The image geolocalization task involves predicting the location an image was taken using visual clues. Traditional large vision-language models (LVLMs) focus on world knowledge and reasoning but often miss a key human strategy: using maps. The authors introduce a method called "Thinking with Map," which incorporates a unique agent-in-the-map loop. This approach uses a two-stage optimization process that includes agentic reinforcement learning (RL) and parallel test-time scaling (TTS). The RL enhances the model's ability to sample efficiently, while TTS allows it to evaluate multiple potential paths before making a final prediction, significantly improving the accuracy of geolocalization. To test their method, the researchers developed MAPBench, a benchmark comprising real-world images for training and evaluation. This dataset provides a comprehensive look at geolocalization challenges, especially within China. It categorizes images into two difficulty levels through a voting process involving advanced models like GPT-3 and GPT-5. The results show that the Thinking with Map method outperforms existing models, particularly by improving accuracy from 8.0% to 22.1% within a 500-meter range compared to the Gemini-3-Pro model. The authors detail the mechanics of their approach, emphasizing the agent's maintenance of a candidate pool of hypotheses during the geolocalization process. They also illustrate how reinforcement learning and parallel test-time scaling work together to optimize performance. Overall, their research illustrates a significant advancement in leveraging map-based reasoning for more effective image geolocation, showcasing a practical application of machine learning in real-world scenarios.