GeoReason

<p align="center"> <h1 align="center">GEOREASON: ALIGNING THINKING AND ANSWERING IN REMOTE SENSING VISION-LANGUAGE MODELS VIA LOGICAL CONSISTENCY REINFORCEMENT LEARNING</h1> <p align="center"> Wenshuai Li</a>&emsp; Xiantai Xiang</a>&emsp; Zixiao Wen</a>&emsp; Guangyao Zhou</a>&emsp; Ben Niu</a>&emsp; Feng Wang</a>&emsp; Lijia Huang</a>&emsp; Qiantong Wang</a>&emsp; Yuxin Hu</a>&emsp; <div align="center"> <a href='http://arxiv.org/abs/2601.04118'><img src='https://img.shields.io/badge/arXiv-2601.04118-brown.svg?logo=arxiv&logoColor=white'></a> <a href='https://huggingface.co/WenshuaiLi/GeoReason'><img src='https://img.shields.io/badge/HuggingFace-Model-yellow.svg?logo=HuggingFace&logoColor=white'></a> <a href='https://huggingface.co/datasets/WenshuaiLi/GeoReason-Bench'><img src='https://img.shields.io/badge/HuggingFace-Datasets-yellow.svg?logo=HuggingFace&logoColor=white'></a> </div> <p align='center'> If you find our work helpful, please consider giving us a ⭐! </p> </p> </p>

---

GeoReason: Overview

GeoReason is a framework designed for Remote Sensing Vision-Language Models (RS-VLMs) to address "logical hallucinations" and "pseudo-reasoning," where models derive correct answers from flawed logic or shortcuts. By introducing a logic-driven dataset (GeoReason-Bench) and employing a consistency-aware reinforcement learning strategy with a novel "Logical Consistency Reward," it compels the model to strictly anchor its final decisions in verifiable reasoning traces, ensuring both accuracy and cognitive reliability.

<p align="center"><img src="assets/pipeline.png" width="80%"></p>

---

Performance

Eevaluating models across Perceptual Tasks (Count, Color, Shape, Scene) and Reasoning Tasks (Reason) to analyze their multi-level understanding.

<p align="center"><img src="assets/result.png" width="80%"></p>

---

Get Started

Environment Installation

conda create -n GeoReason python=3.10
conda activate GeoReason
pip install -r requirements.txt

Infer with GeoReason

You can use GeoReason_infer.py to generate the answers to the questions.

python GeoReason_infer.py --model_path /path/to/model --dataset /path/to/dataset --image_path /path/to/image_path

Citation

@misc{li2026georeasonaligningthinkinganswering, title={GeoReason: Aligning Thinking And Answering In Remote Sensing Vision-Language Models Via Logical Consistency Reinforcement Learning}, author={Wenshuai Li and Xiantai Xiang and Zixiao Wen and Guangyao Zhou and Ben Niu and Feng Wang and Lijia Huang and Qiantong Wang and Yuxin Hu}, year={2026}, eprint={2601.04118}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2601.04118}, }