Research map / Adversarial robustness and security / Attacks on text and graphs
ARMs: Adaptive Red-Teaming Agent against Multimodal Models with Plug-and-Play Attacks
Zhaorun Chen, Xun Liu, Mintong Kang, Jiawei Zhang, Minzhou Pan, Shuang Yang, Bo Li
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 layered memory retrieval (top-k=3) to ARMs (ablation) | ARMs · StrongReject vs Claude-3.7-Sonnet | ASR: 89.4 -> 95.2 | Improved |
| retrieve 7 memories instead of 3 (ablation) | ARMs · StrongReject vs Claude-3.7-Sonnet | ASR: 95.2 -> 85.6 | Hurt |
| add epsilon-greedy decay (lambda=1) instead of pure exploration (ablation) | ARMs · StrongReject vs Claude-3.7-Sonnet | ASR: 86.0 -> 95.2 | Improved |
| add vision modality input to the backbone (ablation) | ARMs · StrongReject vs Claude-3.7-Sonnet | ASR: 92.1 -> 95.2 | Improved |
| add reasoning instructions to the backbone (ablation) | ARMs · StrongReject vs Claude-3.7-Sonnet | ASR: 82.9 -> 95.2 | Improved |
| use GPT-4o instead of Qwen3-235B as backbone (ablation) | ARMs · StrongReject vs Claude-3.7-Sonnet | ASR: 80.6 -> 95.2 | Improved |
| use ARMs multi-step orchestration instead of brute-force oracle over all strategies | ARMs · StrongReject vs Claude-3.7-Sonnet | ASR: 84.0 -> 95.2 | Improved |
| use o3-mini judge instead of GPT-4o judge (ablation) | ARMs · EU AI Act vs GPT-4o victim | ASR: 76.9 -> 100.0 | Improved |
| use Claude-3.7-Sonnet backbone instead of GPT-4o (ablation) | ARMs · StrongReject vs GPT-4o victim | ASR: 93.1 -> 99.4 | Improved |
| fine-tune InternVL3-38B with ARMs-Bench safety data | InternVL3-38B · StrongReject ARMs ASR | ASR: 98.5 -> 69.6 | Improved |
| fine-tune InternVL3-38B with ARMs-Bench safety data | InternVL3-38B · Policy-based ARMs ASR | ASR: 87.7 -> 29.2 | Improved |
| fine-tune InternVL3-38B with JailbreakV safety data | InternVL3-38B · StrongReject ARMs ASR | ASR: 98.5 -> 98.0 | No clear effect |
Closest papers by meaning
- ReaLM: A Unified Red-Teaming Benchmark for Physical-World VLMs (NeurIPS 2026 · similarity 0.93)
- RedVLA: Physical Red Teaming for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- TreeTeaming: Autonomous Red-Teaming of Vision-Language Models via Hierarchical Strategy Exploration (CVPR 2026 · similarity 0.91)
- TRUST-VLM: Thorough Red-Teaming for Uncovering Safety Threats in Vision-Language Models (ICML 2025 · similarity 0.91)
- DiveR-CT: Diversity-enhanced Red Teaming Large Language Model Assistants with Relaxing Constraints (AAAI 2025 · similarity 0.90)
- Learning Diverse Attacks on Large Language Models for Robust Red-Teaming and Safety Tuning (ICLR 2025 · similarity 0.90)
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.
