Latent Block-Diffusion Temporal Point Processes

Pytorch implementation of the paper "Latent Block-Diffusion Temporal Point Processes: A Semi-Autoregressive Framework for Asynchronous Event Sequence Generation".

Paper: arXiv:2606.24982

Overview

LBDTPP is a semi-autoregressive framework for asynchronous event sequence generation:

• Inter-block autoregression: event blocks are generated autoregressively, preserving temporal dependencies and supporting variable-length generation;
• Intra-block diffusion: multiple event representations are sampled in parallel through Gaussian diffusion in a continuous latent space, reducing event-by-event error accumulation.

This design combines the length flexibility of autoregressive TPPs with the high-quality parallel generation capability of non-autoregressive diffusion models.

Installation

1. Install the dependencies

conda create --name bdtpp python=3.10
conda activate bdtpp
pip install -r requirements.txt

2. Unzip the data

unzip data.zip

Dataset

The six real-world datasets are from CDiff.

Running Experiments

We provide runnable experiment scripts:

• exp_conditional.sh: commands for conditional generation
• exp_unconditional.sh: commands for unconditional generation

Citation

If you find this repository useful, please cite:

@article{zhang2026latent,
  title={Latent Block-Diffusion Temporal Point Processes: A Semi-Autoregressive Framework for Asynchronous Event Sequence Generation},
  author={Zhang, Shuai and Chen, Yancheng and Zhou, Chuan and Liu, Yang and Lin, Xixun and Zhao, Xiangyu and Zhu, Jun and Ma, Zhi-Ming},
  journal={arXiv preprint arXiv:2606.24982},
  year={2026}
}