Research map / Video / Video-language understanding
Perceptual Projection Pruning: Diversity-Aware Video Token Pruning for Multimodal Large Language Models
Zhuangqiu Huang, Minxin Lai, Shuo Liu, Yu Zhang, Jiaqi Wang
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 |
|---|---|---|---|
| cluster on 3D saliency instead of high-dim raw features (ablation) | P3 (adaptive) · 4 video benchmarks avg, 25% retention | average delta: +0.10 for raw features | No clear effect |
| add dynamic saliency (SM) (ablation) | P3 · 4 video benchmarks avg, 25% retention | average delta: -1.40 when removed | Improved |
| add regional saliency (SR) (ablation) | P3 · 4 video benchmarks avg, 25% retention | average delta: -0.75 when removed | Improved |
| add focal saliency (SF) (ablation) | P3 · 4 video benchmarks avg, 25% retention | average delta: -0.43 when removed | Improved |
| use K-Means instead of DPC-KNN (ablation) | P3 · 4 video benchmarks avg, 25% retention | average: 62.1 -> 64.3 | Improved |
| uniform temporal step 4 instead of step 2 (ablation) | P3 · 4 video benchmarks avg, 25% retention | average: 64.2 -> 65.2 | Improved |
| adaptive instead of uniform partitioning on multi-turn (ablation) | P3 · MT-Video-Bench | accuracy: 62.5 -> 63.0 | Improved |
| vary cluster count K from 5 to 80 (ablation) | P3 · 4 video benchmarks | accuracy | No clear effect |
| query-agnostic P3 vs query-aware CDPruner at 25% retention | P3 · Qwen3-VL-8B 4 video benchmarks | avg retained (%): 98.1 -> 97.3 | Hurt |
| P3 on query-dependent temporal localization | P3 · Charades-STA | score: 47.9 (CDPruner) -> 44.8 | Hurt |
| P3 on spatial reasoning | P3 · VSI-Bench | accuracy: 55.4 (vanilla) -> 59.5 | Improved |
| P3 for 1024-frame inputs | P3 · 1024-frame, 25% retention | total latency (s): 19.57s prefill -> 2.22s total | Improved |
Closest papers by meaning
- Beyond Attention or Similarity: Maximizing Conditional Diversity for Token Pruning in MLLMs (NeurIPS 2025 · similarity 0.92)
- An Empirical Study of Attention and Diversity for Adaptive Visual Token Pruning in Large Vision-Language Models (ICLR 2026 · similarity 0.92)
- DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models (CVPR 2025 · similarity 0.91)
- FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.91)
- TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model (CVPR 2025 · similarity 0.90)
- Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization (NeurIPS 2025 · similarity 0.90)
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.
