TextSeal: A Localized LLM Watermark for Provenance & Distillation Protection
2026-05-12 • Cryptography and Security
Cryptography and SecurityComputation and LanguageMachine Learning
AI summaryⓘ
The authors present TextSeal, a new method to watermark text generated by large language models without reducing quality. It uses advanced sampling and scoring techniques to help detect AI-generated text accurately, even when mixed with human writing. TextSeal works efficiently during text generation and remains undetectable in terms of quality by human readers. Additionally, the watermark can be identified even if the model was copied or simplified. The authors tested it thoroughly and found it preserves model performance while enabling strong, reliable detection.
watermarkinglarge language modelsGumbel-max samplingentropy weightingspeculative decodingmodel distillationprovenance detectioninference overheadtext generationlocalized detection
Authors
Tom Sander, Hongyan Chang, Tomáš Souček, Tuan Tran, Valeriu Lacatusu, Sylvestre-Alvise Rebuffi, Alexandre Mourachko, Surya Parimi, Christophe Ropers, Rashel Moritz, Vanessa Stark, Hady Elsahar, Pierre Fernandez
Abstract
We introduce TextSeal, a state-of-the-art watermark for large language models. Building on Gumbel-max sampling, TextSeal introduces dual-key generation to restore output diversity, along with entropy-weighted scoring and multi-region localization for improved detection. It supports serving optimizations such as speculative decoding and multi-token prediction, and does not add any inference overhead. TextSeal strictly dominates baselines like SynthID-text in detection strength and is robust to dilution, maintaining confident localized detection even in heavily mixed human/AI documents. The scheme is theoretically distortion-free, and evaluation across reasoning benchmarks confirms that it preserves downstream performance; while a multilingual human evaluation (6000 A/B comparisons, 5 languages) shows no perceptible quality difference. Beyond its use for provenance detection, TextSeal is also ``radioactive'': its watermark signal transfers through model distillation, enabling detection of unauthorized use.