Research map / Large language models / Instruction and fine-tuning
EVA: Editing for Versatile Alignment Against Jailbreaks
Yi Wang, Hongye Qiu, Yue Xu, Sibei Yang, Zhan Qin, Minlie Huang, Wenjie Wang
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 |
|---|---|---|---|
| apply EVA (text) editing | Vicuna-7B-v1.5 · HarmBench AutoDAN | ASR: 0.69 -> 0.04 | Improved |
| apply EVA to a VLM | LLaVA-1.5-7B · ADV-16 on HarmBench | ASR: 0.71 -> 0.00 | Improved |
| apply EVA to a VLM | LLaVA-1.5-7B · MultiTrust Crossmodal | ASR: 0.88 -> 0.00 | Improved |
| apply EVA to a VLM | Qwen2.5-VL-7B · MM-SafetyBench SD+TYPO | ASR: 0.51 -> 0.00 | Improved |
| add visual-token editing to text-only EVA (ablation) | LLaVA-1.5-7B · MultiTrust Multimodal | ASR: 0.42 -> 0.13 | Improved |
| add visual-token editing to text-only EVA (ablation) | LLaVA-1.5-7B · MMStar | score: 27.3 -> 32.3 | Improved |
| adaptive attacks regenerated on edited model | EVA (text) · Vicuna-7B-v1.5 avg ASR | ASR: 0.068 -> 0.112 | Hurt |
| adaptive attacks regenerated on edited model | EVA (text) · Qwen2.5-7B avg ASR | ASR: 0.012 -> 0.113 | Hurt |
| adaptive ADV-16 attack on edited VLM | EVA · LLaVA-1.5-7B ADV-16 avg ASR | ASR: 0.000 -> 0.125 | Hurt |
| select image token randomly instead of by attention (ablation) | EVA · Qwen2.5-VL-7B MM-Vet-v2 | score: 55.5 -> 10.4 | Hurt |
| select first (fixed) image token instead of by attention (ablation) | EVA · Qwen2.5-VL-7B MM-Vet-v2 | score: 55.5 -> 12.1 | Hurt |
| use attention-based token selection rather than L2 distance (ablation) | EVA · Qwen2.5-VL-7B MM-SafetyBench SD+TYPO | ASR: 0.08 -> 0.00 | Improved |
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)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.90)
- Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models (ICLR 2025 · similarity 0.89)
- Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications (ICML 2024 · similarity 0.88)
- MMJ-Bench: A Comprehensive Study on Jailbreak Attacks and Defenses for Vision Language Models (AAAI 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.
