Research map / Vision-language and multimodal / Multimodal LLMs
Prune Once: Retraining-Free Task-Agnostic Pruning for Vision-Language Models
Kang MinSeok, Hyunwoo Kim, Chanyoung Kim, Minwoo Kim, Jaekoo Lee, Dahuin Jung
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 |
|---|---|---|---|
| replace Wanda magnitude score with PORTA variance score (uniform sparsity) (ablation) | PORTA · CLIP 60% sparsity CIFAR-100 | accuracy: 70.28 -> 74.45 | Improved |
| replace Wanda magnitude score with PORTA variance score (uniform sparsity) (ablation) | PORTA · CLIP 60% sparsity ImageNet | accuracy: 57.35 -> 59.35 | Improved |
| add PORTA ratio allocation to PORTA score (ablation) | PORTA · CLIP 60% sparsity CIFAR-100 | accuracy: 74.45 -> 75.59 | Improved |
| add PORTA ratio allocation to PORTA score (ablation) | PORTA · CLIP 60% sparsity ImageNet | accuracy: 59.35 -> 59.66 | Improved |
| add PORTA ratio allocation to Wanda score (ablation) | Wanda · CLIP 60% sparsity CIFAR-100 | accuracy: 70.28 -> 74.29 | Improved |
| use Multiflow ratio allocation with either score (ablation) | PORTA / Wanda · CLIP 60% sparsity CIFAR-100 | accuracy: 74.45 -> 1.40 | Hurt |
| drop zero-mean assumption (ablation) | PORTA · CLIP 60% sparsity retrieval | avg recall: 63.62 -> 63.64 | No clear effect |
| drop diagonal covariance approximation (ablation) | PORTA · CLIP 60% sparsity retrieval | avg recall: 63.62 -> 63.73 | No clear effect |
| PORTA vs Wanda on CLIP at 65% sparsity | PORTA · CLIP 65% sparsity CIFAR-100 | accuracy: 28.86 -> 62.29 | Improved |
| PORTA vs SparseGPT on BLIP Flickr30k at 45% sparsity | PORTA · BLIP 45% sparsity Flickr30k | mean recall: 89.95 -> 90.26 | Improved |
| PORTA vs Wanda on Qwen2-VL ScienceQA at 50% sparsity | PORTA · Qwen2-VL ScienceQA 50% | average accuracy: 53.43 -> 54.30 | Improved |
| PORTA vs SparseGPT time cost | PORTA · CLIP pruning time | seconds: 773.35 -> 195.34 | Improved |
Closest papers by meaning
- MULTIFLOW: Shifting Towards Task-Agnostic Vision-Language Pruning (CVPR 2024 · similarity 0.93)
- ECoFLaP: Efficient Coarse-to-Fine Layer-Wise Pruning for Vision-Language Models (ICLR 2024 · similarity 0.92)
- QuietPrune: Query-Guided Early Token Pruning for Vision-Language Models (CVPR 2026 · similarity 0.91)
- TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model (CVPR 2025 · similarity 0.91)
- IF-Prune: Information-Flow Guided Token Pruning for Efficient Vision-Language Models (CVPR 2026 · similarity 0.91)
- LearnPruner: Rethinking Attention-based Token Pruning in Vision Language Models (ICLR 2026 · similarity 0.91)
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.
