Research map / Large language models / Language, safety and interpretability
Furina: Fragmented Uncertainty-Driven Refusal Instability Attack
Tongxi Wu, Jian Zhang, Yang Gao
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 semantic probes (ablation) | Furina · MM-SafetyBench Hate Speech (GPT-4o-mini) | ASR (%): 0.00 -> 91.41 | Improved |
| add auxiliary synthesis of probe answers (ablation) | Furina · MM-SafetyBench Hate Speech (GPT-4o-mini) | ASR (%): 7.98 -> 91.41 | Improved |
| add scene interpretation (ablation) | Furina · MM-SafetyBench Hate Speech (GPT-4o-mini) | ASR (%): 88.34 -> 91.41 | Improved |
| use dispersed probes instead of progressive ladder reformulation (ablation) | Furina · MM-SafetyBench Hate Speech (GPT-4o-mini) | ASR (%): 80.98 -> 91.41 | Improved |
| render scene as typography instead of diffusion image (ablation) | Furina · MM-SafetyBench on Claude-Haiku-4.5 | ASR (%): 75.48 -> 77.20 | Improved |
| add LlamaGuard-3-8B input filtering of probes as defense | Furina · HarmBench (GPT-4o-mini) | ASR (%): 94.0 -> 93.5 | No clear effect |
| add aggressive p95 perplexity filtering as defense | Furina · HarmBench (GPT-4o-mini) | ASR (%): 94.0 -> 86.5 | Hurt |
| add p99 perplexity filtering as defense | Furina · HarmBench (GPT-4o-mini) | ASR (%): 94.0 -> 93.0 | No clear effect |
| increase rewrite strength from original to semantic rewrite | LLaMA-2-7B (AdvBench) · AdvBench rewrite ladder | ASR: 0.01 -> 0.42 | Improved |
| increase rewrite strength from original to semantic rewrite | LLaMA-3.2-3B (AdvBench) · AdvBench rewrite ladder | ASR: 0.02 -> 0.90 | Improved |
| suppress refusal-direction alignment in last 4 layers (activation patching) (ablation) | LLaMA-2-7B-Chat · Original rewrite | ASR: 0.01 -> 0.04 | Improved |
| suppress refusal-direction alignment in last 4 layers (activation patching) (ablation) | LLaMA-2-7B-Chat · Semantic rewrite | ASR: 0.42 -> 0.42 | No clear effect |
Closest papers by meaning
- Fail-Closed Alignment for Large Language Models (NeurIPS 2026 · similarity 0.88)
- Improving LLM Safety Alignment with Dual-Objective Optimization (ICML 2025 · similarity 0.88)
- Refusal Direction is Universal Across Safety-Aligned Languages (NeurIPS 2025 · similarity 0.87)
- Why Jailbreaks Succeed in Diffusion Language Models: An Energy Landscape Analysis (NeurIPS 2026 · similarity 0.86)
- One Model Transfer to All: On Robust Jailbreak Prompts Generation against LLMs (ICLR 2025 · similarity 0.86)
- Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment (CVPR 2025 · similarity 0.86)
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.
