Research map / Vision-language and multimodal / Multimodal LLMs
Instruction-Guided Cross-Modal Clustering for Training-Free Visual Token Pruning in Vision-Language Models
Yunqian Yu, Biao Chen, Yunya Zhang, Tonglan Xie, Mengmeng Jing, Lin Zuo
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a training-free visual token pruning method for vision-language models that uses instruction-guided cross-modal clustering to remove redundant visual tokens and cut inference cost.
Problem
LVLMs have far more visual tokens than text tokens, causing redundancy and high inference cost; existing pruning relies on unimodal information or cross-modal attention, which have drawbacks.
Approach
Combines instruction (text) guidance with clustering over visual tokens in a cross-modal space to select representative visual tokens for pruning, without any training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CLS-attention token selection over random selection (ablation) | ICCTP pruning (LLaVA-1.5-7B, 88.9% pruning) · TextVQA | accuracy: +5.0% | Improved |
| add instruction-noun clustering on top of CLS-only selection (ablation) | ICCTP pruning (LLaVA-1.5-7B, 88.9% pruning) · POPE | F1: +5.0% | Improved |
| add instruction-noun clustering on top of CLS-only selection (ablation) | ICCTP pruning (LLaVA-1.5-7B, 88.9% pruning) · MME | score: +65.6 | Improved |
| vary global retention ratio rho_g among 0.4, 0.6, 0.8 (ablation) | ICCTP (LLaVA-1.5-7B, 88.9% pruning) · ten benchmarks | accuracy | No clear effect |
| vary semantic-context factor alpha among 0.2, 0.3, 0.4 (ablation) | ICCTP (LLaVA-1.5-7B, 88.9% pruning) · ten benchmarks | accuracy | No clear effect |
| ICCTP vs VisionZip at 94.4% pruning | LLaVA-1.5-7B · POPE | accuracy: 68.7 -> 80.4 | Improved |
| ICCTP vs VisionZip at 94.4% pruning | LLaVA-NeXT-7B · TextVQA | accuracy: 54.7 -> 45.1 | Hurt |
| ICCTP vs DART at 94.4% pruning | LLaVA-1.5-7B · MM-Vet / MMBench-CN | score: MMBCN 47.1 vs 50.0; MMVet 22.2 vs 25.0 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VisionZip — VisionZip: Longer is Better but Not Necessary in Vision Language Models (CVPR 2025) · better on 10 of 12 shared results
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)
- Mostly Text, Smart Visuals: Asymmetric Text-Visual Pruning for Large Vision-Language Models (CVPR 2026 · similarity 0.92)
- QuietPrune: Query-Guided Early Token Pruning for Vision-Language Models (CVPR 2026 · similarity 0.92)
- Lifting the Veil on Visual Information Flow in MLLMs: Unlocking Pathways to Faster Inference (CVPR 2025 · similarity 0.92)
- Rethinking Visual Token Reduction in LVLMs Under Cross-Modal Misalignment (AAAI 2026 · similarity 0.92)
- Don't Just Chase “Highlighted Tokens” in MLLMs: Revisiting Visual Holistic Context Retention (NeurIPS 2025 · similarity 0.91)
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.
