Research map / Large language models / Instruction and fine-tuning
Multilingual Safety Alignment Via Sparse Weight Editing
Jiaming Liang, Zhaoxin Wang, Handing Wang
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 sparse weight edit | Llama-3.2-1B · Multi-StrongREJECT Bengali | unsafe responses of 313: 179 -> 144 | Improved |
| apply sparse weight edit | Qwen2-0.5B · Multi-StrongREJECT English | unsafe responses of 313: 224 -> 176 | Improved |
| apply sparse weight edit | Qwen2.5-7B · Multi-StrongREJECT Bengali | unsafe responses of 313: 98 -> 60 | Improved |
| combine sparse weight edit with MPO | MPO on Qwen2.5-7B · Multi-StrongREJECT Bengali | unsafe responses of 313: 25 -> 11 | Improved |
| combine sparse weight edit with MPO | MPO on Qwen2-0.5B · Multi-StrongREJECT English | unsafe responses of 313: 108 -> 56 | Improved |
| apply sparse weight edit | Llama-3.2-1B Hebrew · Multi-StrongREJECT Hebrew | unsafe responses of 313: 109 -> 115 | Hurt |
| apply sparse weight edit | Qwen2.5-3B · MGSM | MGSM accuracy: 31.02 -> 30.91 | No clear effect |
| apply sparse weight edit | Qwen2.5-7B · MGSM | MGSM accuracy: 32.00 -> 31.56 | No clear effect |
| apply sparse weight edit | Qwen2-0.5B · MGSM | MGSM accuracy: 7.75 -> 5.27 | Hurt |
| use probe-based neuron selection instead of proposed extraction (ablation) | Sparse Weight Editing Llama-3.2-1B · Multi-StrongREJECT avg | ASR (%): 17.53 -> 14.93 | Improved |
| add UtilityAnchor to Regular anchor (ablation) | Sparse Weight Editing Llama-3.2-1B · Multi-StrongREJECT | ASR (%): 68.57 -> 17.53 | Improved |
| add UtilityAnchor to Regular anchor (ablation) | Sparse Weight Editing Llama-3.2-1B · MGSM | MGSM accuracy: 11.02 -> 18.36 | Improved |
Closest papers by meaning
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.91)
- One-Shot Safety Alignment for Large Language Models via Optimal Dualization (NeurIPS 2024 · similarity 0.91)
- Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements (ICLR 2025 · similarity 0.91)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.90)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.90)
- Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback (NeurIPS 2025 · similarity 0.89)
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.
