Towards Reliable Marking and Verification of AI-Generated Text via Geometry-aware Sentence-level Watermarking

Large generative models raise growing concerns about provenance, misinformation, and impersonation. Digital watermarking offers a principled solution, yet extending it to natural language remains challenging due to text discreteness and sensitivity to semantic perturbations. Existing text watermarking methods either operate at the token level requiring white-box access and remaining fragile to paraphrasing, or at the sentence level, which supports black-box deployment but suffers from low Watermark Success Rate (WSR). We show that low WSR in sentence-level watermarking primarily stems from low injection success probability caused by a mismatch between posterior embedding distributions and semantic accept regions. Based on this insight, we propose \textbf{X-Guard}, a geometry-aware sentence-level watermarking framework that improves injection success by jointly optimizing embedding distributions and semantic space partitioning. X-Guard learns a more isotropic embedding space and introduces \textbf{A$^2$PQ}, a centroid-aligned partitioning scheme that approximately equalizes probability mass across regions. Extensive experiments across multiple models, languages, and attack settings demonstrate that X-Guard consistently improves robustness while preserving text fluency and practical deployability.

Generative AI can now produce fluent text at scale, which makes it harder to tell where a piece of text came from and easier to spread misinformation or impersonate others. One way to address this problem is to add an invisible watermark to AI-generated text, so that the text can later be checked for its origin. However, watermarking natural language is difficult because small changes in wording can change the hidden signal or make it disappear. This paper studies sentence-level watermarking, where the watermark is embedded in the meaning of whole sentences rather than in individual word choices. Such methods are easier to use in practical settings, but existing approaches often fail to embed the watermark reliably. We find that this failure happens because the possible meanings produced by a language model do not align well with the semantic regions used to encode the watermark. To solve this, we propose X-Guard, a method that reorganizes the sentence meaning space so that watermark signals can be inserted more reliably. It also partitions this space more evenly, making each hidden signal easier to recover. Experiments across different models, languages, and attacks show that X-Guard makes sentence-level watermarks more robust while keeping the generated text fluent and practical to use.