ArviZ: a modular and flexible library for exploratory analysis of Bayesian models

When working with Bayesian models, a range of related tasks must be addressed beyond inference itself.These include diagnosing the quality of Markov chain Monte Carlo (MCMC) samples, model criticism, model comparison, etc.We collectively refer to these activities as exploratory analysis of Bayesian models.In this work, we present a redesigned version of ArviZ, a Python package for exploratory analysis of Bayesian models (EABM).The redesign emphasizes greater user control and modularity.This redesign delivers a more flexible and efficient toolkit for exploratory analysis of Bayesian models.With its renewed focus on modularity and usability, ArviZ is well-positioned to remain an essential tool for Bayesian modelers in both research and applied settings. Statement of needProbabilistic programming has emerged as a powerful paradigm for statistical modeling, accompanied by a growing ecosystem of tools for model specification and inference.Effective modeling requires robust support for uncertainty visualization, sampling diagnostics, model comparison, and model checking (Gelman et al., 2020;Guo et al., 2024;Martin, 2024).ArviZ addresses this gap by providing a unified, backend-agnostic library to perform these tasks.The original ArviZ paper (Kumar et al., 2019) described the landscape of probabilistic programming tools at the time and the need for a unified, backend-agnostic library for exploratory analysis -a need that has only grown as the ecosystem has expanded.The methods implemented in ArviZ are grounded in well-established statistical principles and provide robust, interpretable diagnostics and visualizations (Dimitriadis et al., 2021; Gelman et al., 2019;Kallioinen et al., 2023;Paananen et al., 2021;Padilla et al., 2021;Säilynoja et al., 2022Säilynoja et al., , 2025;;Vehtari et al., 2017Vehtari et al., , 2021)).Modern Bayesian practice is a rapidly advancing field in which new methodological developments continually extend the range and complexity of models that can be fit in practice.For instance, the methods to compute key ArviZ features such as ess, rhat, loo or compare have been improved between 2019 and now, and new implementations needed significant development effort to adapt to because it wasn't possible to change a part of ArviZ without also adapting everything that interacted with it.The redesign addresses these challenges by modularizing the codebase, allowing individual components to be updated or replaced without affecting the entire system.This modularity not only facilitates maintenance and updates but also encourages community contributions, as developers can focus on specific components without needing to understand the entire codebase.