Research map / Vision-language and multimodal / Multimodal LLMs
VFlowOpt: A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization
Sihan Yang, Runsen Xu, Chenhang Cui, Tai Wang, Dahua Lin, Jiangmiao Pang
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 |
|---|---|---|---|
| use VFlowOpt instead of VisionZip at 10% retention | LLaVA-OneVision-7B · 10 image benchmarks avg | relative avg performance (%): 82.1% -> 85.5% | Improved |
| use VFlowOpt at 50% retention | LLaVA-OneVision-7B · 10 image benchmarks avg | relative avg performance (%): 98.9% (VisionZip) -> 99.9% | Improved |
| add importance calibration (attention from relatively important tokens plus entropy) (ablation) | VFlowOpt · MMStar (retain 25%) | accuracy: 56.2 -> 57.8 | Improved |
| add token recycling (ablation) | VFlowOpt · MMStar (retain 25%) | accuracy: 57.6 -> 57.8 | No clear effect |
| add token recycling (ablation) | VFlowOpt · POPE (retain 25%) | accuracy: 86.8 -> 89.1 | Improved |
| add progressive pruning (three stages) (ablation) | VFlowOpt · MMStar (retain 25%) | accuracy: 56.0 -> 57.8 | Improved |
| optimize on last-token representation instead of mean pooling or first token (ablation) | VFlowOpt Bayesian optimization · MMStar (retain 25%) | accuracy: 56.1 (mean pooling), 54.2 (first token) -> 57.8 | Improved |
| choose optimization data from MathV360K-Geometry3K instead of random samples (ablation) | VFlowOpt · MMStar and others | accuracy: 57.8 -> 57.8 | No clear effect |
| prune 90% of visual tokens with VFlowOpt | LLaVA-OneVision-7B · MME efficiency | latency, KV cache, performance: latency 1040.1 -> 272.1 ms; KV cache -89.0%; MME 1581 -> 1464 (-7.4%) | Mixed |
| prune 50% of visual tokens with VFlowOpt | LLaVA-OneVision-7B · MME | score: 1581 -> 1591 (+0.6%) | Improved |
| apply VFlowOpt at 25% retention to video benchmarks | LLaVA-OneVision-7B · SeedBench video and VideoMME | relative avg performance (%): 99.0% (VisionZip) -> 100% | Improved |
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.
- VisionZip — VisionZip: Longer is Better but Not Necessary in Vision Language Models (CVPR 2025) · better on 8 of 8 shared results
- SparseVLM — SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference (ICML 2025) · better on 2 of 2 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
- TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model (CVPR 2025 · similarity 0.93)
- Hi-Lo Prune: Look at What You'll Lose before Pruning with Hierarchical Token Selection (CVPR 2026 · similarity 0.93)
- QuietPrune: Query-Guided Early Token Pruning for Vision-Language Models (CVPR 2026 · similarity 0.93)
- TransPrune: Token Transition Pruning for Efficient Large Vision-Language Model (CVPR 2026 · similarity 0.92)
- HAWK: Head Importance-Aware Visual Token Pruning in Multimodal Models (CVPR 2026 · similarity 0.92)
- Don't Just Chase “Highlighted Tokens” in MLLMs: Revisiting Visual Holistic Context Retention (NeurIPS 2025 · 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.
