AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

The Devil behind the mask: An emergent safety vulnerability of Diffusion LLMs

Zichen Wen, Jiashu Qu, Zhaorun Chen, Xiaoya Lu, Dongrui Liu, Zhiyuan Liu, Ruixi Wu, Yicun Yang, Xiangqi Jin, Haoyun Xu, Xuyang Liu, Weijia Li and 4 more

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Safety alignment of LLMs · 28 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
add interleaved mask-text prompting (DiJA) over ReNeLLMdLLM jailbreak · HarmBench Dream-InstructASR-e (%): 6.5 -> 57.5Improved
add masked regions to the prompt (ablation)DiJA · HarmBench LLaDA-8B-InstructASR-e (%): 14.0 -> 55.5Improved
add benign separators between masks (ablation)DiJA · HarmBench LLaDA-8B-InstructASR-e (%): 16.8 -> 55.5Improved
add LLM prompt refinement with few-shot demos (ablation)DiJA · HarmBench LLaDA-8B-InstructASR-e (%): 54.8 -> 55.5No clear effect
use GPT-4o instead of Qwen2.5-7B-Instruct as prompt constructorDiJA · HarmBench LLaDA-InstructASR-e (%): 55.5 -> 60.0Improved
increase number of mask spans from 1 to 3 (50 token budget) (ablation)DiJA · HarmBench LLaDA-8B-InstructASR-k (%): 62.3 -> 96.3Improved
increase number of mask spans from 3 to 10 (ablation)DiJA · HarmBench LLaDA-8B-InstructASR-k (%): 96.3 -> 87.5Hurt
apply defenses Self-reminder and RPODiJA · three dLLMs, multiple benchmarksASR-e / SRSNo clear effect
fine-tune with refusal-aware denoising alignmentLLaDA-Instruct · HarmBenchASR-e (%): 55.5 -> 32.5Improved
fine-tune with refusal-aware denoising alignmentLLaDA-Instruct · JailBreakBenchASR-e (%): 81.0 -> 25.0Improved
refusal-aware alignment effect on general capabilityLLaDA-Instruct · general benchmarks (7)accuracy: GPQA 32.4 -> 27.5; GSM8K 78.5 -> 78.9; MMLU-Pro 35.1 -> 36.1Mixed
use too few masked tokens (10) or too many (50) (ablation)DiJA · JailbreakBench and StrongREJECTASRHurt

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