FBA: A New Post-Training Method for Scenario-Specialized Remote Sensing MLLMs

A new method called Filling Before Advancing (FBA) is proposed for post-training remote sensing multimodal large language models (RS-MLLMs) to achieve fine-grained scenario specialization. The approach addresses the challenges of scarce high-quality scenario data and incomplete capability coverage by first filling prerequisite capability gaps before advancing to target-domain adaptation. FBA is instantiated for coastal harbor understanding using the CPRS (Coastal-Port Remote Sensing) dataset, which features a three-layer supervision structure and three ordered training stages: RS semantic anchoring for overhead-view visual-language alignment, domain-bridge convergence for shared RS priors, and scenario specialization. The work is published on arXiv under ID 2607.22205.