Research map / Vision-language and multimodal / Multimodal LLMs
How Many Unicorns Are in This Image? A Safety Evaluation Benchmark for Vision LLMs
Haoqin Tu, Chenhang Cui, Zijun Wang, Yiyang Zhou, Bingchen Zhao, Junlin Han, Wangchunshu Zhou, Huaxiu Yao, Cihang Xie
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 |
|---|---|---|---|
| append counterfactual suffix to OOD questions | 21 VLLMs · OODCV-Counterfactual | Overall accuracy: average drop of 17.1% overall, 33.2% on Yes/No | Hurt |
| counterfactual descriptions on GPT-4V | GPT-4V · OODCV-Counterfactual | Overall accuracy: 80.61 -> 69.00 | Hurt |
| use ensemble-free single CLIP ViT attack Mix.Attack instead of AttackBard at eps=64/255 | misleading attack on VLLMs · NIPS17 misleading attack | Missing rate: +8.4% average over AttackBard | Improved |
| use a mix of target text objects (Mix.Attack) instead of a single one (Sin.Attack) (ablation) | CLIP ViT attack · NIPS17 misleading attack | Missing rate: 79.7% -> 87.5% (LLaVA Vicuna-v0-7B) | Improved |
| use less common sketch object categories | VLLMs · Sketchy-VQA vs Sketchy-Challenging | F1: 4.4% average F1 drop | Hurt |
| apply direct vision jailbreak adversarial images to the three source VLLMs | MiniGPT4, LLaVA, InstructBLIP · toxic output generation | Toxic output likelihood: 2.1x higher vs clean images; only ~5% marginal across all VLLMs | Improved |
| vision-language training of LLM into VLLM | LLaVA family · text jailbreak ASR | Attack success rate: vanilla 1.3% -> 6.8%; white-box 77.5% -> 94.8% | Hurt |
| stronger LLM in models with few tunable parameters | MiniGPT4, InstructBLIP · OOD tasks | OOD accuracy: average decrease 2.1% and 0.2% | No clear effect |
| stronger LLM in fully fine-tuned models | LLaVA, mPLUG-Owl · OOD tasks | OOD accuracy: average boost 52.3% and 10.0% | Improved |
| no explicit vision encoder (Fuyu) | Fuyu · adversarial attack ASR | ASR: 33.5% lower ASR than other VLLMs | Improved |
Closest papers by meaning
- Safety Fine-Tuning at (Almost) No Cost: A Baseline for Vision Large Language Models (ICML 2024 · similarity 0.91)
- ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time (ICLR 2025 · similarity 0.90)
- Retention Score: Quantifying Jailbreak Risks for Vision Language Models (AAAI 2025 · similarity 0.89)
- Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language Models (ICLR 2026 · similarity 0.89)
- MVI-Bench: A Comprehensive Benchmark for Evaluating Robustness to Misleading Visual Inputs in LVLMs (ICML 2026 · similarity 0.88)
- ELITE: Enhanced Language-Image Toxicity Evaluation for Safety (ICML 2025 · similarity 0.88)
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.
