Research map / Label-efficient and robust learning / Test-time adaptation
Flatness Guided Test-Time Adaptation for Vision-Language Models
Aodi Li, Liansheng Zhuang, Xiao Long, Houqiang Li, Shafei Wang
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 sharpness-aware prompt tuning (SAPT) (ablation) | CoOp · ImageNet + variants (Avg.) | top-1 accuracy (%): 61.72 -> 62.64 | Improved |
| add sharpness-based test sample selection (STSS) (ablation) | CoOp · ImageNet + variants (Avg.) | top-1 accuracy (%): 61.72 -> 66.48 | Improved |
| add STSS on top of SAPT (full FGA) (ablation) | CoOp+SAPT · ImageNet + variants (Avg.) | top-1 accuracy (%): 62.64 -> 68.04 | Improved |
| add SAPT on top of STSS (ablation) | CoOp+STSS · ImageNet + variants (Avg.) | top-1 accuracy (%): 66.48 -> 68.04 | Improved |
| add SAPT on top of STSS (ablation) | CoOp+STSS · ImageNet OOD (OOD Avg.) | top-1 accuracy (%): 64.60 -> 66.55 | Improved |
| replace test-time prompt tuning with FGA | TPT+CoOp · ImageNet OOD (OOD Avg.) | top-1 accuracy (%): 61.67 -> 66.55 | Improved |
| replace test-time prompt tuning with FGA | TPT+CoOp · Cross-dataset from ImageNet (10 datasets Avg.) | top-1 accuracy (%): 65.66 -> 67.60 | Improved |
| use FGA instead of single-sample SAR | SAR+CoOp · ImageNet OOD (OOD Avg.) | top-1 accuracy (%): 61.75 -> 66.55 | Improved |
| set test perturbation rho' to zero (entropy only, no sharpness) (ablation) | FGA · ImageNet-R | top-1 accuracy (%) | Hurt |
| vary rho' (non-monotonic) (ablation) | FGA · ImageNet-R | top-1 accuracy (%) | Mixed |
| retain more augmented samples (ablation) | FGA · ImageNet-R | top-1 accuracy (%) | Mixed |
| FGA vs online TDA on cross-dataset | TDA · Cross-dataset from ImageNet (10 datasets Avg.) | top-1 accuracy (%): 67.53 -> 67.60 | Improved |
Closest papers by meaning
- The Illusion of Progress? A Critical Look at Test-Time Adaptation for Vision-Language Models (NeurIPS 2025 · similarity 0.94)
- Realistic Test-Time Adaptation of Vision-Language Models (CVPR 2025 · similarity 0.93)
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.92)
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.92)
- Efficient Test-Time Adaptation of Vision-Language Models (CVPR 2024 · similarity 0.91)
- DOTA: Distributional Test-time Adaptation of Vision-Language Models (NeurIPS 2025 · similarity 0.91)
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.
