原帖引用 · 原文
New Meta paper shows that blending 2 secret keys into 1 makes hidden watermarks in AI-written text easier to detect, even when the AI uses a common trick (like speculative decoding) to write faster.
Watermarks hide a secret-key signal in which tokens an LLM samples, so a detector can later flag its text. Meta's earlier TextSeal picked 1 of 2 keys per token at random, so detection had to average both and lost signal.
TextSeal2 blends both keys into 1 virtual key plus fresh noise, so answers stay varied and every token carries full signal. For speculative decoding, it picks replacement tokens from the leftovers of the drafts' random races, so acceptance rates stay unchanged.
On Qwen3.5-9B with speculative decoding and equal diversity, it detected 89% of 200-token answers at a 0.1% false-positive rate. TextSeal caught 71% and textGrain 67%, at the same acceptance rate.