Research map / Vision-language and multimodal / Multimodal LLMs
Unsafe by Reciprocity: How Generation–Understanding Coupling Undermines Safety in Unified Multimodal Models
Kaishen Wang, Heng Huang
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 |
|---|---|---|---|
| generate image of decomposed object via the same UMM and feed back (full RICE) instead of concatenating decomposed text (ablation) | RICE G2U (Bagel) · AdvBench | ASR (%): 77.12 -> 89.62 | Improved |
| generate image of decomposed object (full RICE) instead of concatenating decomposed text (ablation) | RICE G2U (Janus-Pro-7B) · AdvBench | ASR (%): 81.35 -> 92.69 | Improved |
| pair query with random noise image (ablation) | Text-only attack (Bagel) · AdvBench | ASR (%): 47.88 -> 55.77 | Improved |
| pair query with image generated from the original query (Plain) (ablation) | Text-only attack (Bagel) · AdvBench | ASR (%): 47.88 -> 37.88 | Hurt |
| pair query with image generated from the original query (Plain) (ablation) | Text-only attack (Janus-Pro-7B) · JailbreakBench | ASR (%): 35.00 -> 74.00 | Improved |
| pair query with mismatched image (ablation) | Text-only attack (Bagel) · SafeBench | ASR (%): 40.20 -> 27.60 | Hurt |
| pair query with mismatched image (ablation) | Text-only attack (Janus-Pro-7B) · AdvBench | ASR (%): 35.77 -> 54.81 | Improved |
| add visual-only prompt expansion by the understanding function (RICE U2G) (ablation) | Bagel generation · I2P sexual | NudeNet ASR (%): 12.78 -> 42.32 | Improved |
| add visual-only prompt expansion (RICE U2G) (ablation) | Janus-Pro-7B generation · Explicit | MLLM ASR (%): 97.36 -> 97.10 | No clear effect |
| add Self-CoT planning before generation (ablation) | Janus-Pro-7B generation · I2P sexual | NudeNet ASR (%): 6.55 -> 5.37 | Hurt |
| use RICE G2U instead of textual jailbreaks (PAIR, AutoDAN, GCG) | Bagel · HarmBench | ASR (%): 72.00 -> 82.50 | Improved |
Closest papers by meaning
- When Understanding Becomes a Risk: Authenticity and Safety Risks in the Emerging Image Generation Paradigm (CVPR 2026 · similarity 0.88)
- Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models? (CVPR 2025 · similarity 0.88)
- Multimodal Situational Safety (ICLR 2025 · similarity 0.87)
- Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy (CVPR 2025 · similarity 0.87)
- VLSU: Mapping the Limits of Joint Multimodal Understanding for AI Safety (ICLR 2026 · similarity 0.87)
- LLaVAShield: Safeguarding Multimodal Multi-Turn Dialogues in Vision-Language Models (CVPR 2026 · similarity 0.87)
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.
