Research map / Vision-language and multimodal / Multimodal LLMs
NaVLM-PVC: Progressive Visual Compression for Efficient Native-Resolution Encoding in MLLMs
Shichu Sun, Yichen Zhang, Haolin Song, Zonghao Guo, Chi Chen, Yidan Zhang, Yuan Yao, Zhiyuan Liu, Maosong Sun
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 global native-resolution encoding instead of slice-based encoding (ablation) | SigLIP2 + Qwen2-7B MLLM · general benchmarks avg | average accuracy: 67.5% -> 69.6% | Improved |
| use global native-resolution encoding instead of slice-based encoding (ablation) | SigLIP2 + Qwen2-7B MLLM · ShapeGrid spatial perception | average accuracy: 62.6% -> 73.6% | Improved |
| replace avg-pooling WTC with 2 stage compression (256 tokens) vs 1024 tokens baseline (ablation) | ViT-UHD (SigLIP2) · Module ablation | TTFT / OCR&Chart: TTFT 233 -> 82 ms; avg 62.1 -> 59.2 | Mixed |
| use content-adaptive pooling instead of plain average pooling in WTC (ablation) | ViT-UHD (SigLIP2) · Module ablation | average accuracy: 59.2 -> 60.7 | Improved |
| use pixel-unshuffle as WTC at early layers instead of pooling (ablation) | ViT-UHD (SigLIP2) · Module ablation | average accuracy: 60.7 -> 57.5 | Hurt |
| add refined patch embedding (patch 16 to 8) with 3 WTC (ablation) | ViT-UHD (SigLIP2) · Module ablation | average accuracy: 62.1 -> 63.0 | Improved |
| add more than 3 WTC layers or move their positions (ablation) | LLaVA-UHD v3 · TTFT latency | TTFT | No clear effect |
| integrate PVC (RPE + WTC) into SigLIP2 backbone (ablation) | SigLIP2 · PVC transfer | average accuracy: 61.94 -> 64.71 | Improved |
| integrate PVC into MoonViT backbone (ablation) | MoonViT · PVC transfer | average accuracy: 66.17 -> 66.33 | No clear effect |
| apply same pre-alignment to MoonViT and ViT-UHD (PVC) (ablation) | ViT-UHD · 8-benchmark average, max tokens 776 vs 380 | average accuracy: 64.11 -> 64.26 | No clear effect |
Closest papers by meaning
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.90)
- DUET-VLM: Dual stage Unified Efficient Token reduction for VLM Training and Inference (CVPR 2026 · similarity 0.90)
- When MLLMs Meets Compression Distortion: A Coding Paradigm Tailored to MLLMs (ICLR 2026 · similarity 0.90)
- Hybrid Token Compression for Vision-Language Models (CVPR 2026 · similarity 0.90)
- PVC: Progressive Visual Token Compression for Unified Image and Video Processing in Large Vision-Language Models (CVPR 2025 · similarity 0.90)
- VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration (ICLR 2025 · 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.
