Research map / Image generation / Fast diffusion sampling
dgMARK: Decoding-Guided Watermarking for Diffusion Language Models
Pyo Min Hong, Albert No
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 |
|---|---|---|---|
| add 3-beam one-step lookahead search (ablation) | dgMARK · C4 (greedy, LLaDA 1.5) | TPR: 0.540 -> 0.963 | Improved |
| add 3-beam one-step lookahead search (ablation) | dgMARK · C4 (multinomial, LLaDA 1.5) | TPR: 0.929 -> 1.000 | Improved |
| add 3-beam search (ablation) | dgMARK · C4 (greedy, LLaDA 1.5) | perplexity: 4.44 -> 4.75 | Hurt |
| increase beam size to 8 (ablation) | dgMARK · C4 (greedy, LLaDA 1.5) | TPR: 0.963 -> 0.991 | Improved |
| use dgMARK instead of KGW (delta=3) | watermarking · C4 (multinomial, LLaDA 1.5) | perplexity: 7.87 -> 5.27 | Improved |
| use dgMARK instead of KGW | watermarking · GSM8K (multinomial, LLaDA) | accuracy: 0.464 -> 0.735 | Improved |
| use dgMARK instead of PATTERN-MARK | watermarking · HumanEval (multinomial, LLaDA) | Pass@1: 0.073 -> 0.226 | Improved |
| add 3-beam search to dgMARK (ablation) | dgMARK · GSM8K (multinomial, LLaDA) | accuracy: 0.735 -> 0.678 | Hurt |
| increase block length from 8 to 32 (ablation) | dgMARK · C4 (multinomial) | TPR@0.01% FPR: 80.56 -> 95.29 | Improved |
| increase block length from 8 to 32 (ablation) | dgMARK · C4 (greedy) | TPR@0.01% FPR: 29.82 -> 60.96 | Improved |
| increase block length from 8 to 32 (ablation) | dgMARK · C4 (multinomial) | perplexity: 4.98 -> 5.27 | Hurt |
| use mod 3 or mod 4 hashing instead of parity (mod 2) (ablation) | dgMARK · C4 | detectability | Improved |
Closest papers by meaning
- Break the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models (ICLR 2026 · similarity 0.89)
- SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · similarity 0.88)
- SAEMark: Steering Personalized Multilingual LLM Watermarks with Sparse Autoencoders (NeurIPS 2025 · similarity 0.88)
- Black-Box Detection of Language Model Watermarks (ICLR 2025 · similarity 0.87)
- Inevitable Trade-off between Watermark Strength and Speculative Sampling Efficiency for Language Models (NeurIPS 2024 · similarity 0.87)
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.87)
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.
