A preprint by Rafael Mouallem Rosa, Julyan Arbel, and Hien Duy Nguyen introduces "sources of uncertainty" (SoU) inference, a general framework that relaxes the usual Bayesian assumption of uniform treatment across a model. Instead of applying a single probabilistic treatment to all parameters, SoU allows researchers to explicitly set different confidence levels for different components, controlling how much uncertainty each component retains and how that uncertainty propagates through the rest of the model.
The authors show that empirical Bayes and profiling-type inference emerge as limiting cases of this selective treatment. Because confidence operates through a mechanism separate from prior informativeness, the two can be tuned independently, adding a new degree of freedom in Bayesian model design. This separation is what makes the framework useful for sparse estimation: by selectively shrinking certain components, the method induces sparsity in linear and logistic regression, and in Bayesian neural networks it enables self-pruning of unnecessary units.
The paper is a single methodological contribution, so there are no conflicting sources to compare. As a preprint, it has not yet undergone peer review, but it includes 42 pages of derivations, simulations, and real-data illustrations, with code links provided via the arXiv listing. The authors position SoU as a bridge between full Bayesian inference and more pragmatic profiling approaches, offering a flexible middle ground for modelers who want selective uncertainty control.