Research map / Large language models / Alignment and preferences
Towards Context-Invariant Safety Alignment for Large Language Models
Yixu Wang, Yang Yao, Xin Wang, Yifeng Gao, Yan Teng, Xingjun Ma, Yingchun 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 |
|---|---|---|---|
| add AIR to GRPO | GRPO, Qwen-2.5-14B · Average OOD (3 domains) | group accuracy: 23.05% -> 56.54% | Improved |
| add AIR to GSPO | GSPO, Qwen-2.5-14B · Average OOD (3 domains) | group accuracy: 30.22% -> 54.77% | Improved |
| add AIR to GRPO | GRPO, Qwen-2.5-14B · Safety OOD | group accuracy: 13.73% -> 60.78% | Improved |
| add AIR to GRPO | GRPO, Qwen-2.5-14B · Moral OOD | group accuracy: 14.71% -> 47.79% | Improved |
| add symmetric V-REx penalty to GRPO (ablation) | GRPO, Qwen-2.5-14B · Safety ID | group accuracy: 71.15% -> 35.40% | Hurt |
| add symmetric V-REx penalty to GRPO (ablation) | GRPO, Qwen-2.5-14B · Moral ID | group accuracy: 34.31% -> 7.15% | Hurt |
| add symmetric V-REx penalty to GRPO (ablation) | GRPO, Qwen-2.5-14B · Math ID | group accuracy: 64.60% -> 60.71% | Mixed |
| add AIR to GRPO | GRPO, Qwen-2.5-14B · Math ID | group accuracy: 64.60% -> 63.72% | No clear effect |
| add AIR to GSPO | GSPO, Qwen-2.5-14B · Math OOD | group accuracy: 54.87% -> 51.33% | Hurt |
| use AIR coefficient lambda moderate (about 8e-4 to 1e-3) instead of 0 or very large (ablation) | GRPO+AIR · Mixed ID and OOD average | Avg Acc and Avg group accuracy | Mixed |
| add AIR to GRPO | GRPO, Qwen-2.5-14B · Latent dispersion | L2 distance to intent centroid: 86.47 -> 71.54 | Improved |
| add AIR when open-ended reward is a gameable regex proxy | GRPO, Qwen-2.5-14B · Safety, regex-reward stress test | oracle judge safety | Improved |
Closest papers by meaning
- Safety Alignment Can Be Not Superficial With Explicit Safety Signals (ICML 2025 · similarity 0.92)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.91)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.90)
- AlphaAlign: Incentivizing Safety Alignment with Extremely Simplified Reinforcement Learning (ICLR 2026 · similarity 0.90)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 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.
