A new model generates fake time-series data all at once, not one step at a time, sidestepping a compounding-error problem that plagues the current leading approach.
Researchers built SDFlow to generate synthetic time-series data, the kind used to stand in for sensor readings, stock prices, or health-monitoring logs. Most current systems compress a sequence into discrete tokens and predict them one after another, the same way a text model predicts the next word. Small errors early in that process snowball into bigger ones later, a problem called exposure bias. SDFlow skips the step-by-step guessing entirely. It generates a whole sequence in parallel using flow matching, a technique that maps a simple starting distribution directly onto the target data instead of predicting it piece by piece.
That matters because long sequences are exactly where token-by-token models fall apart, and synthetic time series increasingly trains forecasting and anomaly-detection systems when real data is scarce or sensitive. The paper reports state-of-the-art scores on standard benchmarks, notably stronger results on long sequences, and faster inference than the autoregressive baselines it replaces.
This is a familiar pattern. Flow-based and diffusion methods already displaced autoregression in image and audio generation, and the same argument is now showing up for structured data like time series. The benchmark wins come from the paper's own authors, so the real test is independent replication.