Research map / Image generation / Fast diffusion sampling
Watermarking Diffusion Language Models
Thibaud Gloaguen, Robin Staab, Nikola Jovanović, Martin Vechev
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 |
|---|---|---|---|
| use the DLM-specific expected-context watermark instead of the naive ARLM baseline (C={-1}, delta=4) | Llada-8B watermark · Detection | TPR@1: 0.63 -> 0.99 | Improved |
| use the DLM-specific watermark instead of baseline (C={-1}, delta=4) | Dream-7B watermark · Detection | TPR@1: 0.49 -> 0.99 | Improved |
| use the DLM-specific watermark instead of baseline (C={-1,1}, delta=5) | Dream-7B watermark · Detection | TPR@1: 0.74 -> 0.99 | Improved |
| use the DLM-specific watermark instead of baseline on benchmark accuracy | Dream-7B watermark · Benchmark average | accuracy: 35.8 -> 50.1 | Improved |
| use the DLM-specific watermark instead of baseline on benchmark accuracy (C={-1,1}, delta=5) | Llada-8B watermark · Benchmark average | accuracy: 55.7 -> 53.7 | Hurt |
| apply watermark (ours) vs unwatermarked on log PPL | Llada-8B · Text quality | log PPL: 1.56 -> 1.90 | Hurt |
| ours vs baseline on log PPL | Llada-8B · Text quality | log PPL: 1.93 -> 1.90 | No clear effect |
| enlarge context to two previous tokens C={-2,-1} vs Unigram hashing | watermark · Type-1 error | max FPR@1%: 17.0% -> 1.1% | Improved |
| use ours C={-2,-1} vs C={-1} on key-to-key std of FPR (ablation) | watermark · Type-1 error | std FPR@1%: 0.3% -> 0.1% | Improved |
| apply to DreamOn-v0-7B infilling model | watermark · Infilling | TPR@1 | Improved |
| use fewer diffusion steps (ablation) | watermark · Diffusion steps ablation | TPR | Improved |
| raise top-k from 5 to 100 in energy function (ablation) | watermark · Top-k ablation | TPR@1 | Improved |
Closest papers by meaning
- Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach (NeurIPS 2025 · similarity 0.90)
- Black-Box Detection of Language Model Watermarks (ICLR 2025 · similarity 0.90)
- Unbiased Watermark for Large Language Models (ICLR 2024 · similarity 0.89)
- Break the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models (ICLR 2026 · similarity 0.89)
- A Resilient and Accessible Distribution-Preserving Watermark for Large Language Models (ICML 2024 · similarity 0.89)
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.89)
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.
