Research map / Vision-language and multimodal / Cross-modal learning
Co-Steer: Cross-Modal Collaborative Steering for Jailbreaking MLLMs
Jingmin Zhu, Rollin Omari, Tamas Abraham, Junae Kim, Amardeep Kaur, Trung Le, Dinh Q Phung, Qiuhong Ke
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 |
|---|---|---|---|
| add steering alignment loss with coordinated alternating optimization (vs independent optimization) (ablation) | Co-Steer · MM-SafetyBench | ASR (%): 79.62 -> 87.20 (InternVL2); 78.19 -> 84.65 (Qwen2-VL); 86.85 -> 88.94 (MiniGPT-4) | Improved |
| combine visual and textual perturbations rather than a single modality (ablation) | Co-Steer · MM-SafetyBench | ASR (%): 55.28 -> 79.62 (InternVL2, visual-only vs independent) | Improved |
| extract steering vector with SVD instead of cross-modal mean (ablation) | Co-Steer · MM-SafetyBench | ASR (%): 84.58 -> 87.20 (InternVL2); 82.19 -> 84.65 (Qwen2-VL); 86.47 -> 88.94 (MiniGPT-4) | Improved |
| use both modalities when building the steering vector instead of single-modality averaging (ablation) | Co-Steer · MM-SafetyBench | ASR (%): 62.34 / 68.72 -> 84.58 (InternVL2) | Improved |
| steer at middle fusion layer 15 rather than early or late layers (ablation) | Co-Steer · MM-SafetyBench (InternVL2-8B) | ASR (%): 62.34 (L3) / 68.47 (L25) -> 87.20 (L15) | Improved |
| increase SVD sample pairs from 10 to 100 (ablation) | Co-Steer · MM-SafetyBench (InternVL2-8B) | ASR (%): 76.54 -> 87.20 | Improved |
| increase SVD pairs beyond 100 (150, 200) (ablation) | Co-Steer · MM-SafetyBench (InternVL2-8B) | ASR (%): 87.20 -> 87.58 | No clear effect |
| set coupling weight lambda to 1.0-1.5 vs lambda=0 or 2.0 (ablation) | Co-Steer · MM-SafetyBench | ASR (%) | Improved |
| widen steering-loss layer window from 1 to 3 or 5 layers (ablation) | Co-Steer · MM-SafetyBench (Qwen2-VL-7B) | ASR (%): 84.65 -> 86.43 (width 3); 83.39 (width 5) | Mixed |
| reduce visual budget from eps 32/255 to 16/255 (ablation) | Co-Steer · MM-SafetyBench | ASR (%): 84.65 -> 80.89 (Qwen2-VL); 87.20 -> 85.06 (InternVL2) | Hurt |
| Co-Steer vs strongest baseline BAP in black-box transfer | Co-Steer · Black-box commercial models | ASR (%): +7.62 to +8.95 pp over BAP | Improved |
| visual-only perturbation (VAJM) for transfer | VAJM · Black-box commercial models | ASR (%): 10.55-14.69 vs vanilla 8.86-10.83 | No clear effect |
Closest papers by meaning
- Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy (CVPR 2025 · similarity 0.90)
- Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment (CVPR 2025 · similarity 0.90)
- SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore Mechanism (NeurIPS 2025 · similarity 0.90)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.89)
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.89)
- Distraction is All You Need for Multimodal Large Language Model Jailbreaking (CVPR 2025 · similarity 0.89)
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.
