GeoSkill:Experience-Driven Hierarchical Skill Learning with Collaborative Revision forGeospatialAgents

Details

Geospatial agents performing recurring analytical tasks — spatial joins, remote sensing classification, route optimization — need to learn from prior executions rather than plan from scratch each time. Existing memory-augmented approaches struggle with two specific challenges: capturing long-horizon tool-chain experience and constraining LLM self-reflection from making unreliable revisions that corrupt the memory.\n\nGeoSkill addresses both with two components. The Hierarchical Skill Bank (HSB) maintains two layers: a Planning Skill Bank for high-level task strategies and a Tool Skill Bank for tool invocation constraints. This separation allows different retrieval strategies for planning vs tool selection. The Collaborative Trace-driven Skill Revision (CTSR) mechanism uses three roles — Judge, Critic, and Refiner — to locate skill-level defects in execution traces and update skills surgically, preventing misattribution errors.\n\nDuring deployment, the skill bank is frozen for retrieval-only use on unseen tasks — a design that avoids distribution shift during production. Experiments on EarthBench and ThinkGeo show improvements in both end-to-end task accuracy and tool-execution reliability, though specific numbers are not detailed in the abstract.\n\nThe hierarchical skill bank pattern is applicable beyond geospatial tasks to any agent domain with structured tool use and recurring task types.