Research map / Label-efficient and robust learning / Test-time adaptation
Bridging Domain Expertise and Generalization for Performance Estimation
Shuxuan Li, Zhilin Zhao, Quyu Kong, Wei-Shi Zheng
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 dynamic test-time temperature calibration (ablation) | FRAP (CLIP) · Average over 18 datasets | MAE (%): 8.20 -> 6.53 | Improved |
| use learned test-time temperature instead of fixed 0.05 (ablation) | FRAP (CLIP) · Average over 18 datasets | MAE (%): 8.19 -> 6.53 | Improved |
| use learned test-time temperature instead of fixed 0.01 (ablation) | FRAP (CLIP) · Average over 18 datasets | MAE (%): 6.61 -> 6.53 | Improved |
| use test-time calibration instead of raw CLIP similarity (ablation) | FRAP (CLIP) · CIFAR-10 | ECE: 0.590 -> 0.128 | Improved |
| use test-time calibration instead of raw CLIP similarity (ablation) | FRAP (CLIP) · MNIST | ECE: 0.027 -> 0.073 | Hurt |
| use fixed tau=0.01 instead of test-time calibration (ablation) | FRAP (CLIP) · Several datasets | ECE | Mixed |
| add confidence-weighted fusion of foundation and base predictions (ablation) | FRAP · ImageNet natural shift | Delta SAS vs base model: improvement in 100.0% of experiments, mean 0.249 | Improved |
| add confidence-weighted fusion over CLIP alone (ablation) | FRAP · ImageNet natural shift | Delta SAS vs CLIP: improvement 98.6%, mean 0.486 | Improved |
| add confidence-weighted fusion under synthetic corruption (ablation) | FRAP · ImageNet synthetic shift | Delta SAS vs base model: improvement 64.4%, mean 0.102 | Improved |
| use FRAP fusion instead of CLIP pseudo-labels as ground truth (ablation) | FRAP (CLIP) · Average over 18 datasets | MAE (%): 10.32 -> 6.53 | Improved |
| use FRAP fusion instead of SigLIP pseudo-labels as ground truth (ablation) | FRAP (SigLIP) · Average over 18 datasets | MAE (%): 9.21 -> 7.32 | Improved |
| use fixed-temperature calibration instead of TTC with SigLIP (ablation) | FRAP (SigLIP) · Average over 18 datasets | MAE (%): 7.32 -> 5.65 | Hurt |
Closest papers by meaning
- Efficient Inference under Label Shift in Unsupervised Domain Adaptation (JMLR 2026 · similarity 0.87)
- Unraveling Batch Normalization for Realistic Test-Time Adaptation (AAAI 2024 · similarity 0.86)
- Domain-wise Data Acquisition to Improve Performance under Distribution Shift (ICML 2024 · similarity 0.85)
- AETTA: Label-Free Accuracy Estimation for Test-Time Adaptation (CVPR 2024 · similarity 0.85)
- Test-Time Adaptation Induces Stronger Accuracy and Agreement-on-the-Line (NeurIPS 2024 · similarity 0.85)
- Optimal Aggregation of Prediction Intervals under Unsupervised Domain Shift (NeurIPS 2024 · similarity 0.84)
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.
