Training-Free Bayesian Filtering with Generative Emulators 🌍

This repository contains the implementation of the GenCast experiment from the paper Training-Free Bayesian Filtering with Generative Emulators by Thomas Savary, François Rozet, and Gilles Louppe, published at the International Conference on Machine Learning (ICML) in 2026.

Code

The algorithm described in the paper is implemented in the filtering folder. In particular, it contains the following files:

• filtering/wrapper/denoisers.py that implements an MMPS denoiser using the "basic" GenCast denoiser to draw samples from an approximation of the optimal proposal distribution $q(x_{k+1} \mid x^{k}, y^{k+1}) = p(x_{k+1} \mid x^{k}, y^{k+1})$.
• filtering/fa_apf.py that implements the Fully Adapated Auxiliary Particle Filter (FA-APF) with covariance inflation to control the degeneracy of the weights.

To do other experiments, users can modify the configuration files in the config folder, as well as observations parameters (mask, covariance, ...) in the data/observations folder.

Model and data

Our work build on GenCast, a diffusion-based emulator of the atmosphere developed by Google. GenCast's denoisers were trained on ERA5, a global atmospheric reanalysis dataset covering the period from 1940 to present and produced by the ECMWF (the European Centre for Medium-Range Weather Forecasts).

Citation

If you find this work useful in your research, please consider citing:

@inproceedings{
savary2026trainingfree,
title={Training-Free Bayesian Filtering with Generative Emulators},
author={Thomas Savary and Fran{\c{c}}ois Rozet and Gilles Louppe},
booktitle={Forty-third International Conference on Machine Learning},
year={2026},
url={https://openreview.net/forum?id=ibcZNZwKfZ}
}