Research map / Large language models / Language, safety and interpretability
Stability-Aware Feature Design for Robust Watermark Detection in Machine-Generated Text
Sina Mansouri, Mohit Marvania, Abolfazl Safikhani
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 |
|---|---|---|---|
| add z-score moments (6-D) to global z-score (ablation) | global z-score detector · PG-19 1500 tokens D1 | AUC (%): 74.2 -> 83.2 | Improved |
| add lag-1/lag-2 autocorrelations (8-D) (ablation) | z-score moments detector · PG-19 1500 tokens D1 | AUC (%): 83.2 -> 85.0 | Improved |
| add run-length statistics (14-D) (ablation) | autocorrelation detector · PG-19 1500 tokens D1 | AUC (%): 85.0 -> 87.3 | Improved |
| add run-frequency features (full 20-D) (ablation) | run-length detector · PG-19 1500 tokens D8 | AUC (%): 77.2 -> 78.6 | Improved |
| add paraphrase stability score (PSS) to static features (ablation) | static features detector · PG-19 1500 tokens D1 | AUC (%): 88.1 -> 96.1 | Improved |
| add PSS to static features (ablation) | static features detector · PG-19 1500 tokens D8 | AUC (%): 78.6 -> 91.2 | Improved |
| replace global z-score with WinMax local maximum | global z-score · PG-19 1500 tokens D1 | AUC (%): 74.2 -> 82.1 | Improved |
| use PSS + Static instead of deep-learning detector RADAR | RADAR · PG-19 1500 tokens D8 | AUC (%): 40.8 -> 91.2 | Improved |
| increase watermark greenlist ratio gamma from 0.25 to 0.50 | PSS + Static · PG-19 1500 tokens D8 | AUC (%): 91.2 -> 94.2 | Improved |
| train on PG-19 and test on CNN/DM instead of in-domain | PSS + Static · PG-19 to CNN/DM D1 | AUC (%): 96.1 -> 88.6 | Hurt |
| change paraphraser to Gemma for universal classifier | PSS + Static universal classifier · PG-19 D1 | AUC (%): 96.1 -> 89.1 | Hurt |
| change window/stride across 40-60 and 5-15 (ablation) | PSS + Static · PG-19 1500 tokens mean AUC D1-D8 | AUC (%): 99.0 vs 95.95 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- RADAR — RADAR: Robust AI-Text Detection via Adversarial Learning (NeurIPS 2023) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- On the Reliability of Watermarks for Large Language Models (ICLR 2024 · similarity 0.88)
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.87)
- On the Empirical Power of Goodness-of-Fit Tests in Watermark Detection (NeurIPS 2025 · similarity 0.87)
- Robust Multi-bit Text Watermark with LLM-based Paraphrasers (ICML 2025 · similarity 0.86)
- GaussMark: A Practical Approach for Structural Watermarking of Language Models (ICML 2025 · similarity 0.86)
- Provable Robust Watermarking for AI-Generated Text (ICLR 2024 · similarity 0.86)
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.
