Research map / Vision-language and multimodal / Multimodal LLMs
Feather the Throttle: Revisiting Visual Token Pruning for Vision-Language Model Acceleration
Mark Endo, Xiaohan Wang, Serena Yeung-Levy
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 |
|---|---|---|---|
| remove RoPE when computing attention pruning criterion (ablation) | FastV criterion · Localization avg (pruning after layer 3) | accuracy: 5.9 -> 16.7 | Improved |
| remove RoPE when computing attention pruning criterion (ablation) | FastV criterion · Localization avg (pruning after layer 8) | accuracy: 23.3 -> 27.3 | Improved |
| remove RoPE when computing attention pruning criterion (ablation) | FastV criterion · Open-ended VQA avg (pruning after layer 3) | accuracy: 54.8 -> 59.0 | Improved |
| prune after layer 8 instead of layer 3 (ablation) | phi_-R · Localization avg (final method comparison) (pruning layers 3 vs 8) | accuracy: 16.7 -> 27.3 | Improved |
| add uniform sampling to RoPE-free attention criterion (ablation) | phi_-R · Localization avg (pruning after layer 3) | accuracy: 16.7 -> 27.2 | Improved |
| add uniform sampling to RoPE-free attention criterion (ablation) | phi_-R · Localization avg (pruning after layer 8) | accuracy: 27.3 -> 35.6 | Improved |
| combine phi_-R with uniform sampling versus uniform alone (ablation) | phi_uniform · Localization avg (pruning after layer 3) | accuracy: 28.0 -> 27.2 | Hurt |
| combine phi_-R with uniform sampling versus uniform alone (ablation) | phi_uniform · Localization avg (pruning after layer 8) | accuracy: 30.3 -> 35.6 | Improved |
| use tokens selected after layer 8 instead of layer 3 for whole LLM (ablation) | phi_-R · OCID-Ref | accuracy: 23.8 -> 26.7 | Improved |
| replace FastV with FeatHer at similar FLOPS reduction | FastV · Localization avg (final method comparison) | accuracy: 5.9 -> 39.3 | Improved |
| replace PyramidDrop with FeatHer at similar FLOPS reduction | PyramidDrop · Localization avg (final method comparison) | accuracy: 28.9 -> 39.3 | Improved |
| shuffle image token positions (ablation) | unpruned baseline · Localization avg | accuracy: 53.2 -> 8.0 | Hurt |
Closest papers by meaning
- Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization (NeurIPS 2025 · similarity 0.94)
- QuietPrune: Query-Guided Early Token Pruning for Vision-Language Models (CVPR 2026 · similarity 0.93)
- TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model (CVPR 2025 · similarity 0.93)
- LearnPruner: Rethinking Attention-based Token Pruning in Vision Language Models (ICLR 2026 · similarity 0.93)
- ATP-LLaVA: Adaptive Token Pruning for Large Vision Language Models (CVPR 2025 · similarity 0.92)
- 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.
