Research map / Vision-language and multimodal / CLIP and zero-shot
One Last Attention for Your Vision-Language Model
Liang Chen, Ghazi Shazan Ahmad, Tianjun Yao, Lingqiao Liu, Zhiqiang Shen
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 mask regularization L_reg to RAda (ablation) | RAda (EFT) · Base-to-new, 11 datasets average | new-class accuracy: 71.58 -> 74.14 | Improved |
| use attention layer instead of MLP for the mask function (ablation) | RAda (EFT) · Base-to-new, 11 datasets average | harmonic mean: 75.28 -> 77.94 | Improved |
| add image and text features as extra queries to the rational matrix query (ablation) | RAda attention layer · Base-to-new, 11 datasets average | harmonic mean: 77.72 -> 77.94 | Improved |
| adapt the fused rational matrix instead of attaching attention to text and visual encoders (ablation) | RAda (EFT) · Base-to-new, 11 datasets average | new-class accuracy: 72.46 -> 74.14 | Improved |
| use L2 instead of L1 mask regularization (ablation) | RAda (EFT) · Base-to-new, 11 datasets average | harmonic mean: 77.21 -> 77.94 | Improved |
| use L2 instead of L-infinity mask regularization (ablation) | RAda (EFT) · Base-to-new, 11 datasets average | harmonic mean: 76.84 -> 77.94 | Improved |
| stack 2 attention layers instead of 1 (ablation) | RAda (EFT) · Base-to-new, 11 datasets average | harmonic mean: 77.94 -> 77.93 | No clear effect |
| stack 3 attention layers instead of 1 (ablation) | RAda (EFT) · Base-to-new, 11 datasets average | new-class accuracy: 74.14 -> 71.81 | Hurt |
| fine-tune all parameters after rational adapter warmup (RAda-FT vs FT) | FT on CLIP · ImageNet OOD average (FFT) | accuracy: 53.8 -> 61.3 | Improved |
| replace CLIP-Adapter with RAda-FT | FFT with sequential training · ImageNet OOD average (FFT) | accuracy: 60.3 -> 61.3 | Improved |
| add RAda to CLIP in test-time training | CLIP · ImageNet OOD average (TTT) | accuracy: 57.20 -> 63.50 | Improved |
| add RAda to CLIP with 16-shot base-class training | CLIP · Base-to-new, 11 datasets average | base accuracy: 69.34 -> 82.16 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MaPLe — MaPLe: Multi-Modal Prompt Learning (CVPR 2023) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- VladVA: Discriminative Fine-tuning of LVLMs (CVPR 2025 · similarity 0.88)
- Adaptive Parameter Selection for Tuning Vision-Language Models (CVPR 2025 · similarity 0.88)
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.88)
- A Closer Look at the Few-Shot Adaptation of Large Vision-Language Models (CVPR 2024 · similarity 0.88)
- Realistic Test-Time Adaptation of Vision-Language Models (CVPR 2025 · similarity 0.88)
- Connecting the Dots: Collaborative Fine-tuning for Black-Box Vision-Language Models (ICML 2024 · 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.
