NEST: Tackling Dataset-Level Distribution Shifts via Regime-Oriented Mixture-of-Experts

NEST addresses dataset-level distribution shifts in long-horizon forecasting for complex systems. Its two-phase dense Mixture-of-Experts architecture first partitions data into operational regimes using unsupervised clustering in a moment-entropy space. A regime-oriented router generates initial expert weights from temporal content and refines them through geometric modulation relative to regime centroids. Instead of serving as monolithic predictors, experts act as specialized kernels that learn regime-specific dynamics through distinct variate-attention patterns. The authors report state-of-the-art results across benchmarks involving heterogeneous network traffic and physical phenomena, with code and datasets released on GitHub.