worldparticle

WorldParticle is an early code implementation of a research paper of the same name, which has the longer title Unified Simulation of Lagrangian Particle Dynamics via Transformer. The paper was posted on arXiv in 2026 by a team of authors led by Caoliwen Wang, with contributors from a range of universities and labs. In physics simulations, the Lagrangian view tracks individual particles as they move through space, in contrast to fixed grid based approaches. The paper's claim, as reflected in its title, is that a single transformer model can be used to simulate this kind of particle motion across different scenarios in a unified way.

This repository is the work of an independent author who publishes on GitHub as lucidrains and is known for re-implementing recent machine learning papers in Python and PyTorch. The repo is explicitly marked as work in progress at the top of the README, with the abbreviation wip in the heading. There is one figure image referenced at the top, and the rest of the README is dedicated to two citation blocks in BibTeX format.

The README does not contain installation instructions, usage examples, command line flags, model checkpoints, training data sources, or evaluation results. Beyond the title of the paper being implemented and one secondary citation for a method called Long Context Pre-Training with Lighthouse Attention, there is no further description of what the code does or how to run it. Readers who want to understand the method itself need to read the linked arXiv paper directly. The repository is small, has fifteen stars, and the listed language is Python.