RIFT: Rectified log-Intensity Flows for Temporal Point Processes

Generative modeling of temporal point processes (TPPs) has been extensively studied. Existing approaches—autoregressive models and diffusion-based models—typically operate on individual event times, which exposes them to error accumulation and limits their ability to capture global dependencies across the sequence. We propose RIFT (Rectified log-Intensity Flow TPPs), a generative framework that directly learns a velocity field over log-intensity functions, transforming a tractable base intensity (a homogeneous Poisson process) into the target intensity via a continuous ODE. Operating in log space guarantees that intermediate intensities remain non-negative throughout the transport path. A straightness loss further encourages the learned trajectories to be straight in log-intensity space, reducing the number of ODE steps required at inference from the 100 steps typical of diffusion-based methods to as few as K = 5. Experiments on 13 benchmark datasets demonstrate that RIFT achieves competitive log-likelihood while offering substantial improvements in sampling efficiency over autoregressive and diffusion-based baselines.