Student's t Likelihood Improves Bayesian Neural Network Performance

A new arXiv preprint (2607.25376) investigates the impact of likelihood distribution choice on Bayesian neural networks (BNNs). While variational inference with the evidence lower bound (ELBO) is standard, most implementations assume Gaussian distributions for prior, posterior, and likelihood. The authors argue this assumption may cause model misspecification. They propose using a Student's t distribution for the likelihood, which has heavier tails and can better handle outliers. Experiments show that this alternative improves predictive performance and uncertainty quantification on several benchmarks. The work fills a gap by focusing on likelihood distribution choice, whereas prior studies mainly examined priors.