AI Paper Atlas

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

ICLR 2026 · Adversarial robustness and security › Attacks on text and graphs · approach: LLM jailbreak and injection defense · 5 citations (OpenAlex)

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 changedWhereMeasuredEffect
apply A2D token-level EOS alignmentLLaDA · HarmBench average ASRattack success rate: 60.2 -> 6.8Improved
apply A2D token-level EOS alignmentDream · HarmBench DIJAattack success rate: 84.4 -> 0.0Improved
apply A2D token-level EOS alignmentLLaDA · General capabilityaccuracy: 66.6 -> 66.2No clear effect
apply A2D token-level EOS alignmentLLaDA · Math (GSM8K, GPQA)accuracy: 41.4 -> 40.6No clear effect
apply A2D token-level EOS alignmentDream · Math (GSM8K, GPQA)accuracy: 57.2 -> 55.9Hurt
use [EOS] instead of OOD token as refusal token (ablation)A2D · MMLUaccuracy: 60.5 -> 63.2Improved
use [EOS] instead of high-frequency token as refusal token (ablation)A2D · MMLUaccuracy: 55.7 -> 63.2Improved
use [EOS] instead of low-frequency token as refusal token (ablation)A2D · DIJAattack success rate: 1.9 -> 1.3Improved
use A2D instead of VRPOLLaDA-1.5 · HarmBench average ASRattack success rate: 48.6 -> 9.1Improved
use refusal-trained (RT) alignmentLLaDA · HarmBench Prefilling and DIJAattack success rate: Prefilling 69.6 -> 65.8; DIJA 82.9 -> 79.6No clear effect
use safe-only SFTDream · HarmBench average ASRattack success rate: 40.2 -> 59.2Hurt
add early rejection by [EOS] probability threshold 0.9A2D LLaDA-1.5 · AdvBench refusalsrefusal speedup: 6x speedup, 1.6% over-refusalsImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy