Bayesian evidence, the probability of data given a model, is a powerful tool for model comparison and misspecification testing, but it is notoriously difficult to compute. A new paper introduces ENFR (Evidence Networks with Flow-matched References), an amortized method that estimates absolute evidence rather than just Bayes factors, which only give evidence ratios.

The approach builds on the Evidence Network framework by first learning Bayes factors with respect to a flow-matched reference distribution, then recovering the absolute evidence from that ratio. The authors also introduce and prove an architecture that jointly learns the individual evidences of a set of models, improving scalability.

In benchmarks across a variety of scientific applications, ENFR shows improved evidence estimation over the flow-based baseline and achieves the best performance among all density estimation methods considered. This makes it a practical tool for high-dimensional goodness-of-fit and model fidelity tests, which are essential for trusting simulation-based parameter inference.