Hierarchies over Pixels: A Benchmark for Cognitive Geospatial Reasoning for Agents

ICLR 2026 Conference Submission25567 Authors

Keywords: Benchmark evaluation, Remote sensing imagery, Tool-augmented LLMs

TL;DR: We introduce GeoHOP, a benchmark for hierarchical geospatial reasoning in remote sensing, and GeoPlanner, a tool-augmented LLM agent that excels in structured, fault-tolerant geospatial analysis.

Abstract: Beyond perception, reasoning is crucial in remote sensing, enabling advanced interpretation, inference, and decision-making. Recent advances in large language models (LLMs) have given rise to tool-augmented agents that enhance reasoning by leveraging external tools for complex analytical tasks. However, existing research on these agents in remote sensing largely focuses on perception-oriented tasks, with cognitive geospatial reasoning remaining underexplored. In this work, we systematically evaluate the geospatial reasoning capabilities of LLM-powered tool-augmented agents. To this end, we introduce GeoHOP, a benchmark for hierarchical geospatial reasoning. GeoHOP comprises 417 scenario-driven, hierarchy-aware tasks—such as hazard vulnerability assessment, urban heat island analysis, and forest fragmentation dynamics—spanning optical, Synthetic Aperture Radar (SAR), and infrared (IR) imagery. GeoHOP advances evaluation beyond monitoring-based recognition to cognitive-level geospatial analysis. Building upon GeoHOP, we propose GeoPlanner, an agent powered by LLMs that organizes 5 toolkits into functional hierarchies and executes fault-tolerant reasoning pipelines. GeoPlanner enables structured abstraction, robust recovery from tool failures, and stable long-horizon planning. Extensive experiments across diverse geospatial reasoning tasks demonstrate that GeoPlanner excels in hierarchical reasoning, cross-modal transfer, and error handling.

Primary Area: datasets and benchmarks

Submission Number: 25567