ai-kunkun

ai-kunkun / PASA

Public

[ICML 2026] PASA: A Principled Embedding-Space Watermarking Approach for LLM-Generated Text under Semantic-Invariant Attacks

ai-kunkun.github.ioPASA_page
16
0
100% credibility
Found May 13, 2026 at 16 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

This project provides code to watermark AI-generated text by embedding signals in the semantic space, enabling robust detection even after meaning-preserving modifications.

How It Works

1
🔍 Discover PASA

You hear about this clever tool that adds invisible signatures to computer-written stories so you can spot them later, even if tweaked.

2
📥 Get the files

You download the simple folder of tools to your computer from the project page.

3
🛠️ Ready your computer

You set up the basic helpers needed to run the tool, like preparing a workspace.

4
🤖 Add AI thinkers and texts

You tell the tool where to find the smart AI brains and collections of real writing samples.

5
✨ Create marked stories

You press go, and it writes new stories—some plain, some with secret marks—and checks if it can find the marks.

6
📊 Check the scores

You look at the easy report showing high detection scores for marked stories versus real human ones.

✅ Marks work perfectly

You now have proof that your AI stories carry detectable signatures, safe from sneaky changes.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 16 to 16 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is PASA?

PASA watermarks LLM-generated text by tweaking token choices in embedding space, making it detectable even after paraphrasing or editing that preserves meaning. You feed it a prompt via Python API or shell script, pick models like Llama via Hugging Face, and get watermarked output alongside detection scores—handles generation, watermarking, and verification in one go. Built in Python with PyTorch and Transformers, it reproduces ICML 2026 paper results on C4 datasets out of the box.

Why is it gaining traction?

Unlike token-based watermarks that break under synonyms or rewrites, PASA uses semantic clusters for robustness, backed by arXiv paper and project page—perfect for ICML 2026 papers GitHub hunters. Devs dig the dead-simple API for watermark.generate_watermarked(prompt) and instant detection scores, plus experiment.sh for batch runs on local data. It's popping in ICML 2026 Reddit threads and anonymous GitHub repos chasing deadlines.

Who should use this?

LLM researchers testing watermark resilience against attacks, AI ethics teams tracking generated content provenance, or security devs building detectors for enterprise chatbots. Ideal for reproducing ICML 2025/2026 experiments or prototyping in workshops without custom clustering.

Verdict

Early days with 16 stars and 1.0% credibility score—docs shine with repro scripts and demo code, but needs more tests and broader model support. Grab it for academic watermarking plays or ICML 2026 rebuttal prep; skip for production until maturity grows.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.