Research map / Adversarial robustness and security / Attacks on text and graphs
Dynamic Token Reweighting for Robust Vision-Language Models
Tanqiu Jiang, Jiacheng Liang, Rongyi Zhu, Jiawei Zhou, Fenglong Ma, Ting 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 DTR visual token reweighting | llava-llama2-7b · HADES S+T+A | ASR: 56.9% -> 15.9% | Improved |
| apply DTR to a weakly aligned VLM | llava-1.5-vicuna-7b · HADES S+T+A | ASR: 80.8% -> 9.1% | Improved |
| apply DTR to text-driven harmful prompts | DTR · VLGuard safe image + harmful text | LLM-judged harmfulness: 66.5% -> 7.4% | Improved |
| use DTR instead of ShiftDC (image-to-text conversion) | llava-llama2-7b · MM-Vet benign queries | average inference time: 10.66s -> 4.01s | Improved |
| add DTR defense | llava-llama2-7b · MM-Vet OCR | VLC score: 33.4 -> 30.6 | Hurt |
| add DTR defense | llava-llama2-7b · MM-Vet Spatial Awareness | VLC score: 36.8 -> 39.1 | Improved |
| add CoCA defense | llava-llama2-7b · MM-Vet OCR | VLC score: 33.4 -> 17.4 | Hurt |
| increase reference samples n_ref beyond 32 (ablation) | DTR · HADES ASR reduction | ASR reduction | No clear effect |
| increase optimization steps m (ablation) | DTR · HADES and MM-Vet | ASR and VLC | Improved |
| increase token eviction rate from 0% to 50% (ablation) | DTR · HADES ASR reduction | ASR reduction | No clear effect |
| use lambda = 0.1 in the utility-preserving term (ablation) | DTR · HADES and MM-Vet | ASR and VLC trade-off | Improved |
| nullify weights of tokens with small alpha (adaptive attack) (ablation) | DTR · HADES adaptive attack | ASR-R | Hurt |
Closest papers by meaning
- JailBound: Jailbreaking Internal Safety Boundaries of Vision-Language Models (NeurIPS 2025 · similarity 0.91)
- VERA-V: Variational Inference Framework for Jailbreaking Vision-Language Models (ICML 2026 · similarity 0.91)
- When Think-with-Image Meets Safety: What Determines Multimodal Jailbreak Robustness? (NeurIPS 2026 · similarity 0.90)
- SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore Mechanism (NeurIPS 2025 · similarity 0.90)
- MMJ-Bench: A Comprehensive Study on Jailbreak Attacks and Defenses for Vision Language Models (AAAI 2025 · similarity 0.90)
- FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts (AAAI 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.
