New preprint on a comprehensive benchmark of instruction-following vision-language models on Earth observation data.