Research map / Large language models / Alignment and preferences
SafeDPO: A Simple Approach to Direct Preference Optimization with Enhanced Safety
Geon-Hyeong Kim, Yu Jin Kim, Byoungjip Kim, Honglak Lee, Kyunghoon Bae, Youngsoo Jang, Moontae Lee
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 safety-aware pair transformation (swap/discard pairs) instead of plain DPO on helpfulness preferences (ablation) | DPO · PKU-SafeRLHF-30K test (model-based) | harmless ratio (%): 38.6 -> 97.24 | Improved |
| train SafeDPO instead of only filtering pairs with unsafe preferred responses (ablation) | DPO-SAFEBETTER · PKU-SafeRLHF-30K test | harmless ratio | Improved |
| add safety margin Delta (0 to 10) (ablation) | SafeDPO · PKU-SafeRLHF-30K test | harmless ratio | No clear effect |
| increase safety margin to Delta=50 (ablation) | SafeDPO · PKU-SafeRLHF-30K test | helpfulness | Hurt |
| add safety margin Delta=10 to standard DPO helpfulness training (ablation) | DPO-HELPFUL · PKU-SafeRLHF-30K test (model-based) | harmless ratio (%): 38.6 -> 51.63 | Improved |
| add safety margin Delta=10 to DPO-HARMLESS (ablation) | DPO-HARMLESS · PKU-SafeRLHF-30K test (model-based) | harmless ratio (%): 76.82 -> 93.88 | Improved |
| add safety margin Delta=10 to DPO-SAFEBETTER (ablation) | DPO-SAFEBETTER · PKU-SafeRLHF-30K test (model-based) | harmless ratio (%): 50.5 -> 62.13 | Improved |
| scale reference model from 1.5B to 13B (ablation) | SafeDPO · PKU-SafeRLHF-30K test (model-based) | harmless ratio (%): 96.38 -> 97.00 | No clear effect |
| use SafeDPO instead of SafeRLHF | SafeRLHF · Human evaluation | safety: 0.932 -> 0.943 | No clear effect |
| apply SafeDPO instead of SFT | SFT · Human evaluation | helpfulness: 0.868 -> 0.499 | Hurt |
| use SafeDPO hard-constraint training on benign-lookalike prompts | SafeDPO · XSTest | over-refusal (%): 0 -> 12.4 (DPO-HELPFUL vs SafeDPO) | Hurt |
| use SafeDPO instead of DPO-HELPFUL | DPO-HELPFUL · XSTest | harmless ratio (%): 14.5 -> 100 | Improved |
Closest papers by meaning
- Enhancing Safety in Reinforcement Learning with Human Feedback via Rectified Policy Optimization (NeurIPS 2025 · similarity 0.94)
- BSO: Safety Alignment Is Density Ratio Matching (NeurIPS 2026 · similarity 0.94)
- Safe RLHF: Safe Reinforcement Learning from Human Feedback (ICLR 2024 · similarity 0.93)
- One-Shot Safety Alignment for Large Language Models via Optimal Dualization (NeurIPS 2024 · similarity 0.92)
- AlphaDPO: Adaptive Reward Margin for Direct Preference Optimization (ICML 2025 · similarity 0.92)
- Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback (NeurIPS 2025 · similarity 0.91)
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.
