Research map / Vision-language and multimodal / Multimodal LLMs
SWIFT: Spatial-Window Integrated Frequency-aware Token Pruning for Efficient MLLMs on Edge Devices
Guanglai Liu, Jubo Chen, Xiaosheng Yu
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 |
|---|---|---|---|
| add spatial-window integration (SWI) (ablation) | SWIFT on Qwen2.5VL-3B · POPE, 75% reduction | accuracy: 70.9 -> 84.7 | Improved |
| add SWI (ablation) | SWIFT on Qwen2.5VL-3B · Nocaps, 75% reduction | CIDEr: 68.6 -> 100.4 | Improved |
| add SWI (ablation) | SWIFT on Qwen2.5VL-3B · Flickr30k, 75% reduction | CIDEr: 63.3 -> 80.0 | Improved |
| add frequency-aware indicator (FAI) (ablation) | SWIFT on Qwen2.5VL-3B · Nocaps, 75% reduction | CIDEr: 94.8 -> 100.4 | Improved |
| add FAI (ablation) | SWIFT on Qwen2.5VL-3B · MME, 75% reduction | MME score: 1795 -> 1843 | Improved |
| add attention guidance (AG) (ablation) | SWIFT on Qwen2.5VL-3B · POPE, 75% reduction | accuracy: 73.7 -> 84.7 | Improved |
| add SWI (ablation) | SWIFT on Qwen2.5VL-3B · POPE, 90% pruned (text) | accuracy: 60.94 -> 75.27 | Improved |
| add FAI (ablation) | SWIFT on Qwen2.5VL-3B · Nocaps, 90% pruned (text) | CIDEr: 84.75 -> 89.13 | Improved |
| add AG (ablation) | SWIFT on Qwen2.5VL-3B · POPE, 90% pruned (text) | accuracy: 70.72 -> 75.27 | Improved |
| hybrid weight lambda 0.5 versus extremes 0 or 1 (ablation) | SWIFT · MME across models | MME score | Improved |
| replace attention-based pruning (FastV) with SWIFT | LLaVA-1.5-7B · Nocaps, 90% pruned | CIDEr: 74.7 -> 89.1 | Improved |
| replace FastV with SWIFT | LLaVA-1.5-7B · POPE, 90% pruned | accuracy: 66.3 -> 75.3 | Improved |
Closest papers by meaning
- VLM-Pruner: Buffering for Spatial Sparsity in an Efficient VLM Centrifugal Token Pruning Paradigm (CVPR 2026 · similarity 0.90)
- TwinPrune: Density-Aware Two-Phase Token Pruning for Vision-Language Models (NeurIPS 2026 · similarity 0.89)
- Beyond Attention or Similarity: Maximizing Conditional Diversity for Token Pruning in MLLMs (NeurIPS 2025 · similarity 0.89)
- TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model (CVPR 2025 · similarity 0.89)
- Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization (NeurIPS 2025 · similarity 0.89)
- ATP-LLaVA: Adaptive Token Pruning for Large Vision Language Models (CVPR 2025 · similarity 0.88)
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.
