Research map / Reinforcement learning / RL from feedback
Mitigating the Safety Alignment Tax with Null-Space Constrained Policy Optimization
Yifan Niu, Han Xiao, Dongyi Liu, Nuo Chen, Jia Li
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 NSPO safety alignment (null-space projected RL) | Llama3-8B-Instruct · HarmBench | ASR (%): 7.50 -> 0.18 | Improved |
| apply NSPO safety alignment | Llama3-8B-Instruct · SORRY-Bench | ASR (%): 34.15 -> 16.81 | Improved |
| apply NSPO safety alignment | Qwen2.5-7B-Instruct · PKU-SafeRLHF | ASR (%): 2.95 -> 0.52 | Improved |
| apply NSPO safety alignment | Llama3-8B-Instruct · MMLU / GSM8K / MATH | accuracy (%): GSM8K 75.44 -> 76.42; MATH 27.55 -> 27.10; MMLU 63.79 -> 63.64 | No clear effect |
| apply NSPO safety alignment | Qwen2.5-7B-Instruct · MMLU / GSM8K / MATH | accuracy (%): GSM8K 81.65 -> 81.96; MATH 74.80 -> 75.05 | No clear effect |
| apply NSPO safety alignment | Llama3-8B-Instruct · AlpacaEval | win rate (%): 96.15 -> 93.23 | Hurt |
| use NSPO instead of W-DOOR | Qwen2.5-7B-Instruct · GSM8K | accuracy (%): 57.85 -> 81.96 | Improved |
| add null-space projection and remove KL penalty (NSPO) versus GRPO with KL (ablation) | GRPO · safety and general benchmarks | ASR and accuracy | Improved |
| remove KL penalty from GRPO without projection (ablation) | GRPO w/o KL · general capability benchmarks | accuracy | Hurt |
| use mixed-domain data for projection matrix instead of single domain (ablation) | NSPO · safety and utility averages | utility | Improved |
| increase projection-matrix sample size from 500 to 2000 (ablation) | NSPO · safety and utility averages | safety / utility | Mixed |
| train on 100% instead of 40% of PKU-SafeRLHF (ablation) | NSPO · safety performance | performance | No clear effect |
Closest papers by meaning
- BSO: Safety Alignment Is Density Ratio Matching (NeurIPS 2026 · similarity 0.93)
- Enhancing Safety in Reinforcement Learning with Human Feedback via Rectified Policy Optimization (NeurIPS 2025 · similarity 0.93)
- AlphaAlign: Incentivizing Safety Alignment with Extremely Simplified Reinforcement Learning (ICLR 2026 · similarity 0.92)
- One-Shot Safety Alignment for Large Language Models via Optimal Dualization (NeurIPS 2024 · similarity 0.92)
- Stepwise Alignment for Constrained Language Model Policy Optimization (NeurIPS 2024 · similarity 0.90)
- Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements (ICLR 2025 · similarity 0.90)
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.
