Alibaba researchers, led by lead author and AI specialist Yuxiang Ji, have developed "Thinking with Map," an innovative approach to image geolocalization. Unlike traditional models that rely solely on internal weights or basic text searches, this framework employs an "agent-in-the-map" reasoning loop to navigate and analyze geographic data. By mimicking human cognitive processes, the system bridges the gap between visual input and spatial context, allowing for more nuanced location identification.

The framework utilizes a two-stage optimization process to ensure high precision. It combines agentic reinforcement learning, which refines tool-use efficiency, with parallel test-time scaling that explores multiple geographic candidates simultaneously. To facilitate the training and evaluation of these capabilities, the research team also released MAPBench. This comprehensive benchmark contains a diverse set of real-world images designed to test complex geographic reasoning in unpredictable environments.

Performance results indicate that this map-augmented agent represents a significant leap forward in the field. Within a 500-meter radius, the system achieved a 22.1% accuracy rate, more than doubling the 8.0% performance of Gemini-3-Pro when using grounded search modes. These findings suggest that explicit map interaction and Chain-of-Thought reasoning are critical for reducing hallucinations. This advancement highlights the potential for vision-language models to excel in demanding visual-spatial tasks that require high-resolution accuracy.

Alibaba researchers, led by lead author and AI specialist Yuxiang Ji, have developed "Thinking with Map," an innovative approach to image geolocalization. Unlike traditional models that rely solely on internal weights or basic text searches, this framework employs an "agent-in-the-map" reasoning loop to navigate and analyze geographic data. By mimicking human cognitive processes, the system bridges the gap between visual input and spatial context, allowing for more nuanced location identification.

The framework utilizes a two-stage optimization process to ensure high precision. It combines agentic reinforcement learning, which refines tool-use efficiency, with parallel test-time scaling that explores multiple geographic candidates simultaneously. To facilitate the training and evaluation of these capabilities, the research team also released MAPBench. This comprehensive benchmark contains a diverse set of real-world images designed to test complex geographic reasoning in unpredictable environments.

Performance results indicate that this map-augmented agent represents a significant leap forward in the field. Within a 500-meter radius, the system achieved a 22.1% accuracy rate, more than doubling the 8.0% performance of Gemini-3-Pro when using grounded search modes. These findings suggest that explicit map interaction and Chain-of-Thought reasoning are critical for reducing hallucinations. This advancement highlights the potential for vision-language models to excel in demanding visual-spatial tasks that require high-resolution accuracy.