta Superintelligence Labs published TextSeal2, a distortion-free LLM watermarking method that keeps full detection signal even under speculative decoding.

Problem: the earlier TextSeal preserved output diversity by flipping a coin between two keys per token, forcing detectors to average both scores and lose signal.
Method: TextSeal2 merges two or more keys into a single virtual key coupled to fresh noise via AutoMark's Gaussian copula (even public keys can be merged without diluting detection). Under speculative decoding, replacement tokens are drawn from the leftovers of the draft races via one extra memoryless exponential race, so acceptance rates and the output distribution stay unchanged.
Results: on Qwen3.5-9B with speculative decoding at equal diversity, TextSeal2 detects 89% of 200-token answers at a 0.1% false-positive rate — 18 and 22 percentage points above TextSeal (71%) and textGrain (67%) — with no post-logits optimization and no added latency.

Code is open-sourced at facebookresearch/textseal.

Original post →

Meta's TextSeal2 watermark survives speculative decoding, detects 89% of answers

More from Models