GeoFitBench-LLM is a benchmark that focuses on eliciting data requirements for geospatial analytic tasks. It measures whether large language models can infer the needed datasets, supporting resources, and constraints from natural‑language requests. The benchmark is built around two reasoning dimensions—Spatio‑Temporal Relational and Semantic Grounding—capturing long‑standing challenges in geospatial information retrieval. Ground‑truth requirements are encoded in a fixed, provider‑independent schema, enabling reproducible evaluation of resource identification and attribute‑level specification. Experiments on seven models with minimal prompting show that identifying required resources is relatively achievable; GPT-5.4 Thinking recovers the complete set of resources in 66.67% of benchmark queries, achieving the best performance. However, accurately specifying fitness‑for‑use becomes harder as inferential demands increase.
Review: This benchmark addresses a critical gap in GeoAI evaluation and offers a concrete metric for guiding future improvements in LLM‑based geospatial reasoning.