Research map / Label-efficient and robust learning / Test-time adaptation
Align Your Prompts: Test-Time Prompting with Distribution Alignment for Zero-Shot Generalization
Jameel Hassan Abdul Samadh, Hanan Gani, Noor Hazim Hussein, Muhammad Uzair Khattak, Muzammal Naseer, Fahad Khan, Salman Khan
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 token distribution alignment to MaPLe+TPT | MaPLe+TPT · ImageNet OOD variants | Top-1 accuracy (OOD average): 62.31 -> 63.55 | Improved |
| Add token distribution alignment to MaPLe+TPT | MaPLe · ImageNet-Sketch | Top-1 accuracy: 48.16 -> 50.23 | Improved |
| Add token distribution alignment to MaPLe+TPT | MaPLe · PUG-ImageNet (Worlds) | Top-1 accuracy: 42.00 -> 43.24 | Improved |
| Add token distribution alignment to MaPLe+TPT | MaPLe · Cross-dataset evaluation | Top-1 accuracy (average): 66.50 -> 66.92 | Improved |
| Add entropy minimization alongside distribution alignment (ablation) | PromptAlign · ImageNet-A | Top-1 accuracy: 50.85 -> 59.37 | Improved |
| Use distribution alignment alone without entropy minimization (ablation) | PromptAlign · ImageNet-A | Top-1 accuracy: 50.90 -> 50.85 | No clear effect |
| Replace mean-and-variance alignment with fifth-order Central Moment Discrepancy (ablation) | PromptAlign · Cross-dataset evaluation | Top-1 accuracy (average): 66.92 -> 66.99 | No clear effect |
| Replace ImageNet proxy statistics with statistics from a LAION400M subset (ablation) | PromptAlign · Cross-dataset evaluation | Top-1 accuracy (average): 66.92 -> 67.17 | Improved |
| Estimate test statistics from a bag of same-class samples (N=5) rather than one test sample (ablation) | PromptAlign · Cross-dataset evaluation | Top-1 accuracy (average): 66.92 -> 69.59 | Improved |
| Replace source-statistics alignment with regularization to previous prompts (ablation) | PromptAlign · Cross-dataset evaluation | Top-1 accuracy (average): 66.92 -> 62.87 | Hurt |
| Apply alignment loss across visual encoder layers 1 to 3 (ablation) | PromptAlign · ImageNet-A | Top-1 accuracy | Improved |
| Increase the number of augmented views (ablation) | PromptAlign · ImageNet validation and evaluated datasets | Top-1 accuracy | Improved |
Closest papers by meaning
- Understanding prompt engineering may not require rethinking generalization (ICLR 2024 · similarity 0.91)
- Robust Test-Time Adaptation for Zero-Shot Prompt Tuning (AAAI 2024 · similarity 0.91)
- On the Test-Time Zero-Shot Generalization of Vision-Language Models: Do We Really Need Prompt Learning? (CVPR 2024 · similarity 0.90)
- A Similarity Paradigm Through Textual Regularization Without Forgetting (AAAI 2025 · similarity 0.90)
- OSLoPrompt: Bridging Low-Supervision Challenges and Open-Set Domain Generalization in CLIP (CVPR 2025 · similarity 0.90)
- Any-Shift Prompting for Generalization over Distributions (CVPR 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.
