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

• Recurring analyses across seasons, disasters, and domain tasks create demand for agents to accumulate and reuse experience from prior executions.Such experience includes both transferable task-planning patterns and operational knowledge tied to tools, data states, and execution constraints.Source: Introduction · p. 1
• Geospatial experience needs separate representations for transferable workflow plans and tool-level constraints tied to data states, because over-detailing or over-abstraction undermines reuse or reliable execution.Planning knowledge covers task decomposition and tool-chain organization, while tool knowledge includes parameters, I/O dependencies, physical feasibility, CRS compatibility, and resolution matching.Source: Introduction · p. 2
• Geospatial skill revision is difficult because spatial operations alter intermediate states, failures may surface downstream, and rewriting whole skills risks overwriting validated knowledge.Reprojection, resampling, and temporal filtering can affect subsequent analyses, while failures may stem from planning, parameters, or environmental changes.Source: Introduction · p. 2
• GeoSkill uses a Hierarchical Skill Bank for planning and tool skills, and CTSR’s Judge, Critic, and Refiner target revisions that are validated by re-execution before commitment.The framework’s Retrieve–ReAct–Revise–Retain loop evolves skills without updating the underlying LLM parameters.Source: Introduction · p. 2
• Experiments on EarthBench and ThinkGeo show that GeoSkill improves both end-to-end task accuracy and tool-execution reliability.Source: Introduction · p. 2

LLM-based Geospatial Agents

• Geospatial agents span remote-sensing task models, executable workflows, hierarchical decomposition, and process or memory-based recovery, but reusable procedural learning remains unexplored.Early systems support language-driven scene understanding and detection (Liu et al. 2024; Guo et al. 2024; Xu et al. 2024); later work explores executable workflows (Shabbir et al. 2025; Feng et al. 2025), hierarchical abstraction (Li et al. 2025), and process knowledge or memory-based recovery (Chen et al. 2025; Wu et al. 2026).Source: LLM-based Geospatial Agents · p. 2

Experience and Memory Learning for LLM Agents

• LLM experience methods support within-task correction, cross-task memory, and tool-call learning, yet semantic summaries omit geospatial execution constraints and mechanisms for revising faulty knowledge.Reflexion and CRITIC use reflection and feedback for within-task correction (Shinn et al. 2023; Gou et al. 2024); ExpeL, AWM, and A-MEM support cross-task reuse (Zhao et al. 2024; Wang et al. 2024c; Xu et al. 2025), alongside tool-calling work (Qin et al. 2023; Patil et al. 2023; Schick et al. 2023). Existing representations do not encode constraints such as data state, spatial reference system, resolution, or tool parameters, and lack diagnosis and revision mechanisms (Qiao et al. 2024; Wang et al. 2024b; Huang et al. 2022; Zhang et al. 2025).Source: Experience and Memory Learning for LLM… · pp. 2–3

Skill Learning and Agent Capability Evolution

• Skill-learning work formalizes procedural units and progressive optimization, but failure attribution remains weak and general-purpose skills overlook geospatial dependencies and cascading errors.GeoSkill addresses these gaps by separating planning from tool knowledge and using trace-driven defect localization, targeted modification, and re-execution validation through CTSR.Source: Skill Learning and Agent Capability… · p. 3

Method

• During development, GeoSkill iteratively evolves skills from historical tasks; during deployment, it freezes the skill bank for inference, without updating LLM parameters.The framework follows a closed loop of retrieval, execution, revision, and retention.Source: Method · p. 3
• The Hierarchical Skill Bank separates planning-level knowledge from execution-level constraints, while CTSR uses a Judge–Critic–Refiner pipeline for failure attribution, defect localization, and constraint-preserving revision.Together, these are GeoSkill’s two core components.Source: Method · p. 3

Problem Formulation

• GeoSkill formulates geospatial experience learning as a Memory-based Markov Decision Process in which an agent interacts with an executable environment through multi-step reasoning and tool invocation.States include the task objective, available data, prior tool calls, intermediate artifacts, and environment feedback; actions include reasoning, parameterized tool calls, or a final answer.Source: Problem Formulation · p. 3
• At each rollout’s start, hierarchical retrieval selects a stable context containing one planning skill and the tool skills required by its workflow to condition the ReAct trajectory.This avoids changing retrieved context at every decision step.Source: Problem Formulation · p. 3
• A trajectory-level verifier evaluates task success against a semantic reference answer; available reference explanations or trajectories serve as diagnostic evidence, not the only valid execution path..Source: Problem Formulation · p. 3
• Rather than optimizing model parameters, GeoSkill updates external procedural memory between rollouts using task, retrieved skills, trajectory, environment outputs or errors, predicted answer, and available references; CTSR implements the update.The memory update is defined as M^(r+1) = U(M^r, X^r).Source: Problem Formulation · p. 3

Overview of GeoSkill

• During development, GeoSkill retrieves planning and required tool skills, initializes missing candidates, and retains the candidate bundle only after successful execution.If execution fails, CTSR distinguishes skill defects from agent deviations and runtime failures; only skill-attributed defects trigger revision, and unsuccessful bounded re-execution restores the pre-task bank.Source: Overview of GeoSkill · p. 3, Figure 2 · p. 4
• The Hierarchical Skill Bank separates procedural memory into Planning Skills for task planning and Tool Skills for tool-use constraints.The planning skill encodes a tool chain that determines which tool skills are retrieved.Source: Overview of GeoSkill · p. 3, Figure 2 · p. 4
• At deployment, GeoSkill freezes the skill bank and retrieves skills to guide unseen-task execution without further updates.Source: Figure 2 · p. 4
• CTSR uses Judge, Critic, and Refiner to attribute failures, localize skill defects from execution traces, and make targeted, constraint-preserving revisions.Source: Figure 2 · p. 4

Hierarchical Skill Bank

• GeoSkill separates reusable task-level workflows from tool-specific execution constraints into Planning and Tool Skill Banks, using shared records with different workflow granularity.Records specify a skill identifier, function, use conditions, workflow or invocation template, validation constraints, and termination conditions; planning templates describe tool chains, while tool templates describe parameter schemas, input–output contracts, and execution notes.Source: Hierarchical Skill Bank · p. 4

Planning Skill Retrieval and Initialization

• For a task, GeoSkill retrieves a planning skill as a Plan-and-Execute prior; if none applies, it creates a task-local candidate that must pass rollout verification before retention.The planning template captures stable task-family workflow structure, required tools, stage dependencies, and validation requirements.Source: Planning Skill Retrieval and… · p. 4

Tool Skill Retrieval and Initialization

• GeoSkill retrieves tool skills for operators in the selected workflow, preserving invocation and artifact constraints; missing skills are initialized as candidates, and chain-conditioned retrieval avoids unrelated tool knowledge.Tool skills include parameter names and types, input requirements, output semantics, artifact dependencies, spatial-reference and resolution constraints, and validation conditions.Source: Tool Skill Retrieval and Initialization · pp. 4–5

Skill-Conditioned ReAct

• The complete skill bundle supplies a workflow-level prior through planning skills and constrains parameter generation and artifact propagation through tool skills.The agent remains responsive to environment observations while its tool choices and parameter generation are guided by the activated skills.Source: Skill-Conditioned ReAct · p. 5
• After each rollout, GeoSkill records the task, rendered skills, trajectory, compliance report, environment outputs or errors, predicted answer, and semantic reference answer as diagnostic evidence.Source: Skill-Conditioned ReAct · p. 5
• If the recorded evidence indicates success, the skill bundle proceeds to Retain; otherwise, GeoSkill passes the evidence to CTSR.Source: Skill-Conditioned ReAct · p. 5

Causal Trace-Driven Skill Revision

• CTSR distinguishes skill defects from non-compliance, synthesis, and runtime errors, preventing failures from automatically overwriting verified knowledge.Its three stages run sequentially in one shared LLM conversation context, preserving execution evidence across roles.Source: Causal Trace-Driven Skill Revision · p. 5

Judge: Failure Responsibility Routing

• Judge routes failures to responsibility categories, selects skills for inspection, and records validated components that must remain unchanged.For non-compliance or answer-synthesis errors, it produces an ephemeral reflection for the next rollout without modifying the skill bank.Source: Judge: Failure Responsibility Routing · p. 5

Critic: Skill-Level Causal Diagnosis

• Critic independently diagnoses each target skill, traces the causal path to failure, identifies rewrite targets, and extends the preserve ledger.It checks planning skills for missing stages, tool choice, ordering, dependencies, and termination; tool skills for parameters, contracts, data dependencies, physical constraints, and metric-to-answer mapping.Source: Critic: Skill-Level Causal Diagnosis · p. 5

Refiner: Constraint-Preserving Targeted Revision

• Refiner issues field-level patches rather than regenerating skills, while a deterministic guard rejects replacements or deletions that conflict with preserved components without direct-conflict evidence.When multiple skills are defective, they are diagnosed and revised sequentially in the shared context.Source: Refiner: Constraint-Preserving… · p. 5

Retain and Development Procedure

• The complete bundle of newly initialized or revised skills is re-executed on the original task before any candidate is committed.Plausibility judgments by the LLM alone are not sufficient for commitment.Source: Retain and Development Procedure · p. 5
• If the rollout succeeds, all new or revised working-copy skills are atomically committed to HSB; otherwise, CTSR processes the new trajectory in the next rollout.Source: Retain and Development Procedure · p. 5
• If no rollout succeeds within the bounded budget B, all candidates and patches are discarded and the pre-task bank snapshot is preserved, preventing partial updates from breaking planning–tool skill consistency.Retain acts as both an execution-verification gate and a transactional commit mechanism.Source: Retain and Development Procedure · p. 5

Deployment with a Frozen Skill Bank

• For unseen tasks, GeoSkill retrieves planning and tool skills before skill-conditioned ReAct, following intent analysis and tool-chain parsing.Deployment uses the frozen hierarchical skill bank.Source: Deployment with a Frozen Skill Bank · p. 6
• Deployment uses no reference answers or Judge, Critic, or Refiner, writes back no skills, and falls back to vanilla ReAct if no planning skill applies.Source: Deployment with a Frozen Skill Bank · p. 6

Experimental Setup

• Evaluation spans EarthBench workflows and ThinkGeo perception, orchestration, and logic; experience methods learn on one balanced partition and freeze skills for evaluation on the other.Reasoning-based baselines solve the evaluation partition directly, without prior experience accumulation.Source: Experimental Setup · p. 6

Evaluation Metrics

• Evaluation reports end-to-end Accuracy, EarthBench tool reliability (Tool-Any, Tool-In, Tool-Exact, Parameter Accuracy) and redundancy ratio, plus ThinkGeo Perception, Orchestration, Logic, and answer accuracy.Efficiency measures the ratio between predicted and reference tool calls.Source: Evaluation Metrics · p. 6

Main Results and Analysis

• With Qwen3-8B, GeoSkill achieves 41.13% accuracy versus Reflexion’s 37.10%, and raises Tool-Exact from 8.47% to 22.59% and Parameter Accuracy from 5.09% to 13.88%.The paper attributes these tool-level gains to HSB’s explicit modeling of parameter names, types, and dependencies.Source: Main Results and Analysis · p. 6
• On Qwen3-32B, GeoSkill† reaches 43.55% versus GeoSkill’s 41.13%, but on Qwen3-8B it falls to 38.71% versus 41.13%.The paper attributes the reversal to stronger-model skills encoding multi-step reasoning beyond the 8B model’s execution capacity; same-backbone skill evolution better matches its reasoning budget.Source: Main Results and Analysis · p. 6
• On DeepSeek-V4-Pro, GeoSkill reaches 79.03% EarthBench accuracy versus ReAct’s 58.06% and Reflexion’s 73.98%, while ThinkGeo Logic F1 rises from 41.88% to 59.61%.On EarthBench, the efficiency ratio falls from Reflexion’s 12.93 to 6.08.Source: Main Results and Analysis · p. 6, Table 1 · p. 7

Ablation Study

• On EarthBench with DeepSeek-V4-Pro, adding HSB raises accuracy from 65.32% to 72.58%, and CTSR further raises it to 79.03%.The results support separating planning knowledge from tool constraints and using failure diagnosis with selective revision rather than directly rewriting stored experiences.Source: Ablation Study · pp. 6–7, Table 2 · p. 7

Skill Evolution Analysis

• Accuracy rises monotonically from 72.58% at B=0 to 79.03% at B=2, with the first revision adding 3.23 points; Tool-Exact climbs from 55.80% to 58.37%.The B=0 skill bank already outperforms ReAct’s 58.06% accuracy.Source: Figure 3 · p. 7, Skill Evolution Analysis · p. 7

Conclusion

• GeoSkill enables geospatial agents to distill prior executions into reusable procedural knowledge.Experiments on EarthBench and ThinkGeo confirm improvements in accuracy and reliability.Source: Conclusion · p. 7
• HSB separates planning skills from tool constraints to support transferable yet executable skills.It decouples planning from tool constraints.Source: Conclusion · p. 7
• CTSR attributes errors, localizes skill defects, and validates patches to prevent unreliable skill accumulation.Source: Conclusion · p. 7