Research map / Image generation / Fast diffusion sampling
Toward Safer Diffusion Language Models: Discovery and Mitigation of Priming Vulnerability
Shojiro Yamabe, Jun Sakuma
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 Recovery Alignment training from contaminated intermediate states | LLaDA Instruct · anchoring attack t_inter=16 | ASR (lower better): 88.7 -> 8.3 | Improved |
| add Recovery Alignment | LLaDA Instruct · First-Step GCG | ASR (lower better): 58.0 -> 11.3 | Improved |
| add Recovery Alignment | MMaDA MixCoT · anchoring attack t_inter=16 | ASR (lower better): 98.3 -> 34.3 | Improved |
| train from contaminated states instead of only fully masked sequences (ablation) | RA · LLaDA anchoring t_inter=16 | ASR (lower better): 76.7 -> 8.3 | Improved |
| add Recovery Alignment | LLaDA Instruct · anchoring attack t_inter=32 | ASR (lower better): 96.7 -> 50.7 | Mixed |
| add Recovery Alignment | LLaDA Instruct · PAIR | ASR (lower better): 44.3 -> 10.0 | Improved |
| add Recovery Alignment | MMaDA MixCoT · ReNeLLM | ASR (lower better): 79.3 -> 81.7 | No clear effect |
| add Recovery Alignment | LLaDA Instruct · 11 general benchmarks | average accuracy: 52.2 -> 52.6 | No clear effect |
| add Recovery Alignment | LLaDA Instruct · PIQA | accuracy: 74.4 -> 71.6 | Hurt |
| add Recovery Alignment | MMaDA MixCoT · 11 general benchmarks | average accuracy: 33.2 -> 35.0 | Improved |
| use First-Step GCG instead of Monte Carlo GCG (ablation) | GCG attack · LLaDA Instruct JBB-Behaviors | ASR: 20.0 -> 58.0 | Improved |
| use linear schedule of intervention step instead of const or uniform (ablation) | RA · anchoring attack t_inter=16 | ASR (lower better) | Improved |
Closest papers by meaning
- Why Jailbreaks Succeed in Diffusion Language Models: An Energy Landscape Analysis (NeurIPS 2026 · similarity 0.91)
- MaskForge: Structure-Aware Adaptive Attacks for Jailbreaking Diffusion Large Language Models (NeurIPS 2026 · similarity 0.89)
- Re-evaluating Confidence Remasking in Masked Diffusion Language Models (NeurIPS 2026 · similarity 0.87)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.87)
- Safety Layers in Aligned Large Language Models: The Key to LLM Security (ICLR 2025 · similarity 0.87)
- CORE: Context-Robust Remasking for Diffusion Language Models (ICML 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.
