Research map / Vision-language and multimodal / Multimodal LLMs
AdvDreamer Unveils: Are Vision-Language Models Truly Ready for Real-World 3D Variations?
Shouwei Ruan, Hanqing Liu, Yao Huang, Xiaoqi Wang, Caixin Kang, Hang Su, Yinpeng Dong, Xingxing Wei
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 naturalness reward model (NRM) to adversarial optimization (ablation) | AdvDreamer · ImageNet, OpenCLIP ViT-B/16 | Score_R realism: 1.60 -> 2.52 | Improved |
| add naturalness reward model (NRM) (ablation) | AdvDreamer · ImageNet, OpenCLIP ViT-B/16 | Score_P plausibility: 1.39 -> 2.57 | Improved |
| add naturalness reward model (NRM) (ablation) | AdvDreamer · ImageNet, OpenCLIP ViT-B/16 | top-1 accuracy (attack strength): 48.6 -> 54.0 | Hurt |
| use ISP loss instead of MF_it (ablation) | AdvDreamer · ImageNet p*(Theta), OpenCLIP ViT-B/16 | top-1 accuracy: 58.1% -> 54.0% | Improved |
| use ISP loss instead of MF_ii (ablation) | AdvDreamer · ImageNet Theta, OpenCLIP ViT-B/16 | top-1 accuracy: 22.2% -> 17.4% | Improved |
| use ISP loss instead of MF_ii+MF_tt (ablation) | AdvDreamer · ImageNet p*(Theta), OpenCLIP ViT-B/16 | top-1 accuracy: 54.9% -> 54.0% | No clear effect |
| optimize adversarial distribution p*(Theta) vs random 3D transform | OpenCLIP 149M · ImageNet zero-shot | top-1 accuracy: 62.6 -> 54.0 | Improved |
| use distribution center Theta* instead of sampling p*(Theta) | OpenCLIP 149M · ImageNet zero-shot | top-1 accuracy: 54.0 -> 18.0 | Improved |
| scale up variance of p*(Theta) beyond the optimum (ablation) | AdvDreamer · ImageNet, 5 images x 100 samples | attack success rate | Hurt |
| increase optimization iterations (ablation) | AdvDreamer · ImageNet, OpenCLIP/BLIP ViT-B/16 | attack success rate | Improved |
| transfer attack with OpenCLIP surrogate to OpenAI CLIP using Theta* | AdvDreamer · ImageNet cross-model | top-1 accuracy: 85.3 -> 29.7 | Improved |
| reproduce Adv-3DT samples physically | AdvDreamer · Physical zero-shot classification | accuracy: 100.0 -> 51.3 | Hurt |
Closest papers by meaning
- Improving Zero-Shot Adversarial Robustness in Vision-Language Models by Closed-form Alignment of Adversarial Path Simplices (ICML 2025 · similarity 0.88)
- Unifying 2D and 3D Vision-Language Understanding (ICML 2025 · similarity 0.87)
- Let Language Constrain Geometry: Vision–Language Models as Semantic and Spatial Critics for 3D Generation (ICML 2026 · similarity 0.87)
- VLM-RobustBench: A Comprehensive Benchmark for Robustness of Vision-Language Models (ICML 2026 · similarity 0.87)
- From Thousands to Billions: 3D Visual Language Grounding via Render-Supervised Distillation from 2D VLMs (ICML 2025 · similarity 0.87)
- MonoVLM: Monocular 3D Visual Grounding with 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.
