Research map / Vision-language and multimodal / CLIP and zero-shot
Exploring Interpretability for Visual Prompt Tuning with Cross-layer Concepts
Yubin Wang, XINYANG JIANG, De Cheng, Xiangqian Zhao, Zilong Wang, Dongsheng Li, Cairong Zhao
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 |
|---|---|---|---|
| use 4 prompted layers with decreasing prototype counts [16,13,10,7] vs 1 layer (ablation) | IVPT (DinoV2-B) · CUB-200-2011 | consistency score: 62.5 -> 75.3 | Improved |
| extend from 4 to 5 prompted layers (ablation) | IVPT (DinoV2-B) · CUB-200-2011 | consistency score: 75.3 -> 66.6 | Hurt |
| decreasing prototype count per layer vs constant 10 (ablation) | IVPT (DinoV2-B, 4 layers) · CUB-200-2011 | stability score: 71.2 -> 75.9 | Improved |
| decreasing vs constant prototype count per layer (ablation) | IVPT (DinoV2-B, 4 layers) · CUB-200-2011 | consistency score: 75.6 (constant) vs 75.3 (decreasing) | No clear effect |
| add concept prototypes to VPT-Deep | VPT-Deep (DinoV2-B) · CUB-200-2011 | consistency score: 14.6 -> 70.2 | Improved |
| add concept prototypes to VPT-Deep | VPT-Deep (DinoV2-B) · CUB-200-2011 | accuracy (%): 89.1 -> 90.3 | Improved |
| full IVPT (cross-layer, fusion) vs VPT-Deep with prototypes | DinoV2-B · CUB-200-2011 | consistency score: 70.2 -> 75.3 | Improved |
| IVPT vs Huang et al. on small backbone | DinoV2-S · CUB-200-2011 | consistency score: 65.7 -> 63.5 | Hurt |
| add intra-region feature aggregation (ablation) | IVPT baseline · CUB-200-2011 | accuracy (%): over 1% increase | Improved |
| add cross-layer prototypes and fine-to-coarse fusion (ablation) | IVPT baseline · CUB-200-2011 | consistency and stability scores | Improved |
Closest papers by meaning
- VIPAMIN: Visual Prompt Initialization via Embedding Selection and Subspace Expansion (NeurIPS 2025 · similarity 0.90)
- Learning to Learn Better Visual Prompts (AAAI 2024 · similarity 0.89)
- IPO: Interpretable Prompt Optimization for Vision-Language Models (NeurIPS 2024 · similarity 0.88)
- Revisiting the Power of Prompt for Visual Tuning (ICML 2024 · similarity 0.88)
- Draw-and-Understand: Leveraging Visual Prompts to Enable MLLMs to Comprehend What You Want (ICLR 2025 · similarity 0.88)
- Prompt Tuning In a Compact Attribute Space (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.
