Spatial-Agent: Agentic Geo-spatial Reasoning with Scientific Core Concepts

TL;DR

Spatial-Agent addresses a core gap in geospatial reasoning by grounding geo-analytical questions in spatial information theory and representing workflows as executable GeoFlow Graphs $G=(V,E,\lambda,\rho)$. It introduces a compositional, template-driven approach plus two-stage geographic constraint learning (SFT and DPO) to ensure well-formed graphs that can be executed via GIS tools, yielding verifiable intermediate results. Empirical results on MapEval-API and MapQA show substantial gains over strong baselines across both closed- and open-source LLMs, with notable improvements in routing, trip planning, and POI tasks. The work highlights that the primary bottlenecks are external data quality and API reliability rather than reasoning, underscoring the value of grounded, interpretable spatial workflows for practical, tool-assisted geospatial analytics.

Abstract

Geospatial reasoning is essential for real-world applications such as urban analytics, transportation planning, and disaster response. However, existing LLM-based agents often fail at genuine geospatial computation, relying instead on web search or pattern matching while hallucinating spatial relationships. We present Spatial-Agent, an AI agent grounded in foundational theories of spatial information science. Our approach formalizes geo-analytical question answering as a concept transformation problem, where natural-language questions are parsed into executable workflows represented as GeoFlow Graphs -- directed acyclic graphs with nodes corresponding to spatial concepts and edges representing transformations. Drawing on spatial information theory, Spatial-Agent extracts spatial concepts, assigns functional roles with principled ordering constraints, and composes transformation sequences through template-based generation. Extensive experiments on MapEval-API and MapQA benchmarks demonstrate that Spatial-Agent significantly outperforms existing baselines including ReAct and Reflexion, while producing interpretable and executable geospatial workflows.

Figures (4)

• Figure 1: LLM-intuitive but incorrect workflow (left) vs. correct concept transformation (right). The incorrect applies spatial constraints after aggregation; the correct computes crime rates within spatial context first.
• Figure 2: Overview of Spatial-Agent: (A) Spatial information theory analysis extracts core concepts and assigns functional roles; (B) Concept transformation drafting composes templates from the library; (C) GeoFlow Graph construction produces an ordered and constrained graph; (D) Graph factorization maps to executable tools for execution and answer generation.
• Figure 3: Distribution of error categories in SpatialAgent. Data Quality Issues (45.6%) and Search Result Mismatch (33.8%) account for the majority of errors, both occurring during the execution stage.
• Figure 4: Average latency per query (seconds) across task types. All methods use GPT-4o-mini.

Theorems & Definitions (2)

• Definition 1: Core Spatial Concept Space
• Definition 2: Functional Role Set