Research map / Vision-language and multimodal / Cross-modal learning
UPop: Unified and Progressive Pruning for Compressing Vision-Language Transformers
Dachuan Shi, Chaofan Tao, Ying Jin, Zhendong Yang, Chun Yuan, Jiaqi Wang
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 |
|---|---|---|---|
| Unified search across modalities and structures (vs manual per-component ratios) (ablation) | BLIP compression, NLVR2 · NLVR2 2x | Test Acc: +3.88 over mask-based pruning (80.32 vs 76.44) | Improved |
| Unified search instead of manual per-component ratios (ablation) | BLIP compression, NLVR2 · NLVR2 3x | Test Acc: Mask-based fails to converge; unified reaches 71.66 | Improved |
| Progressive pruning added to unified search (ablation) | BLIP compression, NLVR2 · NLVR2 2x | Test Acc: +0.81 (81.13 vs 80.32) | Improved |
| Progressive pruning added to unified search (ablation) | BLIP compression, NLVR2 · NLVR2 3x | Test Acc: unified 71.66 -> UPop 77.61 (+5.95) | Improved |
| Progressive pruning added to unified search (ablation) | BLIP compression, NLVR2 · NLVR2 4x-10x | convergence: UPop converges at 4x, 5x, 10x; unified and mask-based fail at 4x | Improved |
| Progressive pruning (search with no retraining) (ablation) | BLIP 2x, NLVR2 · NLVR2 no retrain | convergence: UPop reaches 77.84 test; unified and mask-based do not converge | Improved |
| Progressive pruning with one retrain epoch (ablation) | BLIP 2x, NLVR2 · NLVR2 one retrain epoch | Test Acc: UPop 80.08 vs unified 75.30 vs mask-based 63.35 | Improved |
| Adaptive unified ratio assignment vs manual assignment (ablation) | BLIP 2x, NLVR2 · NLVR2 2x | Test Acc: adaptive 81.13 vs manual range 61.31 to 80.01 | Improved |
| Unified + progressive UPop vs mask-based pruning | BLIP retrieval COCO and captioning/VQA · COCO retrieval, Flickr30K retrieval, VQA, Caption 2x | R@1 / accuracy / CIDEr: e.g. COCO I->T R@1 77.4 vs 61.7; VQA 76.3 vs 71.6 | Improved |
| Unified + progressive vs mask-based at 4x | BLIP retrieval COCO · COCO 4x | R@1 (Image->Text): UPop 62.9 vs mask-based not converged | Improved |
| Progressive pruning added to unified search (ablation) | BLIP retrieval COCO · COCO 4x | R@1 (Image->Text): 40.3 -> 62.9 | Improved |
| Replace progressive pruning with iterative pruning with STE (ablation) | BLIP NLVR2 2x · NLVR2 2x | Test Acc: 80.54 (iterative STE) vs 81.13 (progressive) | Hurt |
Closest papers by meaning
- Collaborative Multi-Mode Pruning for Vision-Language Models (CVPR 2026 · similarity 0.89)
- MADTP: Multimodal Alignment-Guided Dynamic Token Pruning for Accelerating Vision-Language Transformer (CVPR 2024 · similarity 0.87)
- ECoFLaP: Efficient Coarse-to-Fine Layer-Wise Pruning for Vision-Language Models (ICLR 2024 · similarity 0.86)
- MoPE-CLIP: Structured Pruning for Efficient Vision-Language Models with Module-wise Pruning Error Metric (CVPR 2024 · similarity 0.86)
- MULTIFLOW: Shifting Towards Task-Agnostic Vision-Language Pruning (CVPR 2024 · similarity 0.85)
- Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization (NeurIPS 2025 · similarity 0.85)
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.
