Structured pruning usually asks which model dimensions to remove, and it typically does this in the model's original representation. But the effect of deleting a coordinate depends on the basis in which the deletion happens. That is the starting point for WUSPruner, a new method described in a recent arXiv paper.
The paper proposes using second-moment-optimal transforms to change the basis before pruning. By transforming the representation, the method aims to make coordinate deletion more principled than simply cutting dimensions in the original coordinate system. The abstract does not give full experimental details, but the core idea is that the choice of basis is a key part of the pruning problem, not a fixed background.
If the approach holds up, it could shift how structured pruning is framed: instead of only deciding which dimensions to remove, you also decide how to represent the model first. The paper is available on arXiv under identifier 2610.15394.