Research map / Large language models / Language, safety and interpretability
Linear Ensembles Wash Away Watermarks: On the Fragility of Distributional Perturbations in LLMs
Zhihao Wu, Gracia Gong, Qinglin Zhu, Yudong Chen, Runcong Zhao
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 |
|---|---|---|---|
| average output distributions of 3 independently watermarked models (WASH) | six watermarking schemes · generation-time z-score | z-score: 5-300 -> below 2 | Improved |
| increase ensemble size from N=3 to N=5 (ablation) | WASH · AAR z-score, Llama3.1-8B | z-score: 3.699 -> 2.232 | Improved |
| share the same coordinated watermark across all ensemble models (ablation) | WASH ensemble · z-score vs ensemble size | z-score: signal persists above z > 4; AAR 25.11, Exp-Edit 17.44 | Hurt |
| remove fluency-aware routing (naive distribution averaging) (ablation) | WASH · GSM8K | accuracy: 0.698 -> 0.339 | Hurt |
| replace routing with post-hoc rewrite by unwatermarked model (ablation) | Naive Avg · GSM8K | accuracy: 0.339 -> 0.205 | Hurt |
| replace routing with post-hoc rewrite by watermarked model (ablation) | Naive Avg · detection TPR@5%FPR | TPR@5%FPR: 41.3 -> 76.8 | Hurt |
| WASH ensemble versus watermarked baseline on generation quality | Llama3.1-8B · GSM8K | accuracy: 0.511 -> 0.695 | Improved |
| use WASH instead of De-mark/ToBlend for removal speed | GSM8K generation · relative runtime | time: 38.80x -> 2.19x (De-mark vs WASH N=3, Llama3.1-8B) | Improved |
| use WASH instead of RandomWalk final-text rewriting | WritingBench · WritingBench | score: 3.61 -> 4.26 | Improved |
| sequential specialist loading (WASH-Seq.) instead of parallel (ablation) | WASH · serving cost | peak memory / token latency: memory 40.87 GB -> 15.76 GB; latency 56.9 ms -> 165.5 ms | Mixed |
| fluency-aware routing on specialised vocabulary (medical, legal) | WASH · MMLU medical and legal subsets | detection z: z 0.84 (medical), 1.08 (legal); routed tokens 3.2%, 2.7% | 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.
- De-mark — De-mark: Watermark Removal in Large Language Models (ICML 2025) · better on 3 of 3 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
- Unbiased Watermark for Large Language Models (ICLR 2024 · similarity 0.90)
- Break the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models (ICLR 2026 · similarity 0.90)
- Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach (NeurIPS 2025 · similarity 0.90)
- Black-Box Detection of Language Model Watermarks (ICLR 2025 · similarity 0.89)
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.89)
- GaussMark: A Practical Approach for Structural Watermarking of Language Models (ICML 2025 · similarity 0.89)
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.
