Research map / Image generation / Fast diffusion sampling
The Safety-Aware Denoiser for Text Diffusion Models
Amman Yusuf, Zhejun Jiang, Mijung Park
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 SAD guidance | MDLM · RTP malicious prompts | unsafe rate (%): 38.4 -> 32.6 | Improved |
| add SAD guidance | LLaDA-8B-Base · RTP malicious prompts | unsafe rate (%): 19.8 -> 14.8 | Improved |
| use ToxiGen instead of RTP as negation set (ablation) | SAD on MDLM · RTP malicious prompts | unsafe rate (%): 32.6 -> 32.8 | Hurt |
| use ToxiGen instead of RTP as negation set (ablation) | SAD on LLaDA · RTP malicious prompts | unsafe rate (%): 14.8 -> 17.6 | Hurt |
| use BeaverTails instead of RTP as negation set (ablation) | SAD on LLaDA · RTP malicious prompts | unsafe rate (%): 14.8 -> 16.0 | Hurt |
| add SAD guidance | MDLM · RTP benign prompts | BERTScore: 0.554 -> 0.555 | No clear effect |
| add SAD guidance (ablation) | MDLM with ToxiGen negation set · RTP benign prompts | BERTScore: 0.554 -> 0.531 | Hurt |
| add SAD guidance under PAD attack | LLaDA-1.5 · WildJailbreak | HarmBench ASR (%): 43.6 -> 29.0 | Improved |
| add SAD guidance under PAD attack | LLaDA-Instruct · WildJailbreak | HarmBench ASR (%): 45.6 -> 29.6 | Improved |
| add SAD guidance under DIJA attack | LLaDA-1.5 · WildJailbreak | HarmBench ASR (%): 64.3 -> 62.8 | No clear effect |
| add SAD guidance under DIJA attack | Dream-Instruct · WildJailbreak | HarmBench ASR (%): 61.7 -> 54.8 | Improved |
| add SAD to DiffuGuard under PAD attack | LLaDA-1.5 · WildJailbreak | HarmBench ASR (%): 45.2 -> 28.6 | Improved |
Closest papers by meaning
- Training-Free Safe Text Embedding Guidance for Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Detect-and-Guide: Self-regulation of Diffusion Models for Safe Text-to-Image Generation via Guideline Token Optimization (CVPR 2025 · similarity 0.89)
- Denoising Time Matters: Diverse Generation in Diffusion Language Models (NeurIPS 2026 · similarity 0.89)
- Training-Free Safe Denoisers for Safe Use of Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Towards Resilient Safety-driven Unlearning for Diffusion Models against Downstream Fine-tuning (NeurIPS 2025 · similarity 0.88)
- Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (ICLR 2025 · 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.
