ChronoSSM Improves Temporal Awareness in State Space Models.

Modern sequence models, including Transformers and State Space Models (SSMs), excel at generative modeling but typically prioritize predicting "what" happens, treating "when" it happens as secondary. In data-mining scenarios where events have explicit timing information, this separation can hinder temporal reasoning, anomaly detection, and accurate reconstruction of event chronologies. A common practice is to use a two-stage approach, where a separate timing model is trained on representations learned solely for event prediction, assuming these representations already contain sufficient temporal structure. Researchers introduce ChronoSSM, an autoregressive State Space Model designed to jointly model both events and their timestamps using a shared backbone. This model is trained with combined objectives for both token and temporal generation. The study compares this joint training regime, where temporal supervision directly updates the backbone, with the traditional two-stage approach where timing is learned from frozen event representations. Across four diverse domains with varying timestamp supervision densities, joint training consistently made inter-arrival information more recoverable from the learned representations. Crucially, this improvement in temporal informativeness was achieved without any systematic degradation in the quality of content generation. The results demonstrate that direct temporal supervision can lead to more temporally aware representations without compromising the primary task of autoregressive event modeling.