Research map / Large language models / Language, safety and interpretability
Differentiated Directional Intervention: A Framework for Evading LLM Safety Alignment
Peng Zhang, Peijie Sun
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 DBDI intervention | Llama-2-7B · AdvBench (official template) | ASR (%): 0 -> 95.96 | Improved |
| intervene on both harm and refusal directions instead of harm only (ablation) | DBDI · AdvBench | ASR (%): 20.00 -> 97.88 | Improved |
| intervene on both directions instead of refusal only (ablation) | DBDI · AdvBench | ASR (%): 2.11 -> 97.88 | Improved |
| use differentiated (projection + steering) instead of symmetric projection (ablation) | DBDI · StrongREJECT | mean harmfulness score: 0.045 -> 0.784 | Improved |
| use differentiated instead of symmetric steering (ablation) | DBDI · StrongREJECT | mean harmfulness score: 0.004 -> 0.784 | Improved |
| add classifier-guided sparsification (keep 25-50% of neurons) (ablation) | DBDI · AdvBench | ASR (%): 90.96 -> 97.88 | Improved |
| use 100 instead of 10 calibration prompt pairs (ablation) | DBDI · AdvBench (simplified template) | ASR (%): 94.23 -> 97.88 | No clear effect |
| apply projection nullification before steering rather than reversed (ablation) | DBDI · AdvBench | ASR (%): 2.11 -> 97.88 | Improved |
| select critical layer l*=16 by linear separability (ablation) | DBDI · AdvBench | ASR (%) | Improved |
| DBDI vs TwinBreak | DBDI · HarmBench | ASR (%): 94.00 -> 91 | Hurt |
| DBDI vs TwinBreak | DBDI · AdvBench | ASR (%): 94.62 -> 95.46 | Improved |
| DBDI vs Directional Ablation (ORTHO) | Llama-2-7B · HarmBench | ASR (%): 22.6 -> 91.8 | Improved |
Closest papers by meaning
- The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Analysis of Orthogonal Safety Directions (ICML 2025 · similarity 0.92)
- LLMs Encode Harmfulness and Refusal Separately (NeurIPS 2025 · similarity 0.91)
- Improving LLM Safety Alignment with Dual-Objective Optimization (ICML 2025 · similarity 0.91)
- Fail-Closed Alignment for Large Language Models (NeurIPS 2026 · similarity 0.90)
- Refusal Direction is Universal Across Safety-Aligned Languages (NeurIPS 2025 · similarity 0.90)
- Expected Harm: Rethinking Safety Evaluation of (Mis)Aligned LLMs (NeurIPS 2026 · similarity 0.90)
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.
