Geospatial reasoning underpins critical applications from city planning to emergency response, yet current large language models frequently struggle with accurate spatial computation, often resorting to unreliable web searches or superficial pattern recognition. Riyang Bao from Emory University, Cheng Yang and Zhexiang Tang from Rutgers University, alongside Dazhou Yu, Gengchen Mai and Liang Zhao, introduce Spatial-Agent, a novel framework grounded in established spatial information science. This research formalises geo-analytical question answering as a process of concept transformation, translating natural language into executable workflows called GeoFlow Graphs , essentially, a blueprint for spatial analysis. By rigorously extracting spatial concepts and composing logical transformation sequences, Spatial-Agent demonstrably outperforms existing models like ReAct and Reflexion on benchmarks such as MapEval-API and MapQA, offering not only improved accuracy but also transparent, interpretable geospatial workflows.

GeoFlow Graphs for Geospatial Reasoning represent a powerful

Scientists have unveiled Spatial-Agent, a novel AI agent designed to perform genuine geospatial reasoning, addressing a critical limitation in current large language model (LLM)-based systems. Existing agents frequently rely on web searches or pattern matching, often hallucinating spatial relationships instead of conducting accurate geospatial computation, but this new approach is fundamentally different. The research team formalised geo-analytical question answering as a concept transformation problem, translating natural language questions into executable workflows represented as GeoFlow Graphs, directed acyclic graphs where nodes signify spatial concepts and edges denote transformations. This innovative framework draws directly from spatial information theory, enabling the agent to extract spatial concepts, assign functional roles with principled ordering, and compose transformation sequences using template-based generation.
Spatial-Agent establishes a crucial bridge between linguistic understanding and computational GIS, moving beyond merely descriptive answers to verifiable, operational analyses. The agent meticulously identifies core spatial entities, objects, events, fields, and networks, and then selects appropriate spatial operators such as buffering, overlay, routing, and aggregation, ordering them into a coherent, executable workflow. This process isn’t simply about recognising terms; it’s about understanding the geometric, topological, and spatiotemporal relationships inherent in geographical phenomena, something previous LLM-based agents have struggled to achieve. By grounding abstract instructions in concrete GIS tools like PostGIS, ArcGIS, and QGIS, Spatial-Agent delivers results rooted in computational verification rather than relying solely on parametric knowledge.

Extensive experiments conducted on the MapEval-API and MapQA benchmarks demonstrate that Spatial-Agent significantly outperforms existing baselines, including ReAct and Reflexion. The study reveals a substantial improvement in correctness, interpretability, and the generation of executable geospatial workflows, effectively bridging the gap between natural language reasoning and computational GIS. Specifically, the agent extracts spatial concepts and instantiates them as nodes within the GeoFlow Graph, then identifies functional roles to impose ordering constraints, and finally composes transformation edges through a template-based approach that captures recurring geo-analytical patterns. This compositional approach, leveraging macro-templates, enhances structural validity through template matching and IO-port composition, ensuring a robust and reliable workflow.

The work opens exciting possibilities for applications in urban analytics, transportation planning, environmental monitoring, disaster response, and public health. By accurately interpreting and executing complex geo-analytical questions, Spatial-Agent promises to empower users with natural-language interfaces capable of handling sophisticated spatial data analysis. This breakthrough not only addresses the limitations of current LLM-based agents but also establishes a new paradigm for geospatial AI, grounded in foundational theories of spatial information science and focused on delivering verifiable, executable results, a significant step towards truly intelligent geospatial systems.

GeoFlow Graphs for Spatial Question Answering

Scientists developed Spatial-, a novel approach to geospatial reasoning grounded in spatial information science, to address limitations in current large language model (LLM) based systems. The research team formalised geo-analytical question answering as a concept transformation problem, enabling the parsing of natural language questions into executable workflows represented as GeoFlow Graphs, directed acyclic graphs where nodes signify spatial concepts and edges denote transformations. This innovative method extracts spatial concepts and assigns functional roles with principled ordering constraints, composing transformation sequences via template-based generation, fundamentally shifting how LLMs approach spatial computation. Experiments employed two key benchmarks: MapEval-API, spanning 180 cities across 54 countries and encompassing Place Info, Nearby, Routing, and Trip tasks, and MapQA, an open-domain geospatial QA dataset with 3,154 question-answer pairs sourced from OpenStreetMap data covering Southern California and Illinois.

The study pioneered a rigorous comparative analysis against Direct LLM, ReAct, Reflexion, and Plan-and-Solve, utilising diverse LLM backbones including GPT-4o-mini, GPT-5, Qwen2.5-72B-Instruct, Qwen2.5-32B-Instruct, and LLaMA-70B. Researchers meticulously measured performance across task categories, recording overall accuracy and task-specific improvements to quantify the effectiveness of Spatial-. The team achieved an overall accuracy of 64.51% with Spatial- and GPT-4o-mini on MapEval-API, representing a 96.30% relative improvement over the baseline of 23.00%. Notably, performance gains were particularly pronounced on Place Info (+149.91%) and Nearby (+133.26%) tasks, demonstrating the system’s strength in structured tool invocation and spatial reasoning.

When utilising GPT-5, Spatial- reached a peak overall accuracy of 71.88%, with exceptional results on Routing (75.76%) and Trip (77.61%) tasks, which demand complex multi-step planning. Further analysis revealed that Spatial- (GPT-4o-mini) attained 61.45% overall accuracy on MapQA, significantly surpassing Direct LLM (13.55%), ReAct (43.79%), and Reflexion (53.79%). Error analysis, conducted on 68 incorrect predictions, identified Data Quality Issues (45.6%) and Search Result Mismatch (33.8%) as primary limitations, occurring during the execution stage, while confirming that no errors originated from the GeoFlow Graph construction itself. Latency measurements, using GPT-4o-mini, showed Spatial- achieving competitive speeds, 7.5 seconds on Routing, and comparable performance to ReAct on Nearby (8.3s vs 7.6s) and Trip (10.4s vs 12.3s).

GeoFlow Graphs enable language-driven geospatial workflows

Scientists have developed Spatial-, a novel approach to geospatial reasoning grounded in spatial information science. The research team formalised geo-analytical question answering as a concept transformation problem, successfully parsing natural language into executable workflows represented as GeoFlow Graphs, directed acyclic graphs defining spatial concepts and transformations. Experiments revealed that Spatial- accurately extracts spatial concepts, assigns functional roles with principled ordering, and composes transformation sequences using template-based generation, establishing a crucial link between language and computation. This breakthrough delivers interpretable and executable geospatial workflows, moving beyond simple pattern matching or web searches often seen in existing large language model (LLM) based systems.

The team measured significant performance improvements on both the MapEval-API and MapQA benchmarks. Specifically, Spatial- substantially outperforms existing baselines, including ReAct and Reflexion, demonstrating a marked advancement in correctness and workflow generation. The core of this achievement lies in the GeoFlow Graph, an intermediate representation that explicitly bridges natural language with computational GIS, allowing for verifiable results rather than relying solely on parametric knowledge. Researchers constructed these graphs by extracting spatial concepts and instantiating them as nodes, then identifying functional roles to impose ordering constraints on the transformations.

Data shows that Spatial-Agent’s compositional GeoFlow Graph generation, based on macro-templates, captures recurring geo-analytical patterns and improves structural validity through template matching and IO-port composition. The system operates through a multi-stage pipeline, beginning with Spatial Information Theory Analysis to extract core concepts and group them by functional roles. Following this, Concept Transformation Drafting retrieves templates to define transformation patterns, and GeoFlow Graph Construction assembles these transformations into an ordered graph adhering to role-based precedence constraints. Tests prove that GeoFlow Graph Factorization & Tool Mapping converts the graph into an executable form with concrete operators, enabling precise spatial computation. Spatial-Agent addresses key challenges in geo-analytical reasoning, including distinguishing between semantic and spatial domains, differentiating cognitive and spatial reasoning, and moving from spatial orchestration to spatial execution. The work establishes a principled intermediate representation, enabling the agent to uncover the implicit structure of spatial questions and generate coherent, executable workflows for complex geospatial tasks.

GeoFlow Graphs enhance geospatial AI reasoning with structured

Scientists have developed Spatial-, a novel geospatial AI agent grounded in the principles of spatial information science, to improve geo-analytical question answering. The research formalises the process as a concept transformation problem, utilising GeoFlow Graphs, directed acyclic graphs representing spatial concepts and their transformations, to enable structured reasoning over geographic workflows. This approach leverages template-based compositional generation, drawing on recurring geo-analytical patterns, and is enhanced by supervised fine-tuning and direct preference optimisation to internalise geographic constraints. Extensive experimentation on MapEval-API and MapQA benchmarks demonstrates that Spatial- significantly outperforms existing agent baselines, such as ReAct and Reflexion, while generating interpretable and executable geospatial workflows.

Error analysis indicates that limitations primarily stem from the reliability of external geospatial APIs during execution, rather than the reasoning components themselves, validating the effectiveness of the structured spatial reasoning methodology. The authors acknowledge that the framework’s accuracy is currently constrained by the quality of data from these external APIs and the coverage of the template library, which may require expansion for novel question types. Scaling the fine-tuning approach also demands substantial annotated data, and the current evaluation is limited to English-language urban environments, leaving performance in specialised geographic domains unexplored. Future work could focus on addressing these limitations by improving API reliability, expanding the template library, and developing more efficient data annotation techniques, potentially broadening the applicability of Spatial- to diverse geographic contexts and languages.