AerialVP preprint reports drone-image gains without model retraining

A preprint introduces AerialVP, a system that improves vision-language models' interpretation of drone images by expanding task prompts without changing model parameters. Researchers report consistent improvements across target localization, visual reasoning and question answering on the evaluated models. A language-model task engine coordinates tools that extract image details, analyze prompts and combine semantic descriptions with spatial constraints. The enhanced prompt and original image go to the unchanged model. The researchers also introduce the AerialSense benchmark and report competitive performance against instruction fine-tuning using LoRA, a method for adapting model weights.