AI Paper Atlas

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

AAAI 2026 · Large language models › Language, safety and interpretability · approach: Safety alignment of LLMs · 10 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
ablate multiple SOM directions instead of single difference-in-means directionrefusal ablation, Llama-3-8B-Instruct · HarmBenchASR (%): 15.09 -> 88.05Improved
ablate multiple SOM directions instead of single directionLlama-2-7B-chat · HarmBenchASR (%): 0.0 -> 59.11Improved
ablate multiple SOM directions instead of single directionQwen2.5-3B-Instruct · HarmBenchASR (%): 88.05 -> 93.71Improved
ablate multiple SOM directions instead of RDO directionGemma2-9B-it · HarmBenchASR (%): 91.82 -> 96.27Improved
use MD refusal ablation instead of SAA jailbreakLlama-3-8B-Instruct · HarmBenchASR (%): 91.20 -> 88.05Hurt
use MD refusal ablation instead of SAA jailbreakMistral-7B-RR · HarmBenchASR (%): 1.6 -> 25.79Improved
increase number of ablated directions from 2 to 7 (ablation)MD on Llama-2-7B-chat · HarmBenchASR (%): 7.5 -> 59.11Improved
increase number of ablated directions from 2 to 7 (ablation)MD on Qwen-14B-Chat · HarmBenchASR (%): 75.47 -> 91.82Improved
increase number of ablated directions from 5 to 7 (ablation)MD on Mistral-7B-RR · HarmBenchASR (%): 25.79 -> 18.25Hurt
increase number of ablated directions from 6 to 7 (ablation)MD on Qwen-7B-Chat · HarmBenchASR (%): 87.42 -> 86.80No clear effect
progressively ablate MD directions (ablation)Llama-2-7B-chat harmful representations · PCA of internalsintra-cluster variance of harmful prompts: 5.85 -> 1.25Improved
progressively ablate MD directions (ablation)Llama-2-7B-chat · PCA of internalsdistance between harmful and harmless centroids: 8.82 -> 2.18Improved

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