Research map / Label-efficient and robust learning / Domain adaptation and generalization
Addressing Exacerbated Attention Sink for Source-Free Cross-Domain Few-Shot Learning
Shuai Yi, Yixiong Zou, Yuhua Li, Ruixuan Li
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 TIR token re-weighting | CLIP-LoRA-Vision · 5-way 5-shot CDFSL average | accuracy: 65.99 -> 68.44 | Improved |
| add TIR token re-weighting | CLIP-LoRA-Vision · 5-way 1-shot CDFSL average | accuracy: 55.92 -> 58.20 | Improved |
| add TIR token re-weighting | CLIP-LoRA-Vision · ISIC (5-way 5-shot) | accuracy: 51.10 -> 56.73 | Improved |
| add TIR token re-weighting | CLIP-LoRA-Vision · ChestX (5-way 1-shot) | accuracy: 21.73 -> 23.98 | Improved |
| add suppression of sink tokens (Sum=5 weight 0) (ablation) | CLIP-LoRA-Vision + TIR · 5-way 5-shot average | accuracy: 65.99 -> 67.66 | Improved |
| add enhancement of discriminative tokens (Sum=1) (ablation) | CLIP-LoRA-Vision + TIR · 5-way 5-shot average | accuracy: 65.99 -> 67.23 | Improved |
| combine inhibit and enhance (ablation) | TIR · 5-way 5-shot average | accuracy: 67.66 -> 68.44 | Improved |
| add TIR to SigLIP2 | SigLIP2 · 5-shot average | accuracy: 62.14 -> 65.40 | Improved |
| add TIR to PE-Core | PE-Core · 1-shot average | accuracy: 56.73 -> 57.89 | Improved |
| add TIR to MaPLe | CLIP + MaPLe · CDFSL average | accuracy: 65.20 -> 67.49 | Improved |
| add Registers instead of TIR | CLIP-LoRA-Vision · 5-way 5-shot average | accuracy: 65.99 -> 65.33 | Hurt |
| add SPARC | CLIP-LoRA-Vision · 5-way 5-shot average | accuracy: 65.99 -> 49.12 | Hurt |
Closest papers by meaning
- Attention Temperature Matters in ViT-Based Cross-Domain Few-Shot Learning (NeurIPS 2024 · similarity 0.91)
- Random Registers for Cross-Domain Few-Shot Learning (ICML 2025 · similarity 0.90)
- A Closer Look at the CLS Token for Cross-Domain Few-Shot Learning (NeurIPS 2024 · similarity 0.90)
- Revisiting Continuity of Image Tokens for Cross-domain Few-shot Learning (ICML 2025 · similarity 0.89)
- Flatten Long-Range Loss Landscapes for Cross-Domain Few-Shot Learning (CVPR 2024 · similarity 0.88)
- Reconstruction Target Matters in Masked Image Modeling for Cross-Domain Few-Shot Learning (AAAI 2025 · similarity 0.88)
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.
