Organizations: School of Geosciences and Info-Physics, Central South University, Changsha 410083, China · School of Earth Sciences and Spatial Information Engineering, Hunan University of Science and Technology, Xiangtan 411201, China · Sanya Institute, Hunan University of Science and Technology, Sanya 572024, China · School of Resources and Environment, University of Electronic Science and Technology of China, Chengdu 611700, China
Large language model-driven remote sensing (RS) agents offer a promising approach to automating geospatial analysis. However, lightweight RS agents based on compact language models struggle with multi-step interactive tasks due to loss of long-horizon states, inefficient environmental feedback utilization, and sparse optimization signals. We propose RS-Claw-Evolution, an environment-feedback-driven framework that progressively improves lightweight agents through three stages. Interaction evolution uses executable code to control observations, maintain intermediate states, and reduce context redundancy. Experience evolution combines failure-aware trajectory generation with error-turn masking to learn from informative failure-recovery experiences without imitating faulty actions. Decision evolution uses reinforcement learning with multi-dimensional environment rewards and turn-level advantage protection to optimize tool-use behaviors and improve credit assignment in long sequences. On Earth-Bench, the optimized Qwen3-4B-based agent achieves 65.9% accuracy in Autonomous Planning mode, outperforming the untrained Qwen3-32B baseline (43.8%) and DeepSeek-V3.1 (60.8%), while approaching GPT-5 (71.6%). These results demonstrate that learning from environmental feedback can improve lightweight agents and narrow their performance gap with larger models in long-horizon RS tasks.
Figures & tables
Fig. 1: Comparison between the original RS-Claw paradigm and the proposed RS-Claw-Evolution framework. The original RS-Claw paradigm treats environmental feedback as execution information during inference, while RS-Claw-Evolution further exploits environmental feedback as evolution signals to progressively improve lightweight agents through three stages: Interaction Evolution via CodeAct-based interaction, Experience Evolution via failure-aware experience learning, and Decision Evolution via feedback-driven policy optimization.
Fig. 2: Overview of the Experience Evolution stage in RS-Claw-Evolution: feedback-driven SFT for lightweight RS agents. (Top) Hint-guided Failure-aware Trajectory Generation extracts task-level hints through hindsight analysis of failed trajectories with reference to successful code trajectories. These hints guide teacher-agent re-exploration, followed by rejection sampling to retain successful trajectories containing intermediate error feedback and recovery behaviors. (Bottom) Error-turn Masking Training removes explicit hints from student training inputs and masks the loss on erroneous action turns while retaining error feedback and subsequent recovery behaviors, enabling lightweight agents to learn error diagnosis and recovery without imitating faulty actions.
Fig. 3: Overview of the Decision Evolution stage in RS-Claw-Evolution: feedback-driven GRPO optimization for long-horizon RS agents. Compared with GRPO using sparse trajectory-level rewards, the proposed approach incorporates multi-dimensional process rewards derived from environmental feedback for finer-grained policy optimization. Furthermore, TAP selectively masks misleading turn-level advantages in successful trajectories containing erroneous exploration turns, reducing erroneous positive reinforcement while protecting recovery behaviors from inappropriate negative penalties. Together, these mechanisms support more reliable and efficient decision-making in long-horizon tool-use tasks.
Model
Paradigm
Accuracy
Tool-Any-Order
Tool-In-Order
AP
IF
AP
IF
AP
IF
Qwen3-4B-Instruct
Code
31.8
35.8
47.6
52.8
40.1
45.8
Qwen3-4B SFT
Code
41.5
44.3
58.6
60.9
49.4
51.8
Qwen3-4B RL (Ours)
Code
65.9
65.9
69.1
70.1
58.9
58.4
Qwen3-8B
Code
19.3
22.7
49.6
50.9
41.1
44.2
Qwen3-8B SFT
Code
46.6
44.3
61.4
63.6
52.2
53.2
TABLE I: Overall Performance Comparison of Language Models with Varying Scales on Earth-Bench. "Code" indicates the programmatic interaction paradigm utilized in our method, while "JSON" represents the traditional structured tool-calling paradigm used in the original RS-Claw framework.
Fig. 4: Comparison of context efficiency between Code and JSON interaction paradigms. The left and right panels respectively show the average Action Tokens and Observation Tokens consumed in a single trajectory by different language models (DeepSeek-V3.1, GPT-5, and Qwen3-32B). The results indicate that compared to the traditional structured JSON tool invocation paradigm, the programmatic interaction paradigm (Code) adopted in this paper effectively reduces context overhead in both AP and IF modes.
Fig. 5: Impact of the Hindsight Hint guidance mechanism on the trajectory generation success rate of 12 training tasks. The solid blue line and dashed red line represent the exploration success rates of the teacher agent with and without the Hindsight Hint guidance, respectively.
Scale
SFT Strategy
Accuracy
Tool-Any-Order
Tool-In-Order
EIR
4B
Before Training
31.8
47.6
40.1
45.8
Oracle
33.5
54.2
44.8
57.4
Unmasked
37.5
58.9
48.6
36.9
Masked (Ours)
41.5
58.6
49.4
31.6
8B
Before Training
19.3
49.6
41.1
53.1
Oracle
35.2
56.7
47.5
57.8
TABLE II: Ablation Study of SFT Strategies for Agent Error Recovery Across Model Scales. Oracle, Unmasked, and Masked denote error-free expert trajectory SFT, full-trajectory SFT, and the proposed error-turn masking strategy, respectively.
Variant
Accuracy
Tool-Any-Order
Tool-In-Order
Avg Turns
EIR
SFT Only
41.5
58.6
49.4
17.426
31.6
Vanilla GRPO
58.5
64.9
53.7
11.688
10.3
w/o MEFR
61.9
68.1
58.1
12.920
8.0
w/o TAP
63.1
68.6
58.6
13.068
7.7
Ours
65.9
69.1
58.9
12.761
6.1
TABLE III: Ablation Study of RL Components Based on the Qwen3-4B-Instruct Model After the Proposed SFT Phase.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Task ID
Application Domain
Workflow Type
3
Vegetation and Water Stress
Composite Analysis
43
Vegetation and Water Stress
Threshold Analysis
14
Land Surface Thermal
Temporal Trend
39
Land Surface Thermal
Threshold Analysis
77
Land Surface Thermal
Dual-object Difference
72
Land Surface Thermal
Composite Analysis
Appendix
TABLE S1: Selected Training Tasks.
Hyperparameter
Value
Training epochs
2
Global batch size
4
Max tokens per GPU
20,480
Max sample tokens
20,480
Learning rate
5×10−6
Minimum learning rate
5×10−7
Appendix
TABLE S2: Hyperparameters for Supervised Fine-Tuning (SFT).
Hyperparameter
Value
Training epochs
5
Learning rate
1×10−6
LR schedule
Constant
Optimizer
Adam ( β1=0.9,β2=0.98 )
Weight decay
0.1
KL loss coefficient
0.001
Appendix
TABLE S3: Hyperparameters for Reinforcement Learning (GRPO).
Metric
Unmasked SFT
Masked SFT (Ours)
Turn-level Statistics
Total Agent Turns
17,816
17,816
Trainable Turns
17,816
14,929
Masked Turns
0
2,887
Turn Retention Ratio
100%
83.80%
Turn Masking Ratio
0%
16.20%
Appendix
TABLE S4: Turn-Level and Token-Level Masking Statistics.
Model
Paradigm
Accuracy
Efficiency
Tool-Any-Order
Tool-In-Order
Tool-Exact-Match
Parameters
AP
IF
AP
IF
AP
IF
AP
IF
AP
IF
AP
IF
Qwen3-4B-Instruct
Code
31.8
35.8
3.030
3.904
47.6
52.8
40.1
45.8
35.4
37.8
23.8
24.2
Qwen3-4B SFT
Code
41.5
44.3
4.650
4.356
58.6
60.9
49.4
51.8
42.9
42.6
23.9
24.8
Qwen3-4B RL (Ours)
Code
65.9
65.9
4.917
6.446
69.1
70.1
58.9
58.4
48.6
48.2
24.1
24.0
Qwen3-8B
Code
19.3
22.7
3.957
3.907
49.6
50.9
41.1
44.2
36.9
37.8
23.8
24.8
Qwen3-8B SFT
Code
46.6
44.3
4.242
5.158
61.4
63.6
52.2
53.2
44.3
42.7
25.3
25.1
Appendix
TABLE S5: Comprehensive Performance Evaluation Results. All metrics except Efficiency are reported as percentages.
The rise of multi-modal large language models (MLLMs) is shifting remote sensing (RS) intelligence from "see" to "action", as OpenClaw-style frameworks enable agents to autonomously operate massive RS image-processing tools for complex tasks. Existing RS agents adopt a passive selection paradigm for tool invocation, relying on either full tool registration (Flat) or retrieval-augmented generation (RAG). However, in the massive and multi-source heterogeneous RS tool ecosystem, such passive mechanisms struggle to dynamically balance "context load" and "toolset completeness" throughout task reasoning, thus exhibiting inherent limitations: full tool registration triggers context space deficits during long-horizon tasks, whereas RAG retrieval may omit critical tools in essential steps. To overcome these bottlenecks, this paper redefines tool selection by arguing that the agent should act as an active explorer within the tool space. Based on this perspective, we propose RS-Claw, a novel RS agent architecture. By leveraging Skill encapsulation technology at the tool end, this architecture hierarchically structures tool descriptions, enabling the agent to execute on-demand sequential decision-making: initially selecting relevant skill branches by reading only tool summaries, then dynamically loading detailed descriptions, and ultimately achieving precise invocation. This active paradigm not only significantly liberates the agent's context space but also effectively ensures the accurate hit rate of critical tools during long-horizon reasoning. Systematic experiments on the Earth-Bench benchmark demonstrate that RS-Claw's active exploration mechanism effectively filters semantic noise and substantially frees up reasoning space, achieving an input token compression ratio of up to 86%, and comprehensively outperforming existing Flat and RAG baselines across complex reasoning evaluations.
Liangtian Liu, Zeyuan Wang, Ziyu Li +8
School of Geosciences and Info-Physics, Central South University, Changsha 410083, China · School of Resources and Environment, University of Electronic Science and Technology of China, Xian 710071, China · School of Earth Sciences and Spatial Information Engineering, Hunan University of Science and Technology, Xiangtan 411201, China +1
Remote sensing visual models have continuously advanced various interpretation tasks. However, the research process behind model improvement still heavily relies on manual expertise, requiring extensive trial-and-error iterations in model design, data processing, and performance diagnosis. Existing agent-based approaches mainly focus on task execution and workflow orchestration, while lacking the capability of autonomous research iteration for continuous performance optimization. To address this issue, we propose RingMoClaw, an experience-inspired self-evolving multi-agent framework for remote sensing visual interpretation. RingMoClaw integrates a research branch, a quality-control branch, and a dual-stream dynamic experience bus to establish a closed-loop optimization process covering strategy generation, experiment execution, independent review, and experience accumulation. The heterogeneous Critic mechanism provides stage-wise diagnosis and feedback, while the dual-stream experience bus incorporates external knowledge and internal experimental experience to guide strategy evolution and eliminate ineffective searches. Extensive experiments on four remote sensing downstream tasks, including object detection, scene classification, semantic segmentation, and change detection, demonstrate the effectiveness and generalization of RingMoClaw. Compared with the corresponding baseline models, RingMoClaw improves performance by 1.84% mAP50 on object detection and achieves consistent gains across the other three tasks, while reducing the required evolution steps by over 40% compared with existing research automation frameworks. These results suggest that RingMoClaw offers a feasible route from task execution toward continuous research driven model evolution in remote sensing.
Kaiyue Kang, Qixuan He, Peijin Wang +9
Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100190, China · School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing 100190, China · University of Chinese Academy of Sciences, Beijing 100190, China +1
Recent advances in large language models and multimodal models have pushed remote sensing (RS) processing from simple perception models to agentic systems designed to tackle complex, long-horizon RS tasks. However, existing systems often rely on monolithic decision-making frameworks, which fail to accommodate the multi-stage, interdependent nature of RS tasks. This centralized approach leads to challenges such as unstable task execution, incorrect tool usage, and error propagation across stages. To address these issues, we propose HiRS-Agent, a hierarchical multi-agent system for long-horizon RS task solving. HiRS-Agent adopts a two-level collaborative architecture: the Manager Layer handles dynamic routing, step-level verification, replanning, and termination control, while the Specialist Layer organizes domain-specific tools according to the RS workflow and is responsible for subtask reasoning and tool execution. To further enhance the system's capability, we introduce a two-stage supervised tuning strategy and a verification-guided hierarchical reinforcement learning stage to jointly optimize coordination and tool-use policies. Experiments on Earth-Agent Benchmark and ThinkGeo show that HiRS-Agent substantially improves long-horizon tool-use capability and final-task correctness, demonstrating the effectiveness of structured multi-agent collaboration for reliable RS agents. The code is publicly available at https://github.com/IntelliSensing/HiRS-Agent.
Boyang Mu, Zhiwei Wei, Mugen Peng +1
Beijing University of Posts and Telecommunications State Key Laboratory of Networking and Switching Technology Beijing, China · Hunan Normal University School of Geographic Sciences Changsha, Hunan, China