An Isotropic Approach to Efficient Uncertainty Quantification with Gradient Norms

ArXi:2603.29466v1 Announce Type: cross Existing methods for quantifying predictive uncertainty in neural networks are either computationally intractable for large language models or require access to