GeoExperts: Scaling Global Geocoding with Region-Aware LoRA Experts

Geocoding means converting addresses or natural-language location descriptions into geographic coordinates and supports applications such as map search, route planning, and logistics. However, existing APIs are costly and impose substantial deployment overhead, while current models struggle to combine global coverage with robustness to complex inputs. To address these issues, we propose GeoExperts, an efficient global geocoding framework that requires no external geographic resources at inference time. To accommodate diverse address systems, GeoExperts adopts an expert specialization strategy. It divides the world into regions based on address systems and trains specialized experts on a shared lightweight language model. For address inputs, an address classification expert routes each request to the appropriate geocoding expert, which maps structured addresses directly to latitude and longitude. For natural-language inputs, a semantic understanding expert converts text into normalized addresses, enabling unified support for standard addresses, address variants, and natural-language location queries. We also introduce GeoMosaic, the first large-scale dataset designed for LLM-based geocoding without external resources, and GeoMosaic-Bench, which contains 40K samples and provides a unified evaluation of direct geocoding, input robustness, and semantic location reasoning. On structured address evaluation, GeoExperts reduces the base model’s median error from 49.85 km to 1.44 km. Its accuracy approaches that of commercial APIs at the 5 km threshold and exceeds it at larger distance thresholds. The whole dataset and code will be public.