Research map / Vision-language and multimodal / Multimodal LLMs
Spectral Evolution-Guided Token Pruning in Large Multimodal Models
Bin Chen, Yuxiang Cai, Yadan Luo, Yi Zhang, Jianwei Yin, Zhi Chen
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 |
|---|---|---|---|
| keep lowest-evolution tokens instead of highest (CLSE-Inertia) (ablation) | CLSE · LLaVA-1.5-7B, 192 tokens | normalized score | Hurt |
| add Gaussian high-pass cutoff (r=0.1) vs r=0 (ablation) | CLSE · LLaVA-1.5-7B, 192 tokens | normalized score | Improved |
| score tokens by cross-layer spectral evolution instead of attention (ablation) | CLSE · LLaVA-1.5-7B, 192 tokens | normalized score | Improved |
| combine CLSE with attention weights (CLSE-Hybrid) (ablation) | CLSE · LLaVA-1.5-7B, 192 tokens | normalized score | No clear effect |
| use spectral score instead of spatial feature residual (SFR) (ablation) | CLSE · LLaVA-1.5-7B, 192 tokens | normalized score | Improved |
| use spectral score instead of mean magnitude difference (MMD) (ablation) | CLSE · LLaVA-1.5-7B, 192 tokens | normalized score | Improved |
| add token merging (recall and cluster pruned tokens) on video (ablation) | CLSE · Video QA avg accuracy | accuracy: 35.7 -> 38.7 | Improved |
| use CLSE instead of FastV at 64 tokens | FastV · LLaVA-1.5-7B | relative avg: 78.0% -> 94.8% | Improved |
| use CLSE instead of FastV on video | FastV · Video-LLaVA-7B | accuracy: 26.1 -> 35.7 | Improved |
| use CLSE rather than FastV for speed | FastV · LLaVA-1.5-7B | prefilling speedup: 1.69x vs 1.50x | 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.
- SparseVLM — SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference (ICML 2025) · better on 6 of 6 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
- TransPrune: Token Transition Pruning for Efficient Large Vision-Language Model (CVPR 2026 · similarity 0.92)
- Hi-Lo Prune: Look at What You'll Lose before Pruning with Hierarchical Token Selection (CVPR 2026 · similarity 0.92)
- Beyond Attention or Similarity: Maximizing Conditional Diversity for Token Pruning in MLLMs (NeurIPS 2025 · similarity 0.92)
- One Layer's Trash is Another Layer's Treasure: Adaptive Layer-wise Visual Token Selection in LVLMs (CVPR 2026 · similarity 0.91)
- SCOPE: Saliency-Coverage Oriented Token Pruning for Efficient Multimodel LLMs (NeurIPS 2025 · similarity 0.91)
- What Kind of Visual Tokens Do We Need? Training-Free Visual Token Pruning for Multi-Modal Large Language Models from the Perspective of Graph (AAAI 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.
