AI Paper Atlas

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

Where to Start Alignment? Diffusion Large Language Model May Demand a Distinct Position

Zhixin Xie, Xurui Song, Jun Luo

AAAI 2026 · Large language models › Language, safety and interpretability · approach: Parallel and early-exit decoding · 7 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
align middle tokens (20th-60th) with safe refusals instead of initial tokens (ablation)LLaDA-8B-Instruct · AdvBench, TAP attack (own judge)ASR (%): 29.6 -> 4.5Improved
add safety alignment on initial tokens only (ablation)LLaDA-8B-Instruct · AdvBench, TAP attack (own judge)ASR (%): 79.1 -> 29.6Improved
add MOSA alignment on Avatar attackLLaDA-8B-Instruct · AdvBench, Avatar attack (own judge)ASR (%): 74.5 -> 14.3Mixed
prefill at later positions (middle) instead of initial position (ablation)LLaDA-8B-Instruct · AdvBench 100 questions, prefill attackattack success rateImproved
GCG optimization targeting middle token instead of first token (ablation)LLaDA-8B-Instruct · AdvBench 100 questions, GCGattack success rate: 33% -> 2%Hurt
add MOSA fine-tuningLLaDA-8B-Instruct · GSM8Kaccuracy: 69.8 -> 68.3No clear effect
add MOSA fine-tuningLLaDA-8B-Instruct · HumanEvalpass rate: 32.8 -> 30.4Hurt
add MOSA fine-tuningLLaDA-8B-Instruct · MMLUaccuracy: 66.4 -> 65.9No clear effect
attack MOSA with middle-token adaptive TAP/EPT/AdvPrefixMOSA · Adaptive attacksASR (%): 5.1 / 3.8 / 4.5No clear effect
GCG on middle token after MOSA alignmentMOSA · AdvBench GCG at 30th tokenattack success rate: under 4%No clear effect

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