[论文解读] Experience-Driven Multi-Agent Systems Are Training-free Context-aware Earth Observers

Pengyu Dai, Weihao Xuan+5 more|arXiv (Cornell University)|Jan 30, 2026

Multimodal Machine Learning Applications被引用 0

一句话总结

GeoEvolver 是一个无需训练、以经验为驱动的多代理系统，能够在记忆库中积累细粒度的 EO 工具执行先验，从而在不更新参数的情况下提升端到端的地球观测任务表现。它将查询分解、探索工具配置，并将失败提炼为可重用的记忆。

ABSTRACT

Recent advances have enabled large language model (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 extbf{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.

研究动机与目标

• Identify why EO failures stem from execution-groundedness rather than planning alone.
• Propose GeoEvolver to acquire EO expertise through structured interaction without updating model parameters.
• Show that memory of execution experiences improves end-to-end EO task success across multiple LLM backbones.

提出的方法

• Decompose each EO query into modular sub-goals assigned to specialized executors.
• Use a retrieval-augmented orchestrator to assemble sub-goals from a memory bank of patterns and failures.
• Allow parallel exploration with multiple variants and retries to find robust tool configurations.
• Judge and validate sub-goal trajectories and propagate success/failure signals to memory.
• Maintain a two-tier memory system: a global Memory Bank and a local Working Memory.
• Iteratively distill successful patterns and failure attributions into the memory bank through single-variant extraction and contrastive distillation.

实验结果

研究问题

• RQ1 Does GeoEvolver improve end-to-end EO task performance across diverse LLM backbones?
• RQ2 How does model capacity affect GeoEvolver's gains across EO benchmarks?
• RQ3 Is GeoEvolver robust across EO benchmarks with different tool–modality couplings?
• RQ4 How does GeoEvolver compare to existing memory-based and multi-agent EO methods?
• RQ5 What is the impact of the number of executors, reasoning variants, and memory items on performance?

主要发现

Method	Tool-A-O ↑	Tool-I-O ↑	Tool-E-M ↑	Efficiency ↓	Accuracy ↑
Expel	32.72	25.94	22.48	1.79	22.58
Zhao et al. (Training-free GRPO)	57.24	44.36	36.44	1.36	31.25
Chase (DeepAgents)	41.67	33.98	25.45	1.06	29.69
Earth-Agent-MAS	32.28	26.96	20.91	1.47	15.87
Ours (GeoEvolver)	57.66	44.66	39.06	1.47	76.56

• GeoEvolver achieves an average end-to-end accuracy gain of 12.56 percentage points across multiple backbones on EO benchmarks.
• Smaller models gain disproportionately from memory-augmented experience, e.g., Qwen3-32B improves from 24.80% to 46.96% (+22.16 pp).
• End-to-end accuracy improvements can accompany decreased step-level scores, indicating functionally correct but non-human trajectories.
• GeoEvolver outperforms memory-based methods and fixed-workflow MAS on Earth-Agent benchmarks (e.g., 76.56% vs. 15.87% in Earth-Agent-MAS).
• Ablations show self-contrast and parallel exploration contribute the largest gains, with notable drops when removed.