NeoST: Synthetic Data Trains Spatio-Temporal Foundation Models

Spatio-Temporal Foundation Models (STFMs) aim to understand and predict complex dynamic systems across space and time. Existing STFMs often struggle with biases present in real-world training data, limitations of autoregressive or diffusion-based architectures, and objectives that overemphasize precise point-wise reconstruction in noisy environments. This research presents NeoST, a groundbreaking STFM that is pre-trained entirely on procedurally generated synthetic data. This approach effectively bypasses the distributional biases inherent in real-world datasets. NeoST employs a novel latent-space reasoning architecture that can generate and iteratively refine multiple future trajectories, avoiding the accumulation of errors common in sequential prediction methods. By focusing on latent-space objectives that prioritize structural dynamics, NeoST demonstrates superior performance and long-horizon stability across a wide range of real-world spatio-temporal benchmarks. Its ability to infer and correct under distribution shifts makes it highly efficient and robust.