SimCRAFT: Distilling Remote Sensing Agents via Synthetic Trajectories and Contextual Retrieval-Augmented Fine-Tuning

SimCRAFT proposes a framework independent of models, aimed at addressing the issues of large language models lacking domain knowledge and high infrastructure requirements. This framework uses a multi-agent synthesis engine and a simulation execution engine to generate the SimRS-14k corpus containing 14,000 constraint-verified workflow plans. Through context retrieval-enhanced fine-tuning (CRAFT) technology, the model can perform analogical reasoning when adapting to standard operating procedures, thereby enabling multi-step remote sensing workflow planning. Experiments show that the SimCRAFT-7B model outperforms open-source large language models significantly, with performance comparable to advanced closed-source models and dedicated remote sensing agents. It can be replicated on three 7B architectures, providing a competitive open-source baseline for lightweight remote sensing autonomous deployment in resource-constrained environments.