Research map / Large language models / Language, safety and interpretability
Towards Reliable Marking and Verification of AI-Generated Text via Geometry-aware Sentence-level Watermarking
Yubing Ren, Ping Guo, Yanan Cao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune BGE-m3 encoder with alignment-uniformity loss (ablation) | X-Guard · High-resource, attack average | average WSR: 0.627 -> 0.643 | Improved |
| fine-tune BGE-m3 encoder with alignment-uniformity loss (ablation) | X-Guard · Low-resource, no attack | WSR: 0.955 -> 0.956 | No clear effect |
| swap encoder to XLM-R or Qwen3-Embedding-0.6B (ablation) | X-Guard · High/low-resource | WSR: 0.616 / 0.641 attack avg (high-res) | No clear effect |
| increase number of partitions |Omega| from 2 to 8 (ablation) | X-Guard · injection success rate | injection success rate: 0.80 -> nearly 1.00 | Improved |
| increase number of partitions |Omega| to 16 (ablation) | X-Guard · attack robustness | robustness: above 0.95 -> around 0.70 | Hurt |
| raise watermarked region ratio gamma (ablation) | X-Guard · Chinese CulturaX 100 sentences | detection vs false positives | Mixed |
| use A2PQ centroid-aligned partitioning with isotropic embeddings instead of SemStamp LSH | X-Guard · Toy injection success probability | injection success probability: 44.9% -> 89.9% | Improved |
| X-Guard vs SemStamp sentence-level watermark | X-Guard · Text continuation, low-resource, optimized adaptive attack | WSR: 25.30% -> 35.97% | Improved |
| X-Guard vs token-level watermarks under paraphrase | X-Guard · Text continuation, high-resource, mono-lingual paraphrase | WSR: 0.15% (KGW) -> 49.61% (X-Guard) | Improved |
| X-Guard vs SemStamp downstream quality | X-Guard · QMsum / MultiNews low-resource | WSR: 0.123 -> 0.490 (QMsum) | Improved |
| X-Guard vs SemStamp generation time | X-Guard · relative TTLT | generation time: x18.0 -> x5.5 | Improved |
Closest papers by meaning
- Learning to Watermark: A Selective Watermarking Framework for Large Language Models via Multi-Objective Optimization (NeurIPS 2025 · similarity 0.88)
- Adaptive Text Watermark for Large Language Models (ICML 2024 · similarity 0.88)
- Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach (NeurIPS 2025 · similarity 0.87)
- A Semantic Invariant Robust Watermark for Large Language Models (ICLR 2024 · similarity 0.87)
- StealthInk: A Multi-bit and Stealthy Watermark for Large Language Models (ICML 2025 · similarity 0.87)
- Provable Robust Watermarking for AI-Generated Text (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
