AI Paper Atlas

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

ICML 2026 · Large language models › Language, safety and interpretability · approach: Text watermarking · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
shrink synchronization window from w=8 to w=3 (ablation)PASA · C4 T5 replacement r=0.5TPR@1%FPR: 0.1508 -> 0.7236Improved
use very fine semantic clusters (K above 500) instead of coarse ones (ablation)PASA · C4 token replacement attacksdetection accuracyHurt
vary cluster count K within 3 to 100 (ablation)PASA · C4 T5 token replacementrobustnessNo clear effect
vary K on clean text (ablation)PASA · C4 clean textdetection accuracyNo clear effect
use PASA semantic-cluster watermark instead of KGW under DIPPER Ord=80Llama-2-13B · DIPPER (Lex=60, Ord=80)TPR@1%FPR: 0.3100 -> 0.5377Improved
use PASA instead of DAWA under T5-Large attackLlama-2-13B · C4 T5-Large token replacementTPR@1%FPR: 0.3300 -> 0.9296Improved
use PASA instead of AWTI under T5-Large attackLlama-2-13B · C4 T5-Large token replacementTPR@1%FPR: 0.9978 -> 0.9296Hurt
use PASA instead of Exp-edit under DIPPER Ord=80Llama-2-13B · DIPPER (Lex=60, Ord=80)TPR@1%FPR: 0.1150 -> 0.5377Improved
use PASA instead of SynthID-Text under OPT-2.7B paraphrasingLlama-2-13B · OPT-2.7B paraphraseTPR@1%FPR: 0.4350 -> 0.9146Improved
use PASA instead of SIR under DIPPERLlama-2-13B · DIPPER (Lex=60, Ord=80)TPR@10%FPR: 0.8809 -> 0.7593Hurt
prepend human prompt to watermarked text before detection (ablation)PASA · C4TPR@1%FPR: 1.0000 -> 0.9899Hurt
use instruction-tuned LLaMA-2-13B-chat as generator instead of basePASA · surrogate LM mismatchROC-AUC: 1.0000 -> 0.9817Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy