Recent advances in vision-language models have made image-based geolocation increasingly accurate, but existing benchmarks are too coarse and linguistically skewed to capture real-world privacy risks. To address this, the authors introduce KoreaGEO, a fine-grained, multimodal benchmark built on Korean street views. It spans four socio-spatial clusters and nine place types, with contextual annotations and two captioning styles designed to simulate how people actually post on social media.

Using a three-path evaluation protocol—image-only, functional captions, and high-risk captions—the researchers tested mainstream VLMs. Their central finding is that input modality has a stronger influence on localization precision and privacy exposure than model scale or architecture. High-risk captions, in particular, substantially boost accuracy, suggesting that text metadata can act as a powerful geolocation cue.

The results also reveal structural prediction biases toward core cities, meaning models are more reliable in urban centers than elsewhere. This uneven performance raises concerns about both location privacy and the fairness of automated geolocation. By building a privacy-aware benchmark with realistic captioning, the paper provides a clearer picture of how multimodal data can expose where people are.