Text watermarking · 2026

SAMark

A self-anchored semantic watermark built to survive paragraph-level paraphrasing.

Jiahao Huo^2,3, Wenjie Qu⁴, Yibo Yan^1,2, Kening Zheng³, Jiaheng Zhang, Xuming Hu^1,2, Philip S. Yu³, Mingxun Zhou¹

¹ The Hong Kong University of Science and Technology · ² The Hong Kong University of Science and Technology (Guangzhou) · ³ University of Illinois Chicago · ⁴ National University of Singapore

Paper / arXiv ↗ Code ↗ Motivation & method ↓

Project overview

SAMark combines self-anchored semantic regions, multi-channel scoring, and diversity-aware filtering.

Motivation

Robust watermarking requires a stable semantic reference.

The central challenge is to preserve detectability after substantial rewriting while maintaining the fluency and semantic diversity of generated text.

Most token-level watermarks encode signals through local vocabulary perturbations; these signals can be weakened by paraphrasing. Existing semantic alternatives improve invariance, but often rely on sequentially propagated states or partitions tied to preceding sentences. Paragraph reordering, insertion, or deletion can therefore desynchronize generation and detection.

SAMark addresses this dependency by deriving the watermark reference from each sentence itself. This self-anchored formulation makes the detection unit locally reproducible and motivates a semantic signal that is both paraphrase-resistant and independent of global sentence order.

TaskSemantic text watermarking

Robustness targetParagraph paraphrase

Core ideaSelf-anchored regions

Main findings

Empirical results

90.2%

Reported TP@FP1% under representative paragraph-level paraphrase attacks.

+30.6 pts

Average gain over the strongest baseline in the reported robustness comparison.

2 signals

Multi-channel semantic scoring gives the detector more stable evidence to aggregate.

Methodology

Self-anchored semantic watermarking

SAMark constructs and verifies the watermark at the sentence level through three coordinated components.

Self-anchored region construction

For each generated sentence, SAMark computes a semantic anchor from the sentence context and uses it to define a reproducible watermark region. Because the anchor is local, detection does not require recovering the original sentence sequence.

Multi-channel hyperbolic scoring

Candidate sentences are represented in hyperbolic semantic space and evaluated through complementary channels. Their aggregated scores determine whether a candidate lies in the designated region and strengthen the statistical evidence available to the detector.

Diversity-aware generation and detection

A diversity-aware filter removes semantically redundant or low-quality candidates before sampling. At inference time, sentence-level scores are recomputed and aggregated into a document-level hypothesis test for watermark detection.

Visual results

Method analysis and experimental results

The figures below show the design intuition, robustness frontier, and behavior under word-level attacks.

SAMark overview diagram — **Overview**Self-anchored generation and paragraph-level detection form one consistent watermarking pipeline.

SAMark robustness quality Pareto frontier — **Robustness–quality frontier**The comparison reports stronger detection under paraphrasing while maintaining competitive generation quality.

SAMark detection trace — **Detection trace**A token- and sentence-level view of the evidence accumulated by the detector.

SAMark geometric semantic landscape — **Semantic geometry**Geometric separation makes the self-anchored signal interpretable and reusable across contexts.

SAMark word substitution attack results — **Word substitution attacks**Performance remains stable as local word substitutions perturb the watermarked text.

SAMark word deletion attack results — **Word deletion attacks**Deletion stress tests complement the paragraph-level paraphrase evaluation.

Citation

BibTeX

@article{huo2026samark,
  title   = {SAMark: A Self-Anchored Text Watermarking with Paragraph-Level Paraphrase Robustness},
  author  = {Huo, Jiahao and Qu, Wenjie and Yan, Yibo and Zheng, Kening and Zhang, Jiaheng and Hu, Xuming and Yu, Philip S. and Zhou, Mingxun},
  journal = {arXiv preprint arXiv:2605.25796},
  year    = {2026}
}