An Isotropic Approach to Efficient Uncertainty Quantification with Gradient Norms

Nils Grünefeld, Jes Frellsen, Christian Hardmeier · 2026

Read An Isotropic Approach to Efficient Uncertainty Quantification with Gradient Norms by Nils Grünefeld, Jes Frellsen, Christian Hardmeier (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-29466.txt.

Description

Grünefeld and co-authors give a cheap way to gauge neural network doubt. Exact methods cost too much for large models. Their trick uses two shortcuts. First, a Taylor step ties the doubt to the slope of the output. Second, they treat weight spread as equal in all directions. One pass yields the model-doubt part from a gradient norm. The data-noise part comes from the prediction itself. Checks against costly sampling matched well. On quiz tests, the signal helped on TruthfulQA but not TriviaQA.

Catalog record

CREATORNils Grünefeld, Jes Frellsen, Christian Hardmeier
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE67316
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.29466

Browse the full library · About the Tower of Babel