Study finds random matrix theory metrics for neural network layers can be arbitrary
An arXiv paper argues that data-free metrics based on random matrix theory do not reliably reflect model quality, because reparameterised networks can keep performance while metrics change.
In plain words
Some analysis methods score neural network layers without looking at data, using random matrix theory, and these scores are used to judge model quality. The paper shows a network's function can stay the same while its scores change arbitrarily, so the scores may not measure what they claim.
What was announced
The authors, in a cross-posted arXiv abstract, examine data-free metrics that random matrix theory (RMT) provides for rating layers of pre-trained neural networks. Such metrics are promoted for predicting model quality layer by layer, spotting pathologies, and guiding pre-training optimisation and post-training compression, with claims of a special link to generalisation. The paper's central finding is that these RMT-based metrics are unrelated to performance or training: functionally preserving reparameterisations of a pre-trained model can carry arbitrary metric values, a result the authors report across a range of architectures and scales.
The construction relies on criticality, described as a well-established phenomenon in which some layers, called robust, can be re-initialised or re-randomised without changing the model's functional behaviour, while others, called critical, cannot. Re-initialising or re-randomising robust layers yields functionally preserving reparameterisations for which the RMT metrics become arbitrary. The authors also report that relationships between metrics offered by RMT are spuriously related to generalisation, and that when many pre-trained models show data-free metrics in a 'good' range, this depends in part on model initialisation. A project page accompanies the paper.
What it does not show
- The evidence is the paper's abstract only; the full article, including methods, experiments and peer review status, is not covered, and no independent confirmation of the findings is available.
Sources
Source-support review completed 14 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 14 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.