LBDTPP: Semi-autoregressive framework beats event sequence generation limits

Researchers from Chinese Academy of Sciences, City University of Hong Kong, Tsinghua University, and other institutions have introduced LBDTPP (Latent Block-Diffusion Temporal Point Processes), a novel framework for generating asynchronous event sequences. Existing autoregressive methods suffer from error accumulation during multi-step generation, while non-autoregressive diffusion models are typically limited to fixed-length outputs. LBDTPP bridges this gap by defining an autoregressive probability distribution over event blocks in latent space and performing Gaussian diffusion within each block. This semi-autoregressive approach preserves the length flexibility of autoregressive TPPs while inheriting the parallel high-quality generation capability of diffusion models.

The team derived Wasserstein error bounds showing that block-wise generation reduces error accumulation compared to event-wise autoregressive generation under suitable assumptions. Extensive experiments on six real-world benchmark datasets demonstrated that LBDTPP outperforms state-of-the-art TPP baselines in both unconditional and conditional generation tasks. Further empirical analyses confirmed benefits of latent-space diffusion and block-wise generation, revealing a trade-off between generation quality and block size. The code is publicly available on GitHub, enabling researchers to apply this framework to domains like social network analysis, medical diagnosis, and financial transaction modeling.