Research map / Large language models / Instruction and fine-tuning
When Style Breaks Safety: Defending LLMs Against Superficial Style Alignment
Yuxin Xiao, Sana Tonekaboni, Walter Gerych, Vinith Menon Suriyakumar, Marzyeh Ghassemi
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 semantically irrelevant style patterns to jailbreak queries | 36 safety-aligned LLMs · 7 jailbreak benchmarks (2,134 queries) | ASR: 32 of 36 models show higher ASR; paired t-test p = 0.0002 | Hurt |
| fine-tune on style-specific data then test on same style | Llama-3.1-8B-Instruct · matched-style jailbreaks | ASR | Hurt |
| mix more style-removed data into fine-tuning set (ablation) | Llama-3.1-8B-Instruct · matched-style jailbreaks | ASR | Improved |
| move style pattern between prefix and suffix (ablation) | Llama-3.1-8B-Instruct · style fine-tuning | ASR | No clear effect |
| SafeStyle: add 50 style-matched safety examples | Llama-3.1-8B-Instruct · poem-style fine-tuning | Delta ASR: +0.394 -> -0.015 | Improved |
| SafeStyle on Shakespeare style | Qwen2.5-3B-Instruct · Shakespeare-style fine-tuning | Delta ASR: -0.060 -> -0.505 | Improved |
| SafeStyle on list style vs original diverse safety data | Llama-3.1-8B-Instruct · list-style fine-tuning | Delta ASR: Vanilla -0.032 vs SafeStyle -0.036 | Mixed |
| SafeStyle on real-world Dolly-15K | Llama-3.1-8B-Instruct · Dolly-15K fine-tuning | Delta ASR: +0.416 -> -0.045 | Improved |
| use style-matched vs style-mismatched safety data (ablation) | Llama-3.1-8B-Instruct · list/poem fine-tuning | ASR | Improved |
| increase safety-data mixing ratio beyond 50 examples (ablation) | SafeStyle · list-style fine-tuning | ASR vs LC_WR | Mixed |
| SafeStyle vs baselines on style adaptation utility | SafeStyle · AlpacaEval styles | Delta LC_WR | No clear effect |
| PTST (safety prompt at inference) defense | Qwen2.5-3B-Instruct · List-style fine-tuning | Delta ASR: +0.269 -> +0.220 | No clear effect |
Closest papers by meaning
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.92)
- Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy (CVPR 2025 · similarity 0.91)
- Look Before You Leap: Enhance Attention and Vigilance Regarding Harmful Content with GuidelineLLM (AAAI 2025 · similarity 0.90)
- Safety Layers in Aligned Large Language Models: The Key to LLM Security (ICLR 2025 · similarity 0.90)
- STAIR: Improving Safety Alignment with Introspective Reasoning (ICML 2025 · similarity 0.90)
- Mission Impossible: A Statistical Perspective on Jailbreaking LLMs (NeurIPS 2024 · 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.
