Research map / Vision-language and multimodal / CLIP and zero-shot
Task Residual for Tuning Vision-Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 a prior-independent task residual to the frozen base classifier (ablation) | CLIP ResNet-50 · 11-dataset average, 16-shot | accuracy: 58.96 -> 75.10 | Helped |
| Add a prior-independent task residual to an enhanced base classifier (ablation) | CLIP ResNet-50 · 11-dataset average, 16-shot | accuracy: 73.09 -> 75.78 | Helped |
| Replace the prior-independent residual with an adapter dependent on the base features (ablation) | TaskRes classifier · ImageNet 1-shot | accuracy: 61.43 -> 61.03 | Hurt |
| Adapt the base classifier directly rather than preserving it as a frozen component (ablation) | CLIP classifier · ImageNet 1-shot | accuracy: 60.33 -> 28.03 | Hurt |
| Apply TaskRes to the image branch rather than the text branch (ablation) | CLIP ResNet-50 · ImageNet 16-shot | accuracy: 64.75 -> 61.60 | Hurt |
| Apply TaskRes to both image and text branches instead of only the text branch (ablation) | CLIP ResNet-50 · ImageNet 16-shot | accuracy: 64.75 -> 64.67 | Hurt |
| Learn the residual scaling factor rather than fixing it at 0.5 (ablation) | TaskRes · 11-dataset average, 1-shot | accuracy: 63.97 -> 64.01 | Helped |
| Use the enhanced base classifier instead of the regular base classifier alongside TaskRes (ablation) | TaskRes · 11-dataset average, 16-shot | accuracy: 75.10 -> 75.78 | Helped |
| Use an enhanced base classifier to improve source-task performance (ablation) | TaskRes · ImageNet-V2, trained on ImageNet 16-shot; ResNet-50 | accuracy: 56.47 -> 57.00 | Hurt |
| Add a task residual to CoOp-style transfer on ImageNet 1-/2-shot | CLIP ResNet-50 · ImageNet 1-/2-shot | accuracy: CoOp underperforms Zero-shot CLIP by 1.03%/0.37% | Helped |
| Increase the residual scaling factor from 0 to 0.1 (ablation) | TaskRes · 11-dataset average, 1-shot | accuracy: 58.96 -> 61.86 | Helped |
| Use TaskRes on OxfordPets and Food101 with very few labeled examples | CLIP ResNet-50 · OxfordPets and Food101, 1-/2-/4-/8-shot | accuracy | Hurt |
Closest papers by meaning
- Vision-Language Model Selection and Reuse for Downstream Adaptation (ICML 2025 · similarity 0.90)
- Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves (CVPR 2025 · similarity 0.90)
- TextRefiner: Internal Visual Feature as Efficient Refiner for Vision-Language Models Prompt Tuning (AAAI 2025 · similarity 0.90)
- Connecting the Dots: Collaborative Fine-tuning for Black-Box Vision-Language Models (ICML 2024 · similarity 0.89)
- Prompt Learning via Meta-Regularization (CVPR 2024 · similarity 0.89)
- Learning to Learn Better Visual Prompts (AAAI 2024 · similarity 0.89)
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.