Brief

Two new research papers propose advancements in Bayesian deep learning, focusing on improving inference methods for neural networks. The first paper argues that sampling-based inference (SAI) has reached computational parity with optimization methods and should become the standard for uncertainty quantification. The second paper introduces a novel, scalable score-based variational inference method that avoids reparameterized sampling and can handle large-scale networks like Vision Transformers, addressing issues like mode collapsing found in other methods. AI

A new research paper questions the effectiveness of deep ensembles for uncertainty quantification in graph neural networks. The study found that ensembles offer minimal improvement over single models, with gains primarily from stabilizing predictions rather than improving uncertainty estimates. This is attributed to "epistemic collapse," where independently trained networks produce overly similar predictions, neutralizing the core advantage of ensembles. AI

A new paper argues that current methods for quantifying uncertainty in large language models (LLMs) are fundamentally flawed, likening them to unsupervised clustering algorithms. These methods primarily measure internal consistency rather than external correctness, making them unable to detect confident hallucinations. The authors advocate for a paradigm shift towards UQ methods that anchor verification in objective truth to ensure model confidence reliably reflects reality. AI