Research map / Architectures and efficiency / Attention and transformers
LoGIC: Multi-LoRA Guided Importance Consensus for Multi-Task Pruning in Vision Transformers
Yu-Hong Chou, Rui Fang, Hsi-Wen Chen, Ming-Syan Chen
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 |
|---|---|---|---|
| LoGIC instead of best competing MTT pruning method | ViT-L, 30% sparsity · Five-task average | average score: 0.6037 -> 0.6943 | Improved |
| LoGIC vs LoRAPrune on FFT backbones | ViT-L, 50% sparsity · NYUv2 segmentation | mIoU: 0.6540 -> 0.7130 | Improved |
| LoGIC vs MTT + Magnitude Pruning | ViT-L, 50% sparsity · Oxford Flowers-102 | accuracy: 0.5581 -> 0.9006 | Improved |
| LoGIC vs Swin-L LoRAPrune at 50% sparsity | Swin-L, 50% sparsity · NYUv2 segmentation | mIoU: 0.4782 -> 0.6201 | Improved |
| 30% LoGIC vs dense LoRA on ViT-L | LoGIC ViT-L 30% · NYUv2 segmentation | mIoU: 0.7531 -> 0.7506 | Hurt |
| LoGIC vs dense FFT on detection at 50% sparsity | ViT-L, 50% sparsity · CPPE-5 | mAP: 0.4457 -> 0.4400 | Hurt |
| Pruning 30% to 50% with LoGIC | ViT-L · Five-task average | average score: 0.6943 -> 0.6580 | Hurt |
| LoGIC vs full FFT in parameters | ViT-L · Deployment size | total parameters: 1520M -> 243.6M | Improved |
Closest papers by meaning
- SAViT: Structure-Aware Vision Transformer Pruning via Collaborative Optimization (NeurIPS 2022 · similarity 0.90)
- Collaborative Multi-Mode Pruning for Vision-Language Models (CVPR 2026 · similarity 0.89)
- Transforming Vision Transformer: Towards Efficient Multi-Task Asynchronous Learner (NeurIPS 2024 · similarity 0.88)
- MULTIFLOW: Shifting Towards Task-Agnostic Vision-Language Pruning (CVPR 2024 · similarity 0.87)
- MADTP: Multimodal Alignment-Guided Dynamic Token Pruning for Accelerating Vision-Language Transformer (CVPR 2024 · similarity 0.86)
- Unified Knowledge Maintenance Pruning and Progressive Recovery with Weight Recalling for Large Vision-Language Models (AAAI 2025 · similarity 0.86)
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.
