ReDiTT: Retrieval Augmented Conditional Diffusion Transformers for Asynchronous Time Series

Asynchronous time series—data where events occur at irregular intervals—are notoriously difficult to forecast because they combine continuous time and discrete event types. Standard autoregressive models often suffer from error accumulation over long horizons, while existing diffusion models struggle to maintain trajectory-specific structure due to a lack of informative global conditioning signals. This paper asks: can we improve long-horizon forecasting by grounding diffusion-based generation in retrieved, historically similar temporal dynamics?

The authors introduce ReDiTT (Retrieval Augmented Conditional Diffusion Transformers). The framework first maps event sequences into a latent space using a pretrained Variational Autoencoder (VAE). During both training and inference, the model retrieves the top-k nearest neighbor sequences from a pre-constructed latent memory bank. These retrieved sequences are then injected into the diffusion transformer blocks via cross-attention. This allows the model to use the retrieved references as structural guidance, effectively anchoring the generation process to concrete, historically observed temporal patterns rather than relying solely on learned parameters.

ReDiTT achieves state-of-the-art performance across seven real-world datasets for both next-event prediction and long-horizon forecasting. By conditioning on retrieved latent priors, the model stabilizes long-term predictions and significantly improves sample diversity. The authors demonstrate that this retrieval-based conditioning is robust and provides a more effective way to guide generation than traditional global conditioning signals, which often lack the specificity required for complex, non-stationary event streams.

This work addresses a fundamental limitation in generative modeling for irregular event data: the tendency of models to drift toward generic, high-probability patterns over long horizons. By combining the generative power of diffusion models with the non-parametric memory of retrieval, ReDiTT provides a scalable way to incorporate historical context into forecasting, making it highly applicable to fields like finance, healthcare, and user behavior modeling where long-term structural consistency is critical.