Verify the severity pattern across all corruption types

Determine whether the reduced ZFGA benefit at extreme corruption severity, observed for Gaussian noise on ResNet-50, also holds for the other six CIFAR-10-C corruption types.

Background

The paper observes an inverted-U pattern for ZFGA under Gaussian noise on ResNet-50: the gain increases through severity 3 and decreases at severity 5. This is interpreted as evidence that extreme corruption may destroy information rather than produce geometric distortions that Fisher alignment can correct.

The observation is based on only one corruption type, and the authors explicitly state that they have not verified whether the same severity dependence applies to the remaining CIFAR-10-C corruptions.

References

We have not verified whether this severity pattern holds for the other six corruption types in our CIFAR-10-C evaluation, and as noted in Appendix~\ref{sec:appendix_ablations}, the severity-3 Gaussian-noise number is not directly comparable to the multi-corruption headline result in Table~\ref{tab:main_results}.

— Technical note on: Zero-Training Feature-Space Alignment via Information Geometry  (2609.37302 - Khan et al., 29 Sep 2026) in Appendix, Section “Extended Analysis: When Does ZFGA Help?”