Research map / Vision-language and multimodal / CLIP and zero-shot
DeAR: Fine-Grained VLM Adaptation by Decomposing Attention Head Roles
Yiming Ma, Hongkun Yang, Lionel Z. Wang, Bin Chen, Weizhi Xian, Jianzhi Teng
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 |
|---|---|---|---|
| Role-based attention mask (vs shielding all deep heads) (ablation) | DeAR · Base-to-novel, 11-dataset average | Harmonic mean: 80.68 -> 82.72 | Improved |
| Role-based attention mask (vs allowing all interactions) (ablation) | DeAR · Base-to-novel, 11-dataset average | Harmonic mean: 80.93 -> 82.72 | Improved |
| Add self-regularization loss Lreg (ablation) | DeAR (CE only) · Base-to-novel, 11-dataset average | Novel accuracy: 74.21 -> 78.45 | Improved |
| Add self-regularization loss Lreg (ablation) | DeAR (CE only) · Base-to-novel, 11-dataset average | Base accuracy: 86.75 -> 85.03 | Hurt |
| Add fusion-weight regularization Lfusion (ablation) | DeAR (CE+Lreg) · Base-to-novel, 11-dataset average | Harmonic mean: 81.61 -> 82.72 | Improved |
| Use DeAR instead of MMRL | CLIP ViT-B/16 · Base-to-novel, 11-dataset average | Harmonic mean: 81.20 -> 82.72 | Improved |
| Use DeAR instead of MMRL | CLIP ViT-B/16 · Cross-dataset (avg of 10 targets) | Accuracy: 67.25 -> 67.60 | Improved |
| Use DeAR instead of TAC | CLIP ViT-B/16 · Domain generalization (average of 4 variants) | Accuracy: 61.63 -> 61.25 | Hurt |
| Use DeAR instead of TAC | CLIP ViT-B/16 · Domain generalization (ImageNet-R) | Accuracy: 78.50 -> 78.83 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- MMRL — MMRL: Multi-Modal Representation Learning for Vision-Language Models (CVPR 2025) · better on 5 of 5 shared results
- PromptKD — PromptKD: Unsupervised Prompt Distillation for Vision-Language Models (CVPR 2024) · better on 2 of 2 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
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.90)
- Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves (CVPR 2025 · similarity 0.89)
- MMA: Multi-Modal Adapter for Vision-Language Models (CVPR 2024 · similarity 0.89)
- Enhancing Target-unspecific Tasks through a Features Matrix (ICML 2025 · similarity 0.89)
- A Stitch in Time Saves Nine: Small VLM is a Precise Guidance for Accelerating Large VLMs (CVPR 2025 · similarity 0.89)
- Vision-Language Models Create Cross-Modal Task Representations (ICML 2025 · 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.
