Relative Repairability: A Calibration-Based Diagnostic for High-Sparsity Post-Pruning Allocation

Qishi Zhan, Liang He, Minxuan Hu, Ziheng Chen, [4pt] Marquette University, [2pt] Corresponding authors · 2026

Read Relative Repairability: A Calibration-Based Diagnostic for High-Sparsity Post-Pruning Allocation by Qishi Zhan, Liang He, Minxuan Hu, Ziheng Chen, [4pt] Marquette University, [2pt] Corresponding authors (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-2605-25508.txt.

Description

Qishi Zhan, Liang He, Minxuan Hu, and Ziheng Chen study where pruning damage lands at high sparsity and whether repair can recover it. They introduce Relative Repairability, a diagnostic weighing raw damage from pruning with residual damage after variance-matching repair, using only unlabeled tuning data. On CIFAR10 and CIFAR100 with ResNet and VGG variants, it is not always dominant. It helps most near a recoverability transition, where standard priors lose reliability but recovery has not collapsed. There it beats ERK mid-band and LAMP near the top.

Catalog record

CREATORQishi Zhan, Liang He, Minxuan Hu, Ziheng Chen, [4pt] Marquette University, [2pt] Corresponding authors
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE77401
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.25508

Browse the full library · About the Tower of Babel