Research map / Label-efficient and robust learning / Test-time adaptation
SwapPrompt: Test-Time Prompt Adaptation for Vision-Language Models
Xiaosong Ma, Jie ZHANG, Song Guo, Wenchao 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 |
|---|---|---|---|
| add swapped prediction loss L_swap on top of augmented pseudo-label loss (ablation) | SwapPrompt · 6 datasets avg | accuracy: 69.39 -> 70.68 | Improved |
| add image augmentation to UPL (ablation) | UPL · 6 datasets | accuracy: 68.65 -> 69.39 | Improved |
| use EMA target prompt (0.99) instead of fixed target prompt (ablation) | SwapPrompt · 6 datasets avg | accuracy: 69.33 -> 70.68 | Improved |
| use EMA target prompt instead of target identical to online prompt (decay 0) (ablation) | SwapPrompt · 6 datasets avg | accuracy: 69.65 -> 70.68 | Improved |
| swapped prediction L_swap versus L1 (no prompt interaction) (ablation) | SwapPrompt · 5 datasets avg | accuracy: 71.32 -> 72.45 | Improved |
| swapped prediction L_swap versus L2 (no cross-view contrast) (ablation) | SwapPrompt · 5 datasets avg | accuracy: 71.24 -> 72.45 | Improved |
| add top-K confident data selection (K=16) (ablation) | SwapPrompt · 5 datasets avg | accuracy: 69.95 -> 72.45 | Improved |
| increase K from 1 to 16 (ablation) | SwapPrompt · 5 datasets avg | accuracy: 68.74 -> 72.45 | Improved |
| increase context length from 4 to 16 tokens (ablation) | SwapPrompt · 5 datasets avg | accuracy: 72.45 -> 71.83 | Hurt |
| hand-crafted or pre-ImageNet prompt initialization instead of random (ablation) | SwapPrompt · 5 datasets avg | accuracy: 72.45 vs 72.33 / 72.36 | No clear effect |
| change alpha/beta weights from 1/1 (ablation) | SwapPrompt · 5 datasets avg | accuracy: 72.45 -> 71.77 / 72.12 | No clear effect |
| online mini-batch test-time adaptation instead of full offline | SwapPrompt · EuroSAT | accuracy: 46.64 -> 42.45 | Mixed |
Closest papers by meaning
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.92)
- Efficient Test-Time Adaptation of Vision-Language Models (CVPR 2024 · similarity 0.90)
- Black-Box Test-Time Prompt Tuning for Vision-Language Models (AAAI 2025 · similarity 0.90)
- On the Test-Time Zero-Shot Generalization of Vision-Language Models: Do We Really Need Prompt Learning? (CVPR 2024 · similarity 0.90)
- The Illusion of Progress? A Critical Look at Test-Time Adaptation for Vision-Language Models (NeurIPS 2025 · similarity 0.90)
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · 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.
