Research map / Vision-language and multimodal / CLIP and zero-shot
SS-TPT: Stability and Suitability-Guided Test-Time Prompt Tuning for Adversarially Robust Vision-Language Models
Sunoh Kim, Daeho Um
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 SS-guided consistency loss (ablation) | SS-TPT · 10 fine-grained datasets | adversarial accuracy (%): 45.5 -> 46.8 | Improved |
| add SS-weighted prediction (ablation) | SS-TPT · 10 fine-grained datasets | adversarial accuracy (%): 42.0 -> 46.8 | Improved |
| add stability score on top of suitability (ablation) | SS-TPT · 10 fine-grained datasets | adversarial accuracy (%): 46.3 -> 46.8 | Improved |
| add suitability score on top of stability (ablation) | SS-TPT · 10 fine-grained datasets | adversarial accuracy (%): 45.4 -> 46.8 | Improved |
| use selected views instead of all views for consistency loss (ablation) | SS-TPT · Caltech101 | adversarial accuracy (%): 79.1 -> 81.0 | Improved |
| use min-max normalization instead of z-score (ablation) | SS-TPT · Caltech101 | adversarial accuracy (%): 78.9 -> 81.0 | Improved |
| add affine transformation to stability augmentations (ablation) | SS-TPT · Caltech101 | adversarial accuracy (%): 77.2 -> 81.0 | Improved |
| add horizontal flip to stability augmentations (ablation) | SS-TPT · Caltech101 | adversarial accuracy (%): 81.0 -> 80.5 | Hurt |
| use inverse mean distance instead of mean similarity for suitability (ablation) | SS-TPT · Caltech101 | adversarial accuracy (%): 80.6 -> 81.0 | No clear effect |
| use Jensen-Shannon instead of symmetric KL divergence for stability (ablation) | SS-TPT · Caltech101 | adversarial accuracy (%): 80.8 -> 81.0 | No clear effect |
| decrease loss balance lambda below 1 (ablation) | SS-TPT · Caltech101 | adversarial accuracy (%) | Hurt |
| combine stability and suitability (alpha=0.4) instead of either alone (ablation) | SS-TPT · Caltech101 | adversarial accuracy (%) | Improved |
Closest papers by meaning
- R-TPT: Improving Adversarial Robustness of Vision-Language Models through Test-Time Prompt Tuning (CVPR 2025 · similarity 0.93)
- One Prompt Word is Enough to Boost Adversarial Robustness for Pre-trained Vision-Language Models (CVPR 2024 · similarity 0.91)
- TAPT: Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models (CVPR 2025 · similarity 0.90)
- C-TPT: Calibrated Test-Time Prompt Tuning for Vision-Language Models via Text Feature Dispersion (ICLR 2024 · similarity 0.89)
- On the Test-Time Zero-Shot Generalization of Vision-Language Models: Do We Really Need Prompt Learning? (CVPR 2024 · similarity 0.89)
- Black-Box Test-Time Prompt Tuning for Vision-Language Models (AAAI 2025 · 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.
