Thinking with Map: Reinforced Parallel Map-Augmented Agent for Geolocalization

Yuxiang Ji^1,2  Yong Wang^2†  Ziyu Ma²  Yiming Hu²  Hailang Huang² 

Xuecai Hu²  Guanhua Chen³  Liaoni Wu¹  Xiangxiang Chu² 

¹Xiamen University    ²AMAP, Alibaba Group    ³Southern University of Science and Technology

^†Project lead    

Page | Arxiv | 🤗 Paper | 🤗 Data | 🤗 Model

Data | Model | X@AK

[!NOTE] This project includes the codebase, datasets and chckpoints for Thinking with Map: a map-augmented agent for geolocalization. Given an image in-the-wild, the agent can conduct reasoning with map to inference the location.

🎬 Demo

The illustration of a complete Thinking with Map process.

Comparison with open- and closed-source models.

News

• [Feb 3, 2026]: 🛠️ Our code and data are realeased now.
• [Jan 12, 2026]: 🔥 We are honored to be featured as HuggingFace Daily Paper #1.
• [Jan 12, 2026]: 📍 Our paper is released on ArXiv and HuggingFace.

Table of contents

• Dataset Access
• Model Zoo
• Quick Start
• Acknowledgement
• Citation

💾 Dataset Access

We release two versions of the dataset. - V1 contains the training and test data used in the paper (lower resolution, will be deperecated). - We also provide V2 with higher-resolution images (recommended): the released training set includes the subset of 6k samples, and the test set is the same size as V1 (~2.5k samples).

V1 (Low Resolution, Deperecated)	V2 (High Resolution, Recommended)
🤗HuggingFace	🤗HuggingFace
ModelScope	ModelScope

📦 Model Zoo

Qwen3-VL-30B-A3B RL-tuned on MAPBench-V2 37k
🤗HuggingFace
ModelScope

🚀 Quick Start

After downloading the dataset, process it into the parquet format.

cd verl/examples/data_preprocess
bash preprocess_thinking_with_map.sh

Installation

If you just want to try the demo without training/evaluating, you can skip the installation part and try Inference part directly.

Please refer to VeRL Installation for more details.

## VeRL Installation 
# use conda for environment management
conda create -n mapagent python==3.12
conda activate mapagent
# install torch and vllm
pip install torch==2.8.0
pip install vllm==0.11.0
# install sglang and other basic packages
cd Thinking-with-Map/verl/scripts
bash install_vllm_sglang_mcore.sh
# install verl
cd Thinking-with-Map/verl
pip install --no-deps -e .

The tool server can share the same environment with VeRL.

## Tool Sever Installation
pip install playwright
pip install "uvicorn[standard]"
pip install json5
pip install fastapi

Inference

After downloading the model, employ it by vllm server.

# at least 2 GPUs with 80GB memory each for the trained Qwen3-VL-30B-A3B model
# try more GPUs when OOM
vllm serve /path/to/released/model \
    --tensor-parallel-size 2 \
    --port 8002

Then try the demo with demo/cookbook_thinking_with_map.ipynb.

Training

Start the tool server with cache on each cluster node.

cd verl/tool_server
# need redis server running on port 6397 for cache
# change your own api key in the bash script
bash run_api_server.sh $RANK

Simply run distributed RL training on each cluster node.

cd verl/geoagent_scripts
# training on MAPBench
bash train_thinking_with_map.sh

The default training is conduct on 8 nodes with 8 GPUs each, while you can change it by modifying N_NODES, train_batch_size and ppo_mini_batch_size respectively to adapt to your own environment.

Evaluation

The evaluation code is also conducted on VeRL.

cd verl/geoagent_scripts
bash test_thinking_with_map.sh

🙏 Acknowledgement

The repo is built upon VeRL, SGLang, vLLM, and Qwen-Agent. We appreciate these open-source communities for their great work.

📌 Citation

@article{ji2026thinking,
  title={Thinking with Map: Reinforced Parallel Map-Augmented Agent for Geolocalization},
  author={Ji, Yuxiang and Wang, Yong and Ma, Ziyu and Hu, Yiming and Huang, Hailang and Hu, Xuecai and Chen, Guanhua and Wu, Liaoni and Chu, Xiangxiang},
  journal={arXiv preprint arXiv:2601.05432},
  year={2026}
}