An ablation removes one component to measure what it contributed.
Without it you cannot tell which part of a method drives the gains.
Often only one or two elements of a complex method matter at all.
Reviewers expect ablations for exactly this reason.
They frequently reveal that the headline idea was not the important one.
An ablation removes one component and measures what happens.
The word comes from surgery, where ablation means removing tissue.
It answers which part of a method actually produces the gain.
A new method usually bundles several changes together.
A new architecture, a new loss, a new schedule and more data all arrive at once.
The headline number tells you the bundle helped and nothing about which part.
Ablating each in turn separates them.
The result is frequently that one change did nearly all the work.
Sometimes the idea named in the title contributes almost nothing.
That is not a failure of the study but the point of doing it.
Reviewers expect ablations, and their absence is a reasonable criticism.
They also guide practice, since knowing what matters tells you what to keep.
Ablate against a fair baseline, not against a deliberately weakened version.
Run each with several seeds, since ablation differences are often small.
The same discipline applies outside research, when a system improves and nobody knows why.
One question on this card. Every option explains itself once you pick, right or wrong.
11 concepts in this module, then a graded test with every answer explained.