New test reveals generative models often get time wrong

A team led by Kiwan Kwon and colleagues from several Korean universities (with Vamsi K. Potluru from outside Korea) published a study on arXiv (July 2026) exposing a major blind spot in how synthetic sequential tabular data is evaluated. Generative models—used for privacy-preserving sharing of time-series data like financial records, patient histories, or sensor logs—can produce timestamps that go backwards, repeat, or create trajectories no real entity followed. Conventional evaluation methods pool records into static distributions, ignoring these temporal anomalies entirely.

The researchers introduce a new evaluation protocol that first characterizes each dataset along four properties: how time is represented, whether observations are regularly sampled, whether trajectories are mutually dependent, and how the schema links entities to their histories. Based on this taxonomy, the protocol measures timestamp validity, cross-sectional structure at aligned time points, within-entity dynamics, and time-varying relational structure. Testing 8 generative models on 13 datasets across 6 domains, they found that rankings under conventional evaluation disagree substantially with those from the temporal protocol, and failures are architecture-coherent rather than random. The work underscores that temporal fidelity must be measured directly, not inferred from pooled record distributions.