Timestep-Conditioned Transformers for Global Weather Forecasting
GEM-3 lets the forecast timestep be chosen at inference with one trained model.
The paper says fixed autoregressive timesteps force a trade-off between sub-daily detail and longer-horizon error accumulation. GEM-3 handles that with explicit multi-timestep inference, so the same weights can run at different timestep settings. The authors also report that mixed-timestep training improves rollout stability compared with timestep-specialist models. ArXiv · AI/CL/LG's note
The paper says fixed autoregressive timesteps force a trade-off between sub-daily detail and longer-horizon error accumulation. GEM-3 handles that with explicit multi-timestep inference, so the same weights can run at different timestep settings. The authors also report that mixed-timestep training improves rollout stability compared with timestep-specialist models. ArXiv · AI/CL/LG's note
score 5