Research map / Architectures and efficiency / Pruning and efficient tuning
Parameter and Computation Efficient Transfer Learning for Vision-Language Pre-trained Models
Qiong Wu, Wei Yu, Yiyi Zhou, Shubin Huang, Xiaoshuai Sun, Rongrong Ji
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 |
|---|---|---|---|
| skip 8 fusion layers instead of 0 with DAS-Fusion (ablation) | DAS-Fusion · VQA test-dev (METER) | accuracy: 75.28 -> 73.70 | Mixed |
| skip 8 fusion layers instead of 0 with DAS-Fusion (ablation) | DAS-Fusion · VQA additional FLOPs (METER) | additional FLOPs: 1.68G -> -24.00G | Improved |
| skip 8 layers instead of 0 with DAS-Global (ablation) | DAS-Global · NLVR2 test-P (METER) | accuracy: 81.28 -> 79.61 | Hurt |
| skip 2 layers instead of 1 with DAS on ViLT (ablation) | DAS · VQA test-dev (ViLT) | accuracy: 69.28 -> 67.64 | Hurt |
| skip 1 layer with DAS on ViLT vs no skipping (ablation) | DAS · VQA test-dev (ViLT) | accuracy: 70.13 -> 69.28 | Hurt |
| skip fusion layers only (DAS-Fusion) instead of whole model (DAS-Global), 4 skipped (ablation) | DAS · VQA additional FLOPs (METER) | additional FLOPs: -4.51G -> -11.16G | Improved |
| RL-based DAS search instead of random skipping (ablation) | DAS · VQA2.0 (METER) | accuracy: +-1.33 deviation for random skipping of 8 layers | Improved |
| DAS4-Fusion instead of full tuning for inference speed | DAS4-Fusion · VQA (METER) | test speed (samples/s): 4.16 -> 4.96 | Improved |
| DAS4-Global instead of full tuning for inference speed | DAS4-Global · VQA (METER) | test speed (samples/s): 4.16 -> 4.57 | Improved |
| Scaled PA instead of full tuning for inference speed | Scaled PA · VQA (METER) | test speed (samples/s): 4.16 -> 3.95 | Hurt |
| DAS4-Fusion instead of Scaled PA | DAS4-Fusion · VQA test-dev (METER) | accuracy: 75.11 -> 74.80 | No clear effect |
| LoRA instead of Adapter | LoRA · VQA test-dev (METER) | accuracy: 74.70 -> 74.00 | Hurt |
Closest papers by meaning
- Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves (CVPR 2025 · similarity 0.89)
- MULTIFLOW: Shifting Towards Task-Agnostic Vision-Language Pruning (CVPR 2024 · similarity 0.88)
- A Closer Look at the Few-Shot Adaptation of Large Vision-Language Models (CVPR 2024 · similarity 0.88)
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.88)
- UniPT: Universal Parallel Tuning for Transfer Learning with Efficient Parameter and Memory (CVPR 2024 · similarity 0.87)
- VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learning (NeurIPS 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.
