Research map / Vision-language and multimodal / Multimodal LLMs
Combating Textual Noise and Redundancy: Entropy-Aware Dense Visual Token Pruning
Xuehui Wang, Xuankun Yang, Wei Shen
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 global guidance to dense guidance (mu 0.0 -> 0.5) (ablation) | EADP · LLaVA-1.5-7B, 128 tokens, 5-benchmark avg | average accuracy: 67.7 -> 68.5 | Improved |
| add spatial smoothing (ablation) | EADP · LLaVA-1.5-7B, 128 tokens, 5-benchmark avg | average accuracy: 67.9 -> 68.5 | Improved |
| add score polarization (beta 0.0 -> 2.0) (ablation) | EADP · LLaVA-1.5-7B, 128 tokens, 5-benchmark avg | average accuracy: 67.8 -> 68.2 | Improved |
| discard high-entropy text tokens (q 0.9 -> 0.3) (ablation) | EADP · LLaVA-1.5-7B, 128 tokens, 5-benchmark avg | average accuracy: 67.6 -> 68.6 | Improved |
| use proposed aggregation instead of simple mean (ablation) | EADP · LLaVA-1.5-7B, 128 tokens, 5-benchmark avg | average accuracy: 68.0 -> 68.6 | Improved |
| replace entropy with central mass ratio or variance to measure dispersion (ablation) | EADP · LLaVA-1.5-7B, 128 tokens, 5-benchmark avg | average accuracy: 68.5 -> 68.4 (B), 68.5 (C) | No clear effect |
| increase smoothing kernel from 3 to 5 or 7 (ablation) | EADP · LLaVA-1.5-7B, 128 tokens, 5-benchmark avg | average accuracy: 68.5 -> 68.5 (size=5), 68.4 (size=7) | No clear effect |
| EADP vs DivPrune at 32 tokens | EADP · LLaVA-1.5-7B, 32 tokens | average accuracy: 58.4 -> 60.9 | Improved |
| EADP vs CDPruner on Qwen3-VL-8B at 256 tokens | EADP · Qwen3-VL-8B DocVQA, 256 tokens | accuracy: 53.0 -> 62.8 | Improved |
| EADP vs DivPrune on Qwen2.5-VL at 512 tokens | EADP · Qwen2.5-VL-7B, 512 tokens | average score: 78.1 -> 78.0 | No clear effect |
| EADP vs DivPrune at 64x16 video tokens | EADP · LLaVA-Video-7B, 64x16 tokens | average accuracy: 54.0 -> 54.8 | Improved |
| EADP overhead vs HiPrune at 128 tokens | EADP · LLaVA-1.5-7B, 128 tokens | prefill time (ms): 99.4 -> 118.8 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DivPrune — DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models (CVPR 2025) · better on 7 of 8 shared results
- VisionZip — VisionZip: Longer is Better but Not Necessary in Vision Language Models (CVPR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Beyond Attention or Similarity: Maximizing Conditional Diversity for Token Pruning in MLLMs (NeurIPS 2025 · similarity 0.92)
- When Token Pruning is Worse than Random: Understanding Visual Token Information in VLLMs (CVPR 2026 · similarity 0.91)
- Prune Redundancy, Preserve Essence: Vision Token Compression in VLMs via Synergistic Importance-Diversity (ICLR 2026 · similarity 0.91)
- An Empirical Study of Attention and Diversity for Adaptive Visual Token Pruning in Large Vision-Language Models (ICLR 2026 · 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)
- Hi-Lo Prune: Look at What You'll Lose before Pruning with Hierarchical Token Selection (CVPR 2026 · 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.
