Research map / Label-efficient and robust learning / Test-time adaptation
Long-tailed Test-Time Adaptation for Vision-Language Models
Xucong Wang, Zhe Zhao, Zekun Wang, Xiaofeng Cao, Xu Wang, Di Wu, Pengkun Wang, Yang 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 rebalancing shortcuts to discriminative prototypes (ablation) | L-TTA with DP, ViT-B/16 · ImageNet (imb=10) | accuracy (%): 68.68 -> 69.76 | Improved |
| add rebalancing shortcuts to exclusive prototypes (ablation) | L-TTA with EP, ViT-B/16 · ImageNet (imb=10) | accuracy (%): 67.54 -> 68.03 | Improved |
| combine DP and EP into synergistic prototypes (ablation) | L-TTA with RS, ViT-B/16 · ImageNet (imb=10) | accuracy (%): 69.76 -> 70.94 | Improved |
| add balanced entropy minimization to SyP+RS (ablation) | L-TTA, ViT-B/16 · ImageNet (imb=10) | accuracy (%): 70.94 -> 71.30 | Improved |
| add balanced entropy minimization to SyP+RS (ablation) | L-TTA, RN50 · ImageNet (imb=10) | accuracy (%): 59.59 -> 59.82 | Improved |
| drop discriminative or exclusive prototypes (ablation) | L-TTA · ImageNet (imb=10) | macro-F1 (%): -3.95% / -3.22% | Hurt |
| add class re-allocation loss (eta=1) to entropy minimization (ablation) | L-TTA · ImageNet (imb=10) | accuracy / macro-F1 (%): +1.19% / +1.64% | Improved |
| use penalty factor beta=1 instead of 0.1 or 8 in BEM (ablation) | L-TTA · ImageNet (imb=10) | accuracy / macro-F1 (%): up to 0.64%/0.85% | Improved |
| increase affinity coefficients lambda1 and lambda2 excessively (ablation) | L-TTA · ImageNet (imb=10) | accuracy | Hurt |
| add BEM to MTA | MTA, ViT-B/16 · ImageNet (imb=10) | accuracy (%): 67.79 -> 68.25 | Improved |
| add BEM loss to other TTA method versus classic LT methods (BS, LA) | MTA, RN50 · ImageNet (imb=10) | accuracy (%): 57.30 -> 57.84 | Improved |
| vary tail-class sampling probability epsilon from 0 to 1 (ablation) | L-TTA · ImageNet and Flowers | accuracy (%): 71.30 -> 71.52 | No clear effect |
Closest papers by meaning
- The Illusion of Progress? A Critical Look at Test-Time Adaptation for Vision-Language Models (NeurIPS 2025 · 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)
- Realistic Test-Time Adaptation of Vision-Language Models (CVPR 2025 · similarity 0.91)
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.91)
- DOTA: Distributional Test-time Adaptation of Vision-Language Models (NeurIPS 2025 · 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.
