Geospatial agents are expected to reuse prior execution experience for recurring and evolving analytical tasks rather than run isolated workflows. GeoSkill addresses this need with an experience‑driven hierarchical skill learning framework consisting of two main components:
-
Hierarchical Skill Bank (HSB): It contains a Planning Skill Bank that abstracts high‑level task‑planning experience and a Tool Skill Bank that captures tool‑level invocation constraints. This design enables structured representation and cross‑task reuse of historical execution knowledge.
-
Collaborative Trace‑driven Skill Revision (CTSR): Three collaborators—Judge, Critic, and Refiner—work together. Judge detects errors, Critic localizes skill‑level defects, and Refiner performs targeted modifications, preventing misattributed revisions from contaminating the skill bank.
During development, GeoSkill learns and validates skills from historical execution traces; during deployment, the skill bank is frozen and used only for retrieval‑based guidance, ensuring reliability on unseen tasks. Extensive experiments on EarthBench and ThinkGeo demonstrate that GeoSkill converts historical experience into reusable hierarchical skills, improving both end‑to‑end task accuracy and tool‑execution reliability.
Review