Research map / Large language models / Alignment and preferences
Revisiting Robustness for LLM Safety Alignment via Selective Geometry Control
Yonghui Yang, Wenjian Tao, Jilong Liu, Xingyu Zhu, Junfeng Fang, Huang Weibiao, Le Wu, Richang Hong, Tat-Seng Chua
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 |
|---|---|---|---|
| add token-level ShaPO selective geometry control to DPO | LLaMA-3-8B DPO · PKU-SafeRLHF-30K IID | win rate: 69.19% -> 85.45% | Improved |
| add selective geometry control vs Dr.DPO | LLaMA-3-8B · OOD safety average | attack success rate: 2.09% -> 1.37% (ShaPO-T) | Improved |
| switch to reward-level ShaPO-R from token-level ShaPO-T | LLaMA-3-8B · OOD safety average | attack success rate: 1.37% -> 0.70% | Improved |
| switch to reward-level ShaPO-R from token-level ShaPO-T | Qwen2.5-7B · OOD safety average | attack success rate: 0.42% -> 0.55% | Hurt |
| random neuron perturbation instead of no control (ablation) | DPO LLaMA-3-8B · OOD average MD | attack success rate: 28.78% -> 28.56% | No clear effect |
| uniform geometry control instead of no control (ablation) | DPO LLaMA-3-8B · OOD average MD | attack success rate: 28.78% -> 26.83% | Improved |
| selective instead of uniform geometry control (ablation) | DPO LLaMA-3-8B · OOD average MD | attack success rate: 26.83% -> 26.31% | Improved |
| selective instead of no control, IID evaluation (ablation) | DPO LLaMA-3-8B · PKU-SafeRLHF-30K IID MD | attack success rate: 26.57% -> 27.32% | No clear effect |
| increase selective neuron ratio from 0.01% to 1% (ablation) | ShaPO LLaMA-3-8B · OOD average ASR | attack success rate | Improved |
| increase selective neuron ratio beyond 1% (ablation) | ShaPO LLaMA-3-8B · OOD average ASR | attack success rate | Hurt |
| add ShaPO-R geometry control on top of reward supervision (DPO) (ablation) | DPO+Reward LLaMA-3-8B · OOD average MD | attack success rate: 17.33% -> 15.98% | Improved |
| add ShaPO-R geometry control on top of reward supervision (Dr.DPO) (ablation) | Dr.DPO+Reward LLaMA-3-8B · OOD average MD | attack success rate: 2.25% -> 1.30% | Improved |
Closest papers by meaning
- LLM Safety Alignment is Divergence Estimation in Disguise (NeurIPS 2025 · similarity 0.89)
- Improving LLM Safety Alignment with Dual-Objective Optimization (ICML 2025 · similarity 0.89)
- Robust LLM Alignment via Distributionally Robust Direct Preference Optimization (NeurIPS 2025 · similarity 0.89)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.89)
- BSO: Safety Alignment Is Density Ratio Matching (NeurIPS 2026 · similarity 0.88)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · 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.
