Scale calibration learns an update for one representation of a pretrained model. A new arXiv preprint asks how much of that update can be reused when the same model is accessed through a different representation. The authors study this transfer problem directly, rather than treating calibration as a one-off fix.
The proposed transfer procedure extracts positive group-wise gains from the original update. This suggests the method is designed to preserve the useful parts of the calibration signal while adapting it to a new representation. The abstract does not yet state the magnitude of the gains or the conditions under which transfer works best.
The work is relevant to anyone who reuses a pretrained model across different feature spaces or downstream tasks, since it may reduce the need to recalibrate from scratch.