Watermark Edit Detection

Modern watermark detectors typically answer whether a document is watermarked. This project asks a harder integrity question: has a watermarked output been edited after generation?

We encode structured redundancy into LLM outputs using error-correcting codes. The resulting verifier can identify inconsistencies introduced by post-generation insertions, deletions, and substitutions while retaining the underlying watermark signal.

Accepted to EMNLP 2026 Findings
Covers method design, implementation, experiments, and evaluation
Read the paper in the ACL Anthology

Error-correcting-code watermark generation with synchronization strings

The watermarking key generates both a VT-code tag and a synchronization string, which jointly structure token selection for edit detection.

(Deng et al., 2026)

References

2026

Detecting Post-generation Edits to Watermarked LLM Outputs via Error-Correcting Codes

Zewei Deng, Muhammad Siddeek, Liyan Xie, and 4 more authors

In Findings of the Association for Computational Linguistics: EMNLP 2026, Oct 2026

Bib HTML PDF

@inproceedings{deng2026detecting,
  title = {Detecting Post-generation Edits to Watermarked {LLM} Outputs via Error-Correcting Codes},
  author = {Deng, Zewei and Siddeek, Muhammad and Xie, Liyan and Seif, Mohamed and Wang, Mengdi and Poor, H. Vincent and Goldsmith, Andrea},
  booktitle = {Findings of the Association for Computational Linguistics: {EMNLP} 2026},
  month = oct,
  year = {2026},
  address = {Budapest, Hungary},
  publisher = {Association for Computational Linguistics},
  pages = {45788--45796},
}