Bayesian model averaging often leaves analysts with a long list of plausible models, each carrying some posterior weight. That list can hide the fact that many models share the same substantive conclusion, and standard marginal summaries do not reveal which configurations are jointly supported. The authors frame the problem as one of finding a compact set of interpretable regions, where each region groups models that share a specified pattern, while retaining a prescribed amount of posterior mass.

To do this, they introduce Adaptive Regional Selection (ARS). The algorithm alternates between exploring the model space and estimating regional weights, and it offers both conservative and optimistic selection procedures. These two modes identify a summary that meets the user-supplied mass target and bound how much further the summary's complexity could be improved.

On the theory side, the paper establishes guarantees for retained mass, total-variation error, and suboptimality, and gives conditions under which optimality can be certified. A notable result is that, when within-region weight variation is bounded, the number of samples needed for regional estimation does not have to grow with the number of represented models, even under adaptive sampling and stopping.

Experiments with Bayesian model averaging and Bayesian networks illustrate the approach, examining retained mass, representation complexity, and certification. The method is accompanied by code and data.