Research map / Vision-language and multimodal / Multimodal LLMs
Structural Pruning of Large Vision Language Models: A Comprehensive Study on Pruning Dynamics, Recovery, and Data Efficiency
Yiran Huang, Lukas Thede, Massimiliano Mancini, Wenjia Xu, Zeynep Akata
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 |
|---|---|---|---|
| use widthwise instead of layerwise pruning (zero-shot) (ablation) | LLaVA-v1.5-7B at 15% pruning · 9-benchmark average | AVG-%: 72.22% -> 90.87% | Improved |
| use widthwise instead of layerwise pruning (zero-shot) (ablation) | Bunny-v1-3B at 15% pruning · AI2D | accuracy: 52.73 -> 65.64 | Improved |
| use widthwise instead of layerwise pruning on free-form tasks (ablation) | Bunny-v1-3B at 15% pruning · GQA, DocVQA, MathVista | relative performance: 2.2% -> 31.5% | Improved |
| use layerwise instead of widthwise pruning after recovery training (ablation) | pruned LVLMs at <=30% ratio · 9-benchmark average | AVG | Improved |
| use widthwise instead of layerwise pruning after recovery training (ablation) | pruned LVLMs at >=45% ratio · 9-benchmark average | AVG | Improved |
| use layerwise instead of widthwise pruning for latency (ablation) | LLaVA at 30% compression · inference latency | latency (ms): 84.4 -> 80.7 | Improved |
| use layerwise instead of widthwise pruning for FLOPs (ablation) | LLaVA at 30% compression · FLOPs | FLOPs (T): 6.89 -> 6.92 | No clear effect |
| compare multimodal vs text-only degradation at mild pruning | LLaVA-v1.5-7B widthwise 15% · multimodal vs text-only inputs | relative degradation: 9.13% -> 1.07% | Mixed |
| finetune only the multimodal projector (ablation) | pruned LVLMs at 15% ratio · 9-benchmark average | AVG-%: over 95% | Improved |
| additionally finetune the pruned LLM with LoRA (ablation) | pruned LVLMs at 45% ratio · 9-benchmark average | AVG-%: over 80% | Improved |
| replace SFT with standalone KD (KL, RKL or L2) (ablation) | pruned LVLMs at high ratios · 9-benchmark average | AVG | Hurt |
| add L2 hidden-state distillation to SFT (ablation) | pruned LVLMs · 9-benchmark average | AVG | Improved |
Closest papers by meaning
- Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization (NeurIPS 2025 · similarity 0.94)
- Unified Knowledge Maintenance Pruning and Progressive Recovery with Weight Recalling for Large Vision-Language Models (AAAI 2025 · similarity 0.93)
- ECoFLaP: Efficient Coarse-to-Fine Layer-Wise Pruning for Vision-Language Models (ICLR 2024 · similarity 0.93)
- ATP-LLaVA: Adaptive Token Pruning for Large Vision Language Models (CVPR 2025 · similarity 0.93)
- Collaborative Multi-Mode Pruning for Vision-Language Models (CVPR 2026 · similarity 0.93)
- ViTCoP: Accelerating Large Vision-Language Models via Visual and Textual Semantic Collaborative Pruning (AAAI 2026 · similarity 0.92)
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.
