Causes of the cross-domain redistribution pattern

Determine the causes of the broad cross-domain pattern in depthwise block-bypass response redistribution during language-model pretraining.

Background

The study finds that changes in block-bypass sensitivity often occur at similar depth positions across general, mathematical, and code evaluation domains within a training trajectory. Controlled perturbation experiments distinguish changes in local missing-update magnitude from changes in downstream response, but they do not explain why the redistribution pattern transfers across domains. The authors therefore leave the causal basis of this broad cross-domain regularity unresolved.