Time series foundation models are increasingly used for general-purpose forecasting, but the abstract notes a key limitation: most top models output marginal quantiles, so the dependence structure of future trajectories is left unspecified. That means they may give accurate per-step uncertainty but not capture how different future points move together.

The paper introduces PathFM, described as a fast generative forecasting model for multivariate trajectories. Generative forecasting produces full sample paths rather than separate quantiles, which can preserve the joint dependence across time steps and variables. The title emphasizes speed, though the abstract excerpt does not include details on the model's architecture or experimental results.