Meta's TextSeal2: LLM watermarking survives speculative decoding, detects 8-22pp more
rohanpaul_ai · x · 2026-11-02
Meta researchers released TextSeal2 (arXiv:2610.40613), an improved distortion-free watermark for LLM text. The predecessor TextSeal routed each token to one of two keys via a coin flip, forcing the detector to average both scores and dilute the signal. TextSeal2 instead merges two or more keys into a single virtual key coupled to fresh noise via AutoMark's Gaussian copula. It keeps watermarking intact under speculative decoding — replacing tokens stay watermarked and a memoryless exponential race reproduces block verification so output follows the target model exactly, with no measurable latency or diversity loss. On three LLMs, at equal diversity and 10^-3 false-positive rate, it detects 8 to 22 percentage points more 100-token answers than TextSeal, and even publicly shared keys can be merged without dilution.
More from Safety
- Gartner survey: AI-discovered cyber vulnerabilities top emerging risks — krishnan · 2026-11-03
- Profile: Thiel protégé Michael Kratsios runs the White House office steering AI policy — saranormous · 2026-11-03
- OpenAI deploys activation classifiers (probes) on frontier models starting from GPT-5.6 — bobabowen · 2026-11-03
- Xbow breaks down the $40K GPT-5.6-sol API bill behind its full-chain Chrome exploit — moyix · 2026-11-03
- Europe's Physical AI Agenda: Gigafactories, Sovereign Compute and AI Act Rules — ingliguori · 2026-11-03
- Cryptographer Matthew Green pushes back on LLM cryptanalysis fears: best AES-128 attacks still need 2^89 work — matthew_d_green · 2026-11-03