Elamur Research⌕
☰
home / papers / 2026 / LR-DWM: Efficient Watermarking for Diffusion Language Models

LR-DWM: Efficient Watermarking for Diffusion Language Models

Paper · 2026 · Ofek Raban, Gal Chechik and Ethan Fetaya

Summary

Most text watermarks for LLMs assume autoregressive decoding, where each token is biased using the tokens already produced to its left. Diffusion Language Models (DLMs) break this assumption because they fill in positions in an unpredictable order. Earlier DLM watermarks, WM-DLM and DMARK, work around this but are slow or memory-hungry.

The authors propose LR-DWM, which treats the left and right neighbours of a position as two separate watermark sources. Each available neighbour is hashed with its own secret key to define a green list, and an additive logit bias δ is applied to each list. Tokens that satisfy both constraints are therefore favoured.

Detection gives each token a ternary score and combines the scores into a Z-statistic calibrated on human-written C4 text. Experiments use LLaDA-8B-Instruct and DREAM-7B-Instruct with WaterBench prompts. LR-DWM runs with time and memory close to the unwatermarked model, and its detectability and quality trade-off is comparable to prior methods. This makes watermarking practical for diffusion-based generation.

Key points

  • Scoring and detection: each token scores m_L + m_R − 1, which is +1, 0 or −1. The sum is standardized with a variance estimated from human text. The threshold was validated on 10,000 C4 texts of 400 tokens to give 1% FPR.
  • Setup: 600 WaterBench prompts, 300-token outputs and 300 diffusion steps. LLaDA used deterministic decoding with block length 25, and DREAM used stochastic decoding. Perplexity was measured with Qwen2.5-32B, and efficiency on one NVIDIA H100.
  • Efficiency: DMARK caches |V|² lookup tables and nearly doubles peak GPU memory. WM-DLM is slower and needs non-zero temperature. LR-DWM stays near the baseline on both measures.
  • Quality on LLaDA (Table 1, as read from the extracted text): at 90% and 99.5% detection, LR-DWM perplexity is about 2.80 and 3.37. This is close to DMARK, while WM-DLM is roughly 5–6.
  • Robustness (MarkLLM): detection stays at 98.8% under 10% word deletion and 99.4% under 10% context-aware substitution. It drops sharply under paraphrasing.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Bar-Ilan UniversityRamat Gan · Israel141 works · 23,225 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksComputer Vision and Pattern Recognition · 2,869 worksHardware and Architecture · 81 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Ofek Raban, Gal Chechik and Ethan Fetaya. 2026. LR-DWM: Efficient Watermarking for Diffusion Language Models. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.2161

@inproceedings{raban2026lr,
  title = {{LR-DWM: Efficient Watermarking for Diffusion Language Models}},
  author = {Ofek Raban, Gal Chechik and Ethan Fetaya},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.2161},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.2161}
}

References

18 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Alibaba Qwen Team. 2024. Qwen2.5: Technical report. arXiv preprint arXiv:2412.15115. find
  2. Jacob Austin, Daniel Johnson, Jonathan Ho, Daniel Tarlow, and Aaron van den Oord. 2021. Structured denoising diffusion models in discrete state-spaces. In Advances in Neural Information Processing Systems.
  3. Thibaud Gloaguen, Robin Staab, Nikola Jovanović, and Martin Vechev. 2025. Watermarking diffusion language models. arXiv preprint arXiv:2509.24368. — [arXiv:2509.24368](https://arxiv.org/abs/2509.24368) find
  4. Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems. find
  5. Xinyu Hu, Yuxin Zhang, Jing Li, Zhi Wang, and Zihan Liu. 2023. Variance-reduced watermarking for large language models. In Advances in Neural Information Processing Systems. find
  6. Jaeyeon Kim, Kulin Shah, Vasilis Kontonis, Sham Kakade, and Sitan Chen. 2025. Train for the worst, plan for the best: Understanding token ordering in masked diffusions. In Proceedings of the 42nd International Conference on Machine Learning. find
  7. Colin Raffel, Noam Shazeer, Adam Roberts, and 1 others. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research.
  8. Subham Sahoo, Marianne Arriola, Yair Schiff, Aaron Gokaslan, Edgar Marroquin, Justin Chiu, Alexander Rush, and Volodymyr Kuleshov. 2024. Simple and effective masked diffusion language models. In Advances in Neural Information Processing Systems.
  9. Yang Song and Stefano Ermon. 2021. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations.
  10. Shangqing Tu, Yuliang Sun, Yushi Bai, Jifan Yu, Lei Hou, and Juanzi Li. 2024. Waterbench: Towards holistic evaluation of watermarks for large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics. find
  11. Linyu Wu, Linhao Zhong, Wenjie Qu, Yuexin Li, Yue Liu, Shengfang Zhai, Chunhua Shen, and Jiaheng Zhang. 2025. Dmark: Order-agnostic watermarking for diffusion large language models. arXiv preprint arXiv:2510.02902. — [arXiv:2510.02902](https://arxiv.org/abs/2510.02902) find
  12. Jiacheng Ye, Zhihui Xie, Lin Zheng, Jiahui Gao, Zirui Wu, Xin Jiang, Zhenguo Li, and Lingpeng Kong. 2025. Dream 7b. https://hkunlp.github.io/ blog/2025/dream. Technical report / project page. ↗
  13. Xuandong Zhao, Zhiyuan Wu, Tianyu Xie, and Yuanzhi Zhang. 2023. Provable robust watermarking for large language models. In Advances in Neural Information Processing Systems. find
  14. John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, and Tom Goldstein. 2023. A watermark for large language models. In Proceedings of the 40th International Conference on Machine Learning.
  15. Aaron Lou and Stefano Ermon. 2024. Discrete diffusion modeling by estimating the ratio of the data distribution to the noise distribution. In Proceedings of the 41st International Conference on Machine Learning. find
  16. Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. 2025. Large language diffusion models. arXiv preprint arXiv:2502.09992.
  17. Junjie Pan, Ruiyi Zhang, Zhen Chen, and Ming Li. 2024a. On the reliability of watermarks for large language models. arXiv preprint. find
  18. Leyi Pan, Aiwei Liu, Zhiwei He, Zitian Gao, Xuandong Zhao, and 1 others. 2024b. Markllm: An open-source toolkit for llm watermarking. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/raban-2026-lr-dwm-efficient-watermarking-diffusion/