Research map / Vision-language and multimodal / Multimodal LLMs
When Large Vision-Language Model Meets Large Remote Sensing Imagery: Coarse-to-Fine Text-Guided Token Pruning
Junwei Luo, Yingying Zhang, Xue Yang, Kang Wu, Qi Zhu, Lei Liang, Jingdong Chen, Yansheng Li
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 RFM text-guided pruning with DIP to LLaVA-1.5 | LLaVA-1.5 · Avg of 4 datasets | Avg accuracy: 27.30 -> 31.59 | Improved |
| add RFM text-guided pruning with DIP to LLaVA-Next | LLaVA-Next-Qwen2 · Avg of 4 datasets | Avg accuracy: 30.80 -> 32.16 | Improved |
| prune 75% of vision tokens by RFM without DIP | LLaVA-Next-p25 · MME-RealWorld-RS | accuracy: 39.65 -> 41.28 | Improved |
| prune 90% instead of 75% of vision tokens (ablation) | w/ prune (Ours) · MME-RealWorld-RS | accuracy: 41.28 -> 39.41 | Hurt |
| use RFM guidance instead of CLIP similarity for pruning (ablation) | anyres-p25 · MME-RealWorld-RS | accuracy: 36.86 -> 41.28 | Improved |
| use RFM guidance instead of RemoteCLIP similarity (ablation) | anyres-p25 · MME-RealWorld-RS | accuracy: 38.77 -> 41.28 | Improved |
| add PruMerge++ pruning (ablation) | anyres-p25 · MME-RealWorld-RS | accuracy: 39.65 -> 34.86 | Hurt |
| add MSE loss on high-res tokens to KL distillation (ablation) | RFM · MME-RealWorld-RS | accuracy: 41.20 -> 41.89 | Improved |
| add KL loss to MSE-only distillation (ablation) | RFM · MME-RealWorld-RS | accuracy: 39.73 -> 41.89 | Improved |
| use dynamic DIP depth instead of fixed 5 layers (ablation) | DIP inference · MME-RealWorld-RS | accuracy: 39.91 -> 41.89 | Improved |
| use select instead of concat of tokens from traversed DIP layers (ablation) | RFM pruning · MME-RealWorld-RS | accuracy: 41.39 -> 41.89 | Improved |
| apply token pruning during SFT training (ablation) | RFM training · MME-RealWorld-RS | accuracy: 41.89 -> 39.09 | 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.
- Geochat* — GeoChat: Grounded Large Vision-Language Model for Remote Sensing (CVPR 2024) · better on 1 of 1 shared result
- LLaVA-UHD-v2* — LLaVA-UHD: an LMM Perceiving any Aspect Ratio and High-Resolution Images (ECCV 2024) · 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
- Quality-Driven Curation of Remote Sensing Vision-Language Data via Learned Scoring Models (NeurIPS 2025 · similarity 0.90)
- GeoLLaVA-8K: Scaling Remote-Sensing Multimodal Large Language Models to 8K Resolution (NeurIPS 2025 · similarity 0.89)
- VRSBench: A Versatile Vision-Language Benchmark Dataset for Remote Sensing Image Understanding (NeurIPS 2024 · similarity 0.89)
- Remote Sensing Vision-Language Foundation Models without Annotations via Ground Remote Alignment (ICLR 2024 · similarity 0.89)
- UHR-BAT: Budget-Aware Token Compression Vision-Language model for Ultra-High-Resolution Remote Sensing (ICML 2026 · similarity 0.89)
- VLM-Pruner: Buffering for Spatial Sparsity in an Efficient VLM Centrifugal Token Pruning Paradigm (CVPR 2026 · similarity 0.89)
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.
