ENFR: Evidence Networks with Flow-Matched References for Amortized Bayesian Evidence Estimation

Forum topic · 小凯 · 2026-10-20

Summary

ENFR (Evidence Networks with Flow-matched References) is a new machine-learning method for estimating the Bayesian evidence, Z = p(x|M), of a data-generating model. The evidence supports model inference, model misspecification tests, and goodness-of-fit assessment as an alternative to standard discrepancy statistics such as chi-squared, but it is notoriously hard to compute even when the likelihood is available. Building on the evidence-network framework of Jeffrey & Wandelt (2024), ENFR first learns Bayes factors relative to a flow-matching reference and then recovers absolute evidence values, unlike prior learned approaches that only measure evidence ratios. The authors show improved evidence estimation over flow-based baselines in all applications and achieve state-of-the-art performance among density-estimation methods on benchmarks. They also introduce an architecture that jointly learns the evidence for an ensemble of models. Results are demonstrated across multiple scientific datasets, with the method fully amortized, accurate, and fast. Paper: arXiv:2610.23825, published 2026-10-16 by Kai Lehman et al.

Paper Overview

Field: Machine Learning
Authors: Kai Lehman, Jameson Dong, Carolina Cuesta-Lazaro, Niall Jeffrey, Benjamin D. Wandelt, Francisco Villaescusa-Navarro, Sven Krippendorf, Jochen Weller
Published: 2026-10-16
arXiv: 2610.23825

Abstract (translated)

The evidence, $\mathcal{Z} = p(x | \mathcal{M})$ , of a data-generating model $\mathcal{M}$ is a powerful tool in Bayesian frameworks: it enables model inference, model misspecification tests, and goodness-of-fit assessment as an alternative to standard discrepancy statistics, e.g. $\chi^2$ . While versatile, the evidence is notoriously difficult to compute, even when the likelihood is available.

In both likelihood-based and simulation-based inference, estimating $\mathcal{Z}$ provides an immediate solution to high-dimensional goodness-of-fit or model misspecification tests: using $\mathcal{Z}(x)$ as the test statistic induces a minimum-volume level- $\alpha$ test. Such tests are essential to assess model fidelity, which is necessary to trust simulation-based parameter inference. Existing learned approaches mainly target Bayes factors and can only measure evidence ratios. In contrast, this work provides accurate, fast, and amortized estimates of the absolute evidence.

The authors propose ENFR: Evidence Networks with Flow-matched References. Building on the evidence network framework of Jeffrey & Wandelt (2024), the method first learns Bayes factors associated with a flow-matching reference, from which absolute evidence is recovered.

Key Results

Improved evidence ( $\mathcal{Z}$ ) estimation over flow-based baselines in all applications.
Best performance among all density-estimation methods on benchmarks.
A new architecture introduced and demonstrated for jointly learning the evidence of an ensemble of models $\{\mathcal{M}_i\}$ .
Performance and scaling behavior demonstrated across a variety of scientific application datasets.

--- *Auto-collected on 2026-10-20*