Experience-Driven Multi-Agent Systems Are Training-free Context-aware Earth Observers

Abstract: Recent advances have enabled LLM agents to solve complex tasks by orchestrating external tools. However, these agents often struggle in specialized, tool-intensive domains that demand long-horizon execution, tight coordination across modalities, and strict adherence to implicit tool constraints. Earth Observation (EO) tasks exemplify this challenge due to the multi-modal and multi-temporal data inputs, as well as the requirements of geo-knowledge constraints (spectrum library, spatial reasoning, etc): many high-level plans can be derailed by subtle execution errors that propagate through a pipeline and invalidate final results. A core difficulty is that existing agents lack a mechanism to learn fine-grained, tool-level expertise from interaction. Without such expertise, they cannot reliably configure tool parameters or recover from mid-execution failures, limiting their effectiveness in complex EO workflows. To address this, we introduce \textbf{GeoEvolver}, a self-evolving multi-agent system~(MAS) that enables LLM agents to acquire EO expertise through structured interaction without any parameter updates. GeoEvolver decomposes each query into independent sub-goals via a retrieval-augmented multi-agent orchestrator, then explores diverse tool-parameter configurations at the sub-goal level. Successful patterns and root-cause attribution from failures are then distilled in an evolving memory bank that provides in-context demonstrations for future queries. Experiments on three tool-integrated EO benchmarks show that GeoEvolver consistently improves end-to-end task success, with an average gain of 12\% across multiple LLM backbones, demonstrating that EO expertise can emerge progressively from efficient, fine-grained interactions with the environment.

• The paper introduces GeoEvolver, a self-evolving multi-agent system that improves EO workflow execution reliability through experience-driven tool-level expertise, enabling parallel exploration and hierarchical memory consolidation.
• GeoEvolver achieves an average end-to-end accuracy improvement of +12.56 percentage points across five backbones on the EarthAgent benchmark, with model-specific accuracy improvements of up to +89.4%, and the contrastive distillation mechanism is the key factor, says the paper.

Motivation and problem statement

This paper addresses a persistent failure mode of LLM-based agents in specialized, tool-intensive scientific domains: the gap between high-level planning competence and low-level execution reliability. The authors ground this argument in Earth Observation (EO) workflows, where pipelines are state-coupled—the validity of each step depends on invisible metadata such as coordinate reference systems, spatial resolution, spatial extent, and temporal coverage. When an agent treats these attributes as symbolic parameters, it can produce "hallucinated execution": a toolchain that runs without errors but whose physical semantics are broken. The paper's central diagnostic claim is that EO agent failures are dominated by execution-grounded errors rather than planning errors, shifting the core challenge from "how to plan" to "how to execute reliably under tool-coupled constraints."

The proposed answer is GeoEvolver, a self-evolving multi-agent system (MAS) in which LLM agents acquire tool-level EO expertise through structured interaction with no parameter updates. The hypothesis is that verified interaction experience—successful patterns and root-cause attributions from failures—can be distilled into a retrievable memory bank serving as in-context demonstrations, thereby substituting for fine-tuning.

Architecture

GeoEvolver formalizes EO workflow generation as a retrieval-conditioned sequential decision problem. A query $q$ is decomposed by an Orchestrator into $N$ sub-goals $\{g_n\}$, each executed independently by a specialized Executor subagent producing a sub-trajectory $\tau_n$; the task succeeds only if all sub-goals succeed ($Y = \bigwedge_n Y_n$). This decomposition localizes failure attribution and enables parallel execution. Four operators form a retrieve–plan–execute–judge closed loop: a Retriever queries the memory bank via embedding similarity to build a strategy context; the Orchestrator assigns interface contracts to sub-goals; Executors run tools; and a Judge emits a binary success label plus auxiliary validity signals (format compliance, numeric matching, trajectory-length penalties).

Two mechanisms distinguish the design:

• Parallel exploration: $K$ variants of the full pipeline run independently with prompt/decoding diversity, each allowed up to $A$ corrective retries. The final solution is selected by prioritizing verified success, using Judge confidence as tie-breaker.
• Hierarchical memory: a global Memory Bank of distilled procedural knowledge (tool-chain patterns and failure-derived guardrails) persists across episodes, while a compressed Working Memory summarizes older interactions and retains only the $L$ most recent raw steps within an episode.

Self-evolution operates at two levels: single-variant extraction from the best solution (analysis patterns on success, error attributions on failure), and contrastive distillation across all $K$ variants, where concatenated candidate trajectories with success labels are prompted into transferable insights. Consolidation uses deterministic key-based deduplication to keep the bank compact. A leakage filter excludes memory entries co-occurring with expected outputs during benchmark evaluation.

Empirical results

Evaluation spans three benchmarks: ThinkGeo (486 multi-step spatial reasoning queries), EarthAgent (236 tasks over 104 heterogeneous tools), and GeoPlan-Bench (996 synthetic long-horizon planning tasks).

Cross-backbone gains. On EarthAgent, GeoEvolver improves end-to-end accuracy by an average of +12.56 pp across five backbones (GPT-5, Gemini-2.5, GPT-4o, DeepSeek-V3.1, Qwen3-32B). The effect is strongest for smaller models: Qwen3-32B rises from 24.80% to 46.96% (+22.16 pp, +89.4% relative), suggesting externalized experience partially compensates for limited parametric capacity. Notably, some step-level metrics decrease under GeoEvolver (e.g., GPT-5 Tool-A-O drops from 71.16% to 67.52%) while end-to-end accuracy rises (63.16% → 70.85%). The authors interpret this decoupling as evidence that functionally correct trajectories need not imitate human-annotated ones—a substantive design claim against trajectory imitation for domain adaptation.

Comparison with existing methods. On the Earth-Agent Benchmark-subset-65, GeoEvolver reaches 76.56% accuracy, versus 31.25% for Training-free GRPO, 30.16% for AFlow, 22.58% for ExpeL, and 15.87% for Earth-Agent-MAS. On ThinkGeo, a GPT-4o-mini instantiation achieves 46.88% accuracy—4.1× GPT-4o's 11.51%—with 90.50% Operation-F1 and 79.03% Logic-F1. On GeoPlan-Bench, GeoEvolver matches Earth-Agent-MAS on key-step F1 (0.63) with higher precision (0.68 vs. 0.65) but lower structural score (0.45 vs. 0.68); the authors attribute the narrower margin there to intentionally low execution complexity.

Ablations. Progressive removal from the full model shows Self-contrast is the largest contributor (−21.87 pp when removed), followed by Parallel Exploration (−18.75 pp further) and Memory (−4.69 pp further); the single-agent baseline attains only 25.00%, for a cumulative gain of +51.56 pp. Sensitivity analysis reveals an interaction effect between executors $N$ and variants $N$0: increasing either alone yields limited gains, whereas joint increases amplify performance, since decomposition granularity exposes tool-specific constraints only if exploration breadth surfaces diverse feedback. Moderate settings ($N$1, $N$2) balance cost and benefit.

Discussion

The paper frames its contribution as "Experience-as-Parameters": domain knowledge is externalized into a dynamic, non-parametric Memory Bank rather than encoded in weights vulnerable to catastrophic forgetting—an approach suited to EO, where sensor characteristics and toolchains evolve faster than foundation models are retrained. A second observation is that exploration itself produces actionable evidence: failed rollouts carry diagnostic signal about the environment's feasible region, and distilling failure-induced constraints prunes invalid actions early. The ablation showing contrastive distillation contributes more than memory storage alone supports this reading.

Limitations

The authors concede three limitations directly. First, parallel exploration scales token usage and wall-clock time approximately linearly in $N$3, which is prohibitive for time-critical applications such as rapid disaster response; the accuracy–latency trade-off is not systematically modeled. Second, the system exhibits a "Blind Agent" limitation: agents manipulate imagery through metadata and file handles rather than pixel data, limiting sensitivity to visual artifacts in intermediate rasters. Third, GeoEvolver is bounded by its underlying toolchain—if a tool fails silently or emits physically implausible output without explicit errors, the agent has limited grounding to detect semantic anomalies. An additional caveat borne by the evaluation itself is that the headline comparison on subset-65 uses a small task sample, and the strongest baseline gaps occur against methods (ExpeL, Training-free GRPO) designed for general-purpose rather than EO-specific memory organization.

Conclusion

GeoEvolver demonstrates that training-free, experience-driven specialization can yield large end-to-end gains in tool-intensive EO workflows, with an average +12.56 pp improvement across backbones and a 76.56% result on the most challenging benchmark—driven primarily by contrastive distillation over parallel explorations rather than by memory volume alone. The step-level versus task-level metric decoupling is the paper's most consequential empirical finding, arguing against trajectory imitation as a domain-adaptation strategy. Open questions include principled modeling of the $N$4-dependent cost–accuracy trade-off, integration of direct visual grounding into the Judge, and whether distilled memory traces can compress this "System 2" reasoning into compact fine-tuned policies.