Thinking with Map: Reinforced Parallel Map-Augemented Agent for Geolocalization (amap-ml.github.io)

A groundbreaking approach to image geolocalization has been introduced with the "Thinking with Map" method, which integrates map usage into the model's reasoning process. Unlike existing large vision-language models (LVLMs) that primarily rely on world knowledge and agentic reasoning, this new model employs a unique agent-in-the-map loop method. It enhances the model's capabilities through a two-stage optimization strategy involving agentic reinforcement learning (RL) to boost sampling efficiency, followed by parallel test-time scaling (TTS) that allows for the exploration of multiple candidate paths before settling on a final location prediction. The significance of this innovation lies in its substantial performance improvement over previous models, evidenced by a notable increase in accuracy from 8.0% to 22.1% at a 500-meter radius, specifically outperforming the Gemini-3-Pro model. Additionally, the introduction of MAPBench—a robust benchmark designed exclusively with real-world images—provides a solid framework for evaluating geolocalization methods in diverse environments. This advancement not only underscores the potential of merging traditional map usage with modern AI techniques but also paves the way for more accurate location predictions in various applications, from autonomous navigation to augmented reality.