Researchers have developed a new method for message passing in variational inference called natural-gradient message passing (NGMP). This approach is based on the Bethe free energy and involves constraining an edge marginal to an exponential family. At a stationary point, the natural parameter of that edge equals the sum of two projected messages from incident factors. NGMP is more accurate than traditional variational message passing when uncertainty persists in non-conjugate factors or parameters are filtered through successive data batches. --- Why it matters: This matters to AI researchers because it provides a more accurate method for handling uncertainty and non-conjugate factors, which can improve the performance of models such as Poisson smoothing, heteroskedastic regression, and hourly ETTh forecasting. Source: https://arxiv.org/abs/2608.15922