Research map / Adversarial robustness and security / Attacks on text and graphs
A2D: Any-Order, Any-Step Safety Alignment for Diffusion Language Models
Wonje Jeung, Sangyeon Yoon, Yoonjun Cho, Dongjae Jeon, Sangwoo Shin, Hyesoo 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 |
|---|---|---|---|
| apply A2D token-level EOS alignment | LLaDA · HarmBench average ASR | attack success rate: 60.2 -> 6.8 | Improved |
| apply A2D token-level EOS alignment | Dream · HarmBench DIJA | attack success rate: 84.4 -> 0.0 | Improved |
| apply A2D token-level EOS alignment | LLaDA · General capability | accuracy: 66.6 -> 66.2 | No clear effect |
| apply A2D token-level EOS alignment | LLaDA · Math (GSM8K, GPQA) | accuracy: 41.4 -> 40.6 | No clear effect |
| apply A2D token-level EOS alignment | Dream · Math (GSM8K, GPQA) | accuracy: 57.2 -> 55.9 | Hurt |
| use [EOS] instead of OOD token as refusal token (ablation) | A2D · MMLU | accuracy: 60.5 -> 63.2 | Improved |
| use [EOS] instead of high-frequency token as refusal token (ablation) | A2D · MMLU | accuracy: 55.7 -> 63.2 | Improved |
| use [EOS] instead of low-frequency token as refusal token (ablation) | A2D · DIJA | attack success rate: 1.9 -> 1.3 | Improved |
| use A2D instead of VRPO | LLaDA-1.5 · HarmBench average ASR | attack success rate: 48.6 -> 9.1 | Improved |
| use refusal-trained (RT) alignment | LLaDA · HarmBench Prefilling and DIJA | attack success rate: Prefilling 69.6 -> 65.8; DIJA 82.9 -> 79.6 | No clear effect |
| use safe-only SFT | Dream · HarmBench average ASR | attack success rate: 40.2 -> 59.2 | Hurt |
| add early rejection by [EOS] probability threshold 0.9 | A2D LLaDA-1.5 · AdvBench refusals | refusal speedup: 6x speedup, 1.6% over-refusals | Improved |
Closest papers by meaning
- MaskForge: Structure-Aware Adaptive Attacks for Jailbreaking Diffusion Large Language Models (NeurIPS 2026 · similarity 0.88)
- Fail-Closed Alignment for Large Language Models (NeurIPS 2026 · similarity 0.88)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.88)
- Inference-Time Vulnerability Beyond Shallow Safety: Alignment Along Generation Trajectories (NeurIPS 2026 · similarity 0.88)
- Safety Layers in Aligned Large Language Models: The Key to LLM Security (ICLR 2025 · similarity 0.88)
- Why Jailbreaks Succeed in Diffusion Language Models: An Energy Landscape Analysis (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.
