GeoMTVR: Learning Multi-Tool Visual Reasoning for Ultra-High-Resolution Remote Sensing Imagery

Forum topic · 小凯 · 2026-07-30

Summary

This paper introduces GeoMTVR, a large-scale geospatial multi-tool visual reasoning dataset for ultra-high-resolution (UHR) remote sensing imagery, and GeoLens, a multimodal large language model (MLLM) trained on it. The authors show via a pilot study on XLRS-Bench that equipping MLLMs with a single zoom-in tool only partially helps: it resolves easy and medium-level tasks with locally recoverable evidence but saturates on hard cases requiring global search, multi-region comparison, path planning, or dispersed-evidence reasoning. GeoMTVR addresses this with 13K UHR VQA samples featuring interleaved reasoning traces, diverse visual tool calls, and returned visual observations, teaching models question decomposition, tool selection, region inspection, object-level grounding, auxiliary visual reasoning, and cross-tool evidence integration. Beyond supervised fine-tuning, the authors propose a tool-attention-focused reinforcement learning algorithm that concentrates optimization on key tool-use decisions: when to invoke a tool, which tool to choose, where to apply it, and how to interpret its output. Experiments show GeoLens consistently outperforms direct inference and single-tool zoom baselines in accuracy, evidence grounding, and tool-use trajectory quality.

Paper Overview

Field: Computer Vision (CV) Authors: Fengxiang Wang, Jiangnan Huang, Mingshuo Chen, Yueying Li, Yang Shi, Junwei Luo, Haoyu Wang, Yansheng Li, Jing Zhang, Haiyan Zhao, Wenjing Yang arXiv: 2607.25993

Ultra-high-resolution (UHR) remote-sensing (RS) imagery provides fine-grained Earth-observation evidence over city-scale scenes, but poses a fundamental challenge for multimodal large language models (MLLMs): task-relevant evidence is often sparse, local, and spatially dispersed across extremely large visual contexts.

A natural solution is to equip MLLMs with zoom-in tools for active local inspection. However, through a pilot study on XLRS-Bench, the authors find that zoom-in is only partially effective: it resolves easy and medium-level tasks with locally recoverable evidence, but saturates on hard cases requiring global search, multi-region comparison, path planning, or dispersed-evidence reasoning.

Key Contributions

GeoMTVR dataset: A large-scale geospatial multi-tool visual reasoning dataset built from wide-area satellite imagery, containing 13K UHR VQA samples with interleaved reasoning traces, diverse visual tool calls, and returned visual observations. It enables models to learn question decomposition, tool selection, region inspection, object-level grounding, auxiliary visual reasoning, and cross-tool evidence integration.
Tool-attention-focused RL: Beyond supervised fine-tuning (SFT), the authors propose a reinforcement learning algorithm that concentrates optimization on critical tool-use decisions — when to invoke a tool, which tool to choose, where to apply it, and how to interpret tool outputs.
GeoLens: A multi-tool visual reasoning MLLM for UHR remote sensing, developed by combining SFT on GeoMTVR with the proposed RL algorithm.

Results

Experiments demonstrate that GeoLens consistently outperforms direct inference and single-tool zoom-in baselines, achieving stronger accuracy, better evidence grounding, and more effective tool-use trajectories.