Research map / Large language models / Language, safety and interpretability
Catch-22: On the Fundamental Tradeoff Between Detectability and Robustness in LLM Watermarking
Kuheli Pratihar, Debdeep Mukhopadhyay
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 DIPPER paraphrasing attack to HCW (bias-free) | HCW · LFQA Llama-2-7B | AUROC: .99 -> .91 | Hurt |
| use semantic PMark instead of bias-free HCW under DIPPER | watermark family · LFQA, DIPPER | AUROC: .91 -> .94 | Improved |
| use semantic PMark instead of bias-free HCW under OPT-2.7B | watermark family · LFQA, OPT-2.7B | AUROC: .92 -> .95 | Improved |
| increase edit rate to back-translation (~0.42) for bias-free schemes | bias-free watermarks · LFQA Llama-2-7B | AUROC: .99 -> .55 to .60 | Hurt |
| use semantic PMark instead of HCW under back-translation | watermark family · LFQA, back-translation | AUROC: .59 -> .85 | Improved |
| use hybrid selection instead of best single family under back-translation | Hybrid · LFQA, back-translation | AUROC: .85 -> .86 | Improved |
| use hybrid instead of biased KGW (no attack) | Hybrid · LFQA, no attack | keyless z-score: 30.1 -> −7.8 | Improved |
| use distribution-preserving CGW instead of KGW under DIPPER | CGW · LFQA, DIPPER | AUROC: .86 -> .50 | Hurt |
| use hybrid instead of KGW under DIPPER | Hybrid · LFQA, DIPPER | TPR at 1% FPR: .64 -> .89 | Improved |
| use co-designed DAWA instead of Hybrid under DIPPER | DAWA · LFQA, DIPPER | AUROC: .94 -> .54 | Hurt |
| use conservative upper-confidence-bound selector (ablation) | Hybrid selector · Appendix G sensitivity | worst-case regret | Improved |
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.
- GaussMark — GaussMark: A Practical Approach for Structural Watermarking of Language Models (ICML 2025) · better on 2 of 2 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
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.91)
- On the Reliability of Watermarks for Large Language Models (ICLR 2024 · similarity 0.91)
- Provable Robust Watermarking for AI-Generated Text (ICLR 2024 · similarity 0.91)
- Black-Box Detection of Language Model Watermarks (ICLR 2025 · similarity 0.91)
- Break the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models (ICLR 2026 · similarity 0.91)
- Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach (NeurIPS 2025 · similarity 0.91)
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.
