Research map / Vision-language and multimodal / Multimodal LLMs
InfoScan: Information-Efficient Visual Scanning via Resource-Adaptive Walks
Yifeng Wu, Huimin Huang, Shangjie Zhou, Yawen Huang, Hao Zheng, Yun Chen, Xian Wu, Ruize Han, Guanhua Chen
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 InfoScan, an information-aware, resource-adaptive scanning mechanism for state-space (Mamba-style) visual backbones to replace fixed scan patterns in high-resolution visual representation learning.
Problem
High-resolution visual representation learning is costly due to quadratic ViT complexity, and fixed scanning patterns in Mamba-based vision models hinder content-adaptive perception.
Approach
An information-aware scanning mechanism for state-space visual backbones that dynamically adapts scan walks to image content and resource budget, instead of using fixed scan orders.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add patch selection module (ablation) | InfoScan · ImageNet-1K 512x512 | Top-1 (%): 82.5 -> 83.4 | Improved |
| add path planning module on top of patch selection (ablation) | InfoScan · ImageNet-1K 512x512 | Top-1 (%): 83.4 -> 85.9 | Improved |
| add adaptive revisitation reward (ablation) | InfoScan scanning policy · ImageNet-1K | Top-1 (%): 80.4 -> 81.1 | Improved |
| add exploration bonus to reward (ablation) | InfoScan scanning policy · ImageNet-1K | Top-1 (%): 81.1 -> 84.2 | Improved |
| add neighborhood information gain to reward (ablation) | InfoScan scanning policy · ImageNet-1K | Top-1 (%): 84.2 -> 85.9 | Improved |
| add Shannon entropy to information scoring (ablation) | Information Scoring Module · ImageNet-1K | Top-1 (%): 79.7 -> 81.6 | Improved |
| add local variance to entropy scoring (ablation) | Information Scoring Module · ImageNet-1K | Top-1 (%): 81.6 -> 85.9 | Improved |
| repeat the fixed raster scan three times (ablation) | Raster scan · ImageNet-1K and ADE-20K | Top-1 / mIoU: Top-1 83.5 -> 83.6; ADE-20K 42.9 -> 41.8 | No clear effect |
| repeat the fixed Hilbert scan three times (ablation) | Hilbert scan · ImageNet-1K, ADE-20K, BraTS | Top-1 / mIoU: Top-1 83.6 -> 84.8; BraTS 17.1 -> 16.8 | Mixed |
| randomize the scan start point (ablation) | Single Hilbert scan · ImageNet-1K | Top-1 (%): 83.6 -> 84.5 | Improved |
| randomize the scan start point (ablation) | InfoScan · ImageNet-1K | Top-1 (%): 84.2 -> 85.9 | Improved |
| replace fixed scan with learned InfoScan policy (fixed start) (ablation) | Triple Hilbert Scan · ImageNet-1K and ADE-20K | Top-1 / mIoU: Top-1 84.8 vs 84.2; ADE-20K 44.2 vs 44.6 | Mixed |
Experimental setup
Builds on: Mamba Vision Mamba VMamba
Closest papers by meaning
- CVSearch: Empowering Multimodal LLMs with Cognitive Visual Search for High-Resolution Image Perception (ICML 2026 · similarity 0.87)
- LookThere! Sparse Vision by Reinforced Selection (NeurIPS 2026 · similarity 0.86)
- BVS: Bayesian Visual Search with Multimodal Large Language Model for Fine-grained Perception (ICML 2026 · similarity 0.84)
- Pathways of Visual Information Flow in Vision-Language Models (NeurIPS 2026 · similarity 0.84)
- ExtraVAR: Stage-Aware RoPE Remapping for Resolution Extrapolation in Visual Autoregressive Models (NeurIPS 2026 · similarity 0.84)
- DeepScan: A Training-Free Framework for Visually Grounded Reasoning in Large Vision-Language Models (CVPR 2026 · similarity 0.84)
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.
