Purplelink
← Why it matters

Manuscript red-team reference · companion to Paper Review

missing ablations

Claims without ablating each architectural choice.

What it is

A method paper that proposes 5 architectural changes and reports the combined effect tells you nothing about which changes mattered. Maybe 4 of them were noise and only 1 drove the gain — or maybe two changes interact destructively and you'd be better off with just one.

Why a reviewer cares

Reviewers ask: which component caused the improvement? Each non-trivial design choice should have an ablation isolating its contribution. The pattern 'we added X, Y, and Z and beat the baseline' is incomplete evidence.

How to fix it

For each architectural or training-recipe change, run an ablation: same model with that one change removed. Report the per-ablation effect. Components that don't survive ablation should either be removed from the paper or kept with the honest framing that they didn't help.


This is one of ~15 canonical methodology explainers Paper Review's red-team report links to. To get a full review of your manuscript, start a Paper Review — $5.