RS-WorldModel: A Unified Benchmark for Fine-Grained Remote Sensing with a Strong Baseline

Remote sensing world modeling requires models that both explain observed Earth-surface changes and generate future scenes under specified conditions at fine granularity, yet existing methods typically decouple interpretation from forecasting and operate at a coarser granularity. To bridge this gap, we construct RSWBench-1.1M, a unified **1.1 million**-sample benchmark whose fine-grained language annotations tie each described change to rich geographic and acquisition metadata, for spatiotemporal change question answering and text-guided future scene forecasting. As a strong baseline, we present RS-WorldModel, a unified autoregressive world model that combines geo-aware generative pre-training, shared instruction tuning, and reference-based verifiable reward alignment to unify understanding and forecasting in *a single model*. With only **2B** parameters, RS-WorldModel surpasses open-source models up to **120**$\times$ larger on most spatiotemporal change question-answering metrics and, on text-guided future scene forecasting, achieves the best PSNR of **17.20** and the best FID of **43.13** among all compared methods, including the closed-source Gemini-2.5-Flash Image.