Research map / Large language models / Language, safety and interpretability
SOM Directions Are Better than One: Multi-Directional Refusal Suppression in Language Models
Giorgio Piras, Raffaele Mura, Fabio Brau, Luca Oneto, Fabio Roli, Battista Biggio
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 |
|---|---|---|---|
| ablate multiple SOM directions instead of single difference-in-means direction | refusal ablation, Llama-3-8B-Instruct · HarmBench | ASR (%): 15.09 -> 88.05 | Improved |
| ablate multiple SOM directions instead of single direction | Llama-2-7B-chat · HarmBench | ASR (%): 0.0 -> 59.11 | Improved |
| ablate multiple SOM directions instead of single direction | Qwen2.5-3B-Instruct · HarmBench | ASR (%): 88.05 -> 93.71 | Improved |
| ablate multiple SOM directions instead of RDO direction | Gemma2-9B-it · HarmBench | ASR (%): 91.82 -> 96.27 | Improved |
| use MD refusal ablation instead of SAA jailbreak | Llama-3-8B-Instruct · HarmBench | ASR (%): 91.20 -> 88.05 | Hurt |
| use MD refusal ablation instead of SAA jailbreak | Mistral-7B-RR · HarmBench | ASR (%): 1.6 -> 25.79 | Improved |
| increase number of ablated directions from 2 to 7 (ablation) | MD on Llama-2-7B-chat · HarmBench | ASR (%): 7.5 -> 59.11 | Improved |
| increase number of ablated directions from 2 to 7 (ablation) | MD on Qwen-14B-Chat · HarmBench | ASR (%): 75.47 -> 91.82 | Improved |
| increase number of ablated directions from 5 to 7 (ablation) | MD on Mistral-7B-RR · HarmBench | ASR (%): 25.79 -> 18.25 | Hurt |
| increase number of ablated directions from 6 to 7 (ablation) | MD on Qwen-7B-Chat · HarmBench | ASR (%): 87.42 -> 86.80 | No clear effect |
| progressively ablate MD directions (ablation) | Llama-2-7B-chat harmful representations · PCA of internals | intra-cluster variance of harmful prompts: 5.85 -> 1.25 | Improved |
| progressively ablate MD directions (ablation) | Llama-2-7B-chat · PCA of internals | distance between harmful and harmless centroids: 8.82 -> 2.18 | Improved |
Closest papers by meaning
- The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independence (ICML 2025 · similarity 0.91)
- Latent-space Attacks for Refusal Evasion in Language Models (NeurIPS 2026 · similarity 0.91)
- LLMs Encode Harmfulness and Refusal Separately (NeurIPS 2025 · similarity 0.90)
- Refusal in Language Models Is Mediated by a Single Direction (NeurIPS 2024 · similarity 0.89)
- Refusal Direction is Universal Across Safety-Aligned Languages (NeurIPS 2025 · similarity 0.89)
- A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models (NeurIPS 2026 · similarity 0.87)
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.
