I've found that when conducting FLIP simulations, using motion to move the collision object results in very good outcomes. However, if I use a deforming object for collisions, the results become very poor. Can anyone explain the underlying principles of using motion in this context?

Do you mean the performance becomes poor? The motion node doesn't support deforming geometry, only transformation. That's why it's faster. If your collision object's mesh isn't deforming, use motion, if it is deforming, use a FileCache to pre-convert each frame to a VDB for better performance rather than having to convert every frame on-the-fly. That's why the Collision object shelf tool generates the node.