Research map / Vision-language and multimodal / CLIP and zero-shot
Regularized Mask Tuning: Uncovering Hidden Knowledge in Pre-Trained Vision-Language Models
Kecheng Zheng, Wei Wu, Ruili Feng, Kai Zhu, Jiawei Liu, Deli Zhao, Zheng-Jun Zha, Wei Chen, Yujun 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 |
|---|---|---|---|
| Learn binary masks on the attention layers of frozen CLIP (R-AMT) instead of zero-shot use | CLIP ViT-B/16 · 11 datasets, 16-shot average | accuracy (%): 65.23 -> 83.96 | Improved |
| Add R-AMT mask to CoOP | CoOP · 11 datasets, 16-shot average | accuracy (%): 79.90 -> 83.16 | Improved |
| Add R-AMT mask to TIP-Adapter | TIP-Adapter · 11 datasets, 16-shot average | accuracy (%): 81.24 -> 84.37 | Improved |
| Add gradient dropout regularity to mask tuning (ablation) | AMT · Base-to-new, new classes | accuracy (%): 69.11 -> 72.15 | Improved |
| Add KL loss directly to the CE loss instead of gradient dropout (ablation) | AMT · ImageNet 16-shot | accuracy (%): 72.60 -> 71.92 | Hurt |
| Use AgreeGrad gradient surgery with AMT (ablation) | AMT · ImageNet 16-shot | accuracy (%): 72.60 -> 68.82 | Hurt |
| Use ProGrad gradient projection with AMT (ablation) | AMT · ImageNet 16-shot | accuracy (%): 72.60 -> 72.70 | Improved |
| Lower leak parameter l from 1.0 to 0.1 (ablation) | R-AMT · ImageNet 16-shot | accuracy (%): 73.07 -> 72.67 | Hurt |
| Mask only attention layers instead of all layers (ablation) | R-PMT · 11 datasets, 16-shot average | accuracy (%): 83.96 vs 83.96, storage 19M -> 6.7M | No clear effect |
| Mask attention layers instead of MLP layers (ablation) | R-MMT · 11 datasets, 16-shot average | accuracy (%): 83.39 -> 83.96 | Improved |
| Set hard threshold alpha to 5e-3 rather than 4e-3 or 6e-3 (ablation) | R-AMT · ImageNet 16-shot | accuracy (%): 72.87 / 72.91 -> 73.07 | Improved |
| Fine-tune whole image encoder instead of mask tuning | CLIP image encoder · ImageNet 16-shot | accuracy (%): 66.73 (zero-shot) -> 64.51 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- TIP-Adapter — Tip-Adapter: Training-Free Adaption of CLIP for Few-Shot Classification (ECCV 2022) · better on 2 of 2 shared results
- CLIP-adapter — CLIP-Adapter: Better Vision-Language Models with Feature Adapters (IJCV 2023) · better on 1 of 1 shared result
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
- Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves (CVPR 2025 · similarity 0.91)
- Understanding and Enhancing Mask-Based Pretraining towards Universal Representations (NeurIPS 2025 · similarity 0.88)
- Connecting the Dots: Collaborative Fine-tuning for Black-Box Vision-Language Models (ICML 2024 · similarity 0.88)
- Enhancing Large Language Model Performance with Gradient-Based Parameter Selection (AAAI 2025 · similarity 0.88)
- Random Masking Finds Winning Tickets for Parameter Efficient Fine-tuning (ICML 2024 · similarity 0.87)
- ZIP: An Efficient Zeroth-order Prompt Tuning for Black-box Vision-Language Models (ICLR 2025 · similarity 0.87)
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.
