GeoEyes: On-Demand Visual Focusing for Evidence-Grounded Understanding of Ultra-High-Resolution Remote Sensing Imagery

Fengxiang Wang¹, Mingshuo Chen², Yueying Li¹, Yajie Yang³, Yifan Zhang^4*, Long Lan¹
Xue Yang⁵, Hongda Sun^6*, Yulin Wang⁷, Di Wang⁸, Jing Zhang⁸, Jun Song^*, Bo Du⁸

¹National University of Defense Technology, ²Beijing University of Posts and Telecommunications
³University of the Chinese Academy of Sciences, ⁴Chinese Academy of Science
⁵Shanghai Jiao Tong University, ⁶Renmin University of China, ⁷Tsinghua University, ⁸Wuhan University

📚 Contents

• 📚Contents
• 🔍Overview
• 🌐UHR-CoZ Dataset
• 🛠️Methodology & Training
• 🚀Evaluation
• 🔗Citation
• 🤝Acknowledgement

🔍Overview

Fig 1. Overview of the AdaZoom-GRPO Framework.

We introduce GeoEyes, a specialized MLLM for Ultra-High-Resolution (UHR) Remote Sensing. Current "thinking-with-images" models suffer from Tool Usage Homogenization—collapsing into rigid, one-size-fits-all zooming patterns that fail to address the task heterogeneity and low evidence density of UHR imagery.

To solve this, we propose a staged training framework:

1. Cold-Start SFT: Initializing the model with UHR-CoZ, a dataset containing diverse "Chain-of-Zoom" trajectories (Global, Single-Zoom, Multi-Step).
2. AdaZoom-GRPO: An Agentic Reinforcement Learning stage with a novel reward system designed to incentivize on-demand zooming and progressive focusing.

Our method achieves 54.23% accuracy on XLRS-Bench, establishing a new state-of-the-art by outperforming larger models like Qwen2.5-VL-72B and domain-specific agents like DeepEyes.

🌐UHR-CoZ Dataset

We construct UHR Chain-of-Zoom (UHR-CoZ), the first large-scale interleaved image-text chain-of-thought dataset specifically for UHR remote sensing. It is built using an automated agentic pipeline (Fig 2) involving GLM-4.5V, which generates multi-round zoom-in trajectories cleaned by a semantic scorer.

Fig 2. Automated data construction pipeline for UHR-CoZ.

Dataset Statistics

Statistics	Value
Total Samples	25,467
Avg. Image Resolution	2,178 × 2,051
Zoom-in Depth 1 (No Zoom)	6.4%
Zoom-in Depth 2	86.7%
Zoom-in Depth $\ge 3$	6.9%
Avg. Reasoning Length	157.8 tokens

🛠️Methodology & Training

Our approach builds upon the DeepEyes framework, introducing a two-stage optimization process.

1. Prepare Data

• UHR-CoZ: Download our constructed SFT dataset with interleaved zoom trajectories through huggingface.
• SuperRS-VQA: Used during the RL stage to enhance task diversity which is included in UHR-CoZ.
• General RL Data: We utilize DeepEyes-47K for general reasoning stability.

2. Training Stages

The code base is develeped using torch2.6/2.8+cu128 and Python3.10/3.11.

Stage 1: Cold-Start SFT

We perform Supervised Fine-Tuning on UHR-CoZ to initialize the policy with basic tool capabilities and stop-conditions.

# 1. Download and prepare sft data from huggingface
# please make sure to modify the absolute image paths in UHR-CoZ.json
# 2. SFT using llamafactory
# We use this specific commit: https://github.com/hiyouga/LlamaFactory/tree/2a822178dea4d1c05f595521dd883a8e4f4e2e77
# if encountered TypeError during dataset preprocess, refer to https://github.com/hiyouga/LlamaFactory/issues/5613
# modify json paths in dataset_info.json and yaml file
llamafactory-cli train config.yaml

Stage 2: AdaZoom-GRPO (Agentic RL)

We optimize the model using Group Relative Policy Optimization (GRPO) with our specific reward formulation:

• Adaptive Efficiency Reward (Penalizes redundant tools on easy tasks).
• Chain-of-Focus Reward (Geometric containment reward for progressive zoom).
• Necessity-Aware Process Verification (LLM-based judge for logical rigor).

# 1. first install DeepEyes following https://github.com/Visual-Agent/DeepEyes
# we also provided a clean requirements.txt without torch package
# 2. download RL data, and modify parquet file paths in the training script/yaml file
# there are 3 parquets from DeepEyes-47k and 1 parquet file from UHR-CoZ HF repo
# 3. follow deepeyes to set LLM judge and start training using
# export LLM_AS_A_JUDGE_BASE="http://{IP}:{PORT}/v1"
python -m verl.trainer.main_ppo \
    --config-path DeepEyes/config \
    --config-name deepeyes_coz

🚀Evaluation

We evaluate on XLRS-Bench, focusing on Perception (e.g., Counting, Object Classification) and Reasoning (e.g., Route Planning, Anomaly Detection) tasks.

Running Evaluation

# 0. execute the prepare_xlrs_data.ipynb to preprocess the evaluation data
# 1. convert model from pt format to hf model
bash s1.sh
# 2. deploy model using vllm (or ray using `serve run ray.yaml`)
bash s21.sh
# 3. prompting vllm
bash s22.sh
# 4. calculate metrics
bash s232.sh

Main Results (XLRS-Bench)

🔗Citation

If you find our work helpful, please consider citing:

@article{wang2026geoeyes,
  title={GeoEyes: On-Demand Visual Focusing for Evidence-Grounded Understanding of Ultra-High-Resolution Remote Sensing Imagery},
  author={Wang, Fengxiang and Chen, Mingshuo and Li, Yueying and Yang, Yajie and Zhang, Yifan and Lan, Long and Yang, Xue and Sun, Hongda and Wang, Yulin and Wang, Di and others},
  journal={arXiv preprint arXiv:2602.14201},
  year={2026}
}

🤝Acknowledgement

This repo benefits from DeepEyes and LLaMA-Factory. Thanks for their wonderful works.