Research map / Large language models / Language, safety and interpretability
PASA: A Principled Embedding-Space Watermarking Approach for LLM-Generated Text under Semantic-Invariant Attacks
Zhenxin Ai, Haiyun He
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 |
|---|---|---|---|
| shrink synchronization window from w=8 to w=3 (ablation) | PASA · C4 T5 replacement r=0.5 | TPR@1%FPR: 0.1508 -> 0.7236 | Improved |
| use very fine semantic clusters (K above 500) instead of coarse ones (ablation) | PASA · C4 token replacement attacks | detection accuracy | Hurt |
| vary cluster count K within 3 to 100 (ablation) | PASA · C4 T5 token replacement | robustness | No clear effect |
| vary K on clean text (ablation) | PASA · C4 clean text | detection accuracy | No clear effect |
| use PASA semantic-cluster watermark instead of KGW under DIPPER Ord=80 | Llama-2-13B · DIPPER (Lex=60, Ord=80) | TPR@1%FPR: 0.3100 -> 0.5377 | Improved |
| use PASA instead of DAWA under T5-Large attack | Llama-2-13B · C4 T5-Large token replacement | TPR@1%FPR: 0.3300 -> 0.9296 | Improved |
| use PASA instead of AWTI under T5-Large attack | Llama-2-13B · C4 T5-Large token replacement | TPR@1%FPR: 0.9978 -> 0.9296 | Hurt |
| use PASA instead of Exp-edit under DIPPER Ord=80 | Llama-2-13B · DIPPER (Lex=60, Ord=80) | TPR@1%FPR: 0.1150 -> 0.5377 | Improved |
| use PASA instead of SynthID-Text under OPT-2.7B paraphrasing | Llama-2-13B · OPT-2.7B paraphrase | TPR@1%FPR: 0.4350 -> 0.9146 | Improved |
| use PASA instead of SIR under DIPPER | Llama-2-13B · DIPPER (Lex=60, Ord=80) | TPR@10%FPR: 0.8809 -> 0.7593 | Hurt |
| prepend human prompt to watermarked text before detection (ablation) | PASA · C4 | TPR@1%FPR: 1.0000 -> 0.9899 | Hurt |
| use instruction-tuned LLaMA-2-13B-chat as generator instead of base | PASA · surrogate LM mismatch | ROC-AUC: 1.0000 -> 0.9817 | Hurt |
Closest papers by meaning
- A Semantic Invariant Robust Watermark for Large Language Models (ICLR 2024 · similarity 0.91)
- Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach (NeurIPS 2025 · similarity 0.88)
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.88)
- Black-Box Detection of Language Model Watermarks (ICLR 2025 · similarity 0.88)
- Robust Multi-bit Text Watermark with LLM-based Paraphrasers (ICML 2025 · similarity 0.88)
- Adaptive Text Watermark for Large Language Models (ICML 2024 · similarity 0.88)
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.
