Research map / Vision-language and multimodal / CLIP and zero-shot
Auto-Prompting: Layer-Specific Prompt Fusion Discovery via Differentiable Search
Xi Xiao, Xingjian Li, Yunbei Zhang, Cheng Han, Tianming Liu, Tianyang Wang, Runmin Jiang, Jihun Hamm, Xiao Wang, Min Xu
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 |
|---|---|---|---|
| replace fixed cross-attention fusion with searched layer-wise hybrid fusion (ablation) | VPT with frozen ViT-B/16 · VTAB-1k Natural | accuracy: 79.91 -> 82.88 | Improved |
| replace fixed cross-attention with searched fusion (ablation) | VPT with frozen ViT-B/16 · VTAB-1k Structured | accuracy: 58.67 -> 62.55 | Improved |
| replace fixed concat with searched fusion (ablation) | VPT with frozen ViT-B/16 · VTAB-1k Natural | accuracy: 78.42 -> 82.88 | Improved |
| replace fixed add with searched fusion (ablation) | VPT with frozen ViT-B/16 · VTAB-1k Natural | accuracy: 77.63 -> 82.88 | Improved |
| replace fixed affine with searched fusion (ablation) | VPT with frozen ViT-B/16 · VTAB-1k Structured | accuracy: 57.40 -> 62.55 | Improved |
| use searched fusion instead of VPT-Deep concatenation | ViT-B/16 · VTAB-1k mean | accuracy: 69.43 -> 77.01 | Improved |
| searched fusion under MAE pretraining vs VFPT | ViT-B/16 MAE · VTAB-1k Natural | accuracy: 53.59 -> 55.12 | Improved |
| searched fusion under MAE pretraining vs Partial-1 / Bias | ViT-B/16 MAE · VTAB-1k | accuracy: 39.01 vs 47.64 Structured (Partial-1) | Mixed |
| searched fusion under MoCo v3 | ViT-B/16 MoCo v3 · VTAB-1k Natural | accuracy: 77.47 -> 79.60 | Improved |
| searched fusion on Swin-Base vs VPT-Deep | Swin-Base · VTAB-1k Structured | accuracy: +10.51 | Improved |
| searched fusion adds inference cost | ViT-B/16 · VTAB-1k | latency per image: 14.8 -> 15.9 ms | Hurt |
| searched fusion adds search cost | ViT-B/16 · VTAB-1k | training GPU hours: 7.8 -> 10.8 GPU hours | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DA-VPT — DA-VPT: Semantic-Guided Visual Prompt Tuning for Vision Transformers (CVPR 2025) · better on 2 of 2 shared results
- LoR-VP — LoR-VP: Low-Rank Visual Prompting for Efficient Vision Model Adaptation (ICLR 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DA-VPT: Semantic-Guided Visual Prompt Tuning for Vision Transformers (CVPR 2025 · similarity 0.91)
- Facing the Elephant in the Room: Visual Prompt Tuning or Full finetuning? (ICLR 2024 · similarity 0.91)
- Revisiting the Power of Prompt for Visual Tuning (ICML 2024 · similarity 0.90)
- Learning to Learn Better Visual Prompts (AAAI 2024 · similarity 0.89)
- Token Coordinated Prompt Attention is Needed for Visual Prompting (ICML 2025 · similarity 0.89)
- SA²VP: Spatially Aligned-and-Adapted Visual Prompt (AAAI 2024 · similarity 0.89)
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.
