AGCD: Agent-Guided Cross-Modal Decoding for Weather Forecasting

ArXi:2603.15260v1 Announce Type: new Accurate weather forecasting is than grid-wise regression: it must preserve coherent synoptic structures and physical consistency of meteorological fields, especially under autoregressive rollouts where small one-step errors can amplify into structural bias. Existing physics-priors approaches typically impose global, once-for-all constraints via architectures, regularization, or NWP coupling, offering limited state-adaptive and sample-specific controllability at deployment.