Research map / Adversarial robustness and security / Attacks on text and graphs
Disentangling Intent from Role: Adversarial Self-Play for Persona-Invariant Safety Alignment
Jiajia Li, Xiaoyu Wen, Shuyue Hu, Qiaosheng Zhang, Zhen 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 persona-invariant consistency (PIC) loss to DPO+SFT (ablation) | PICL on Qwen2.5-7B-Instruct · OOD persona jailbreak (avg) | ASR: 0.099 -> 0.054 | Improved |
| add PIC loss to DPO+SFT (ablation) | PICL on Llama-3.1-8B-Instruct · OOD persona jailbreak (avg) | ASR: 0.072 -> 0.052 | Improved |
| add PIC loss (ablation) | PICL on Qwen2.5-7B-Instruct · Direct harmful instructions (avg) | ASR: 0.131 -> 0.115 | Improved |
| add PIC loss (ablation) | PICL on Llama-3.1-8B-Instruct · Direct harmful instructions (avg) | ASR: 0.052 -> 0.036 | Improved |
| add PIC loss (ablation) | PICL on Qwen2.5-7B-Instruct · Benign over-refusal (avg) | RtA rate: 0.114 -> 0.091 | Improved |
| add PIC loss (ablation) | PICL on Llama-3.1-8B-Instruct · Benign over-refusal (avg) | RtA rate: 0.185 -> 0.179 | Improved |
| add PIC loss (ablation) | PICL on Qwen2.5-7B-Instruct · General capability (avg) | accuracy: 0.636 -> 0.629 | No clear effect |
| add lineage-based credit propagation (gamma>0) (ablation) | PLE · Persona evolution trajectory | ASR convergence | Improved |
| add UCB-based selection instead of greedy (ablation) | PLE · Persona evolution trajectory | ASR | Improved |
| replace genetic algorithm with PLE | Persona evolution attack · WildGuardTest, Qwen2.5-7B-Instruct | ASR: 24.5% -> 31.6% | Improved |
| replace genetic algorithm with PLE | Persona evolution attack · MaliciousInstruct, Qwen2.5-7B-Instruct | ASR: 0.900 -> 0.990 | Improved |
| use PICL instead of DPO | Qwen2.5-7B-Instruct · OOD persona jailbreak (avg) | ASR: 0.447 -> 0.054 | Improved |
Closest papers by meaning
- WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Language Models (NeurIPS 2024 · similarity 0.86)
- Self-Jailbreaking: Language Models Can Reason Themselves Out of Safety Alignment After Benign Reasoning Training (ICLR 2026 · similarity 0.86)
- Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy (CVPR 2025 · similarity 0.86)
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.86)
- Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment (CVPR 2025 · similarity 0.85)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.85)
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.
