Research map / Label-efficient and robust learning / Test-time adaptation
IMSE: Intrinsic Mixture of Spectral Experts Fine-tuning for Test-Time Adaptation
Sunghyun Baek, Jaemyung Yu, Seunghee Koh, Minsu Kim, Hyeonseong Jeon, Junmo Kim
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 diversity maximization loss to entropy minimization (ablation) | IMSE (TTA) · ImageNet-C 50k, ViT-B | accuracy (%): 67.8 -> 69.1 | Improved |
| Use diversity loss alone without entropy minimization (ablation) | IMSE (TTA) · ImageNet-C 50k, ViT-B | accuracy (%): 69.1 -> 32.7 | Hurt |
| Add diversity maximization loss under non-iid label imbalance (ablation) | IMSE · ImageNet-C non-iid, alpha=0.05 | accuracy (%): 67.33 -> 68.41 | Improved |
| Add diversity maximization loss (ablation) | IMSE-Retrieval (CTTA) · ImageNet-C 5k CTTA | accuracy (%): 59.4 -> 62.2 | Improved |
| Add domain bank retrieval (ablation) | IMSE-Retrieval (CTTA) · ImageNet-C 5k CTTA | accuracy (%): 59.4 -> 62.8 | Improved |
| Combine diversity loss and domain bank (ablation) | IMSE-Retrieval (CTTA) · ImageNet-C 5k CTTA | accuracy (%): 62.2 -> 64.4 | Improved |
| Use mean+variance descriptor instead of mean or variance alone (ablation) | IMSE-Retrieval · ImageNet-C 5k CTTA | accuracy (%): 63.1 / 63.2 -> 64.4 | Improved |
| Tune top-R% instead of bottom-R% singular values (R=20) (ablation) | IMSE · ImageNet-C TTA | accuracy (%): 64.7 -> 67.7 | Improved |
| Tune top vs bottom 80% singular values (ablation) | IMSE · ImageNet-C TTA | accuracy (%): 68.4 -> 69.1 | Improved |
| Adapt all linear layers instead of q,k,v,proj,fc2 (ablation) | IMSE · ImageNet-C TTA | accuracy (%): 69.0 -> 69.2 | Improved |
| Raise domain-change threshold tau from 0.02 to 0.04 (ablation) | IMSE-Retrieval · ImageNet-C Gradual CTTA | accuracy (%): 74.9 -> 75.6 (detection F1 0.77 -> 0.28) | Mixed |
| Scale diversity loss weight lambda_dm from 50 up to 200 (ablation) | IMSE · ImageNet-C TTA | accuracy (%): 69.0 -> ~68 | No clear effect |
Closest papers by meaning
- PALM: Pushing Adaptive Learning Rate Mechanisms for Continual Test-Time Adaptation (AAAI 2025 · similarity 0.86)
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.85)
- pMoE: Prompting Diverse Experts Together Wins More in Visual Adaptation (ICLR 2025 · similarity 0.85)
- Test-Time Spectrum-Aware Latent Steering for Zero-Shot Generalization in Vision-Language Models (NeurIPS 2025 · similarity 0.85)
- Free on the Fly: Enhancing Flexibility in Test-Time Adaptation with Online EM (CVPR 2025 · similarity 0.85)
- DOTA: Distributional Test-time Adaptation of Vision-Language Models (NeurIPS 2025 · similarity 0.85)
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.
