Research map / Vision-language and multimodal / Multimodal LLMs
"Eyes Closed, Safety On: Protecting Multimodal LLMs via Image-to-Text Transformation"
Yunhao Gou, Kai Chen, Zhili LIU, Lanqing Hong, Hang Xu, Zhenguo Li, Dit-Yan Yeung, James Kwok, Yu Zhang
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 |
|---|---|---|---|
| remove images when querying MLLMs with malicious text (ablation) | five MLLMs · VLSafe | harmless rate: about 20% -> nearly 100% | Improved |
| apply ECSO (self-check, query-aware caption, answer without image) | LLaVA-1.5-7B · MM-SafetyBench SD+OCR | harmless rate: 32.1 -> 86.4 | Improved |
| apply ECSO on SD images where Direct is already fairly safe | LLaVA-1.5-7B · MM-SafetyBench SD | harmless rate: 85.0 -> 95.4 | Improved |
| apply ECSO | five MLLMs · MME, MM-Vet, MMBench utility | utility: e.g. LLaVA-1.5-7B MMBench 64.6 -> 64.2; ShareGPT4V MME-C 376.4 -> 380.7 | No clear effect |
| keep image features in addition to the query-aware caption at step 4 (ablation) | ECSO · MM-SafetyBench SD+OCR and VLSafe | harmless rate | Hurt |
| use generic caption instead of query-aware caption (ablation) | ECSO on ShareGPT4V-7B · MME-P and MMBench | utility: MME-P 1567.1 -> 1562.7; MMBench 66.1 -> 65.8 | Hurt |
| add HARMLESS and ETHICAL safety instruction to ECSO final step (ablation) | ECSO · VLSafe | harmless rate: 71.3 -> 91.8 for LLaVA-1.5-7B | Improved |
| detect harm from the model response instead of image and query (ablation) | ECSO step 2 · 1000 labelled responses | detection accuracy: 75.6 -> 95.0 for LLaVA-1.5-7B | Improved |
| use original LLM instead of finetuned LLM in step 4 (ablation) | ECSO · VLSafe | harmless rate: 91.8 -> 93.3 for LLaVA-1.5-7B | Improved |
| SFT on ECSO-generated responses instead of ground-truth VLGuard responses (ablation) | LLaVA-1.5-7B · MM-SafetyBench SD+OCR | harmless rate / utility: utility 1679.7 -> 1727.2 | Improved |
| combine finetuning with ECSO at inference (ablation) | LLaVA-1.5-7B finetuned on VLGuard · MM-SafetyBench SD+OCR | harmless rate: 73.68 -> 89.18 | Improved |
Closest papers by meaning
- GuardAlign: Robust Safety Alignment in Multimodal Large Language Models (ICLR 2026 · similarity 0.93)
- MLLMGuard: A Multi-dimensional Safety Evaluation Suite for Multimodal Large Language Models (NeurIPS 2024 · similarity 0.92)
- Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models? (CVPR 2025 · similarity 0.92)
- SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore Mechanism (NeurIPS 2025 · similarity 0.91)
- Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs (CVPR 2026 · similarity 0.91)
- Risk Awareness Injection: Calibrating Vision-Language Models for Safety without Compromising Utility (ICML 2026 · similarity 0.91)
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.
