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
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 |
|---|---|---|---|
| align middle tokens (20th-60th) with safe refusals instead of initial tokens (ablation) | LLaDA-8B-Instruct · AdvBench, TAP attack (own judge) | ASR (%): 29.6 -> 4.5 | Improved |
| add safety alignment on initial tokens only (ablation) | LLaDA-8B-Instruct · AdvBench, TAP attack (own judge) | ASR (%): 79.1 -> 29.6 | Improved |
| add MOSA alignment on Avatar attack | LLaDA-8B-Instruct · AdvBench, Avatar attack (own judge) | ASR (%): 74.5 -> 14.3 | Mixed |
| prefill at later positions (middle) instead of initial position (ablation) | LLaDA-8B-Instruct · AdvBench 100 questions, prefill attack | attack success rate | Improved |
| GCG optimization targeting middle token instead of first token (ablation) | LLaDA-8B-Instruct · AdvBench 100 questions, GCG | attack success rate: 33% -> 2% | Hurt |
| add MOSA fine-tuning | LLaDA-8B-Instruct · GSM8K | accuracy: 69.8 -> 68.3 | No clear effect |
| add MOSA fine-tuning | LLaDA-8B-Instruct · HumanEval | pass rate: 32.8 -> 30.4 | Hurt |
| add MOSA fine-tuning | LLaDA-8B-Instruct · MMLU | accuracy: 66.4 -> 65.9 | No clear effect |
| attack MOSA with middle-token adaptive TAP/EPT/AdvPrefix | MOSA · Adaptive attacks | ASR (%): 5.1 / 3.8 / 4.5 | No clear effect |
| GCG on middle token after MOSA alignment | MOSA · AdvBench GCG at 30th token | attack success rate: under 4% | No clear effect |
Closest papers by meaning
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.90)
- Safety Depth in Large Language Models: A Markov Chain Perspective (NeurIPS 2025 · similarity 0.89)
- Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models? (CVPR 2025 · similarity 0.89)
- Safety Layers in Aligned Large Language Models: The Key to LLM Security (ICLR 2025 · similarity 0.89)
- Inference-Time Vulnerability Beyond Shallow Safety: Alignment Along Generation Trajectories (NeurIPS 2026 · similarity 0.89)
- Fail-Closed Alignment for Large Language Models (NeurIPS 2026 · similarity 0.88)
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.
