Switching dynamical systems are a common tool for interpretable time-series analysis: they segment data into regimes, each governed by a simple dynamic primitive. But the hard discrete choices cause problems: they cannot represent gradual transitions or overlapping states, and on messy real-world data they often flicker rapidly between regimes.

The Gumbel Dynamical Model (GDM) addresses this by relaxing the discrete state selection into a continuous stochastic variable using Gumbel noise. This lets the model combine reusable dynamic primitives smoothly and capture soft, sticky states, while remaining fully differentiable. As a result, training can use standard gradient descent instead of more complicated discrete optimization.

The authors validate GDM on standard simulations and two real-world datasets, showing it can infer interpretable states in stochastic settings where traditional methods fail. The paper is a single arXiv preprint accepted at NeurIPS 2026; no independent replication is reported in the source.