AI Paper Atlas

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

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Safety alignment after fine-tuning

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 changedWhereMeasuredEffect
add semantically irrelevant style patterns to jailbreak queries36 safety-aligned LLMs · 7 jailbreak benchmarks (2,134 queries)ASR: 32 of 36 models show higher ASR; paired t-test p = 0.0002Hurt
fine-tune on style-specific data then test on same styleLlama-3.1-8B-Instruct · matched-style jailbreaksASRHurt
mix more style-removed data into fine-tuning set (ablation)Llama-3.1-8B-Instruct · matched-style jailbreaksASRImproved
move style pattern between prefix and suffix (ablation)Llama-3.1-8B-Instruct · style fine-tuningASRNo clear effect
SafeStyle: add 50 style-matched safety examplesLlama-3.1-8B-Instruct · poem-style fine-tuningDelta ASR: +0.394 -> -0.015Improved
SafeStyle on Shakespeare styleQwen2.5-3B-Instruct · Shakespeare-style fine-tuningDelta ASR: -0.060 -> -0.505Improved
SafeStyle on list style vs original diverse safety dataLlama-3.1-8B-Instruct · list-style fine-tuningDelta ASR: Vanilla -0.032 vs SafeStyle -0.036Mixed
SafeStyle on real-world Dolly-15KLlama-3.1-8B-Instruct · Dolly-15K fine-tuningDelta ASR: +0.416 -> -0.045Improved
use style-matched vs style-mismatched safety data (ablation)Llama-3.1-8B-Instruct · list/poem fine-tuningASRImproved
increase safety-data mixing ratio beyond 50 examples (ablation)SafeStyle · list-style fine-tuningASR vs LC_WRMixed
SafeStyle vs baselines on style adaptation utilitySafeStyle · AlpacaEval stylesDelta LC_WRNo clear effect
PTST (safety prompt at inference) defenseQwen2.5-3B-Instruct · List-style fine-tuningDelta ASR: +0.269 -> +0.220No clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy