Direct Revision (critique-free ablation) — tests the necessity of critique in → Critique and Revision

direct-revision-ablation asks a pointed question about critique-and-revision’s own design: is the critique step actually doing anything, or is it there for show? The answer is scale-dependent. At smaller model sizes, critiqued revisions score meaningfully more harmless than revisions generated directly with no critique in between; at the largest scale tested (52B), the gap nearly closes, and the ablation’s own inspection found the critiques "often made inaccurate or overstated criticisms" (constitutional-ai, §"3.5 Are Critiques Necessary?", p. 10). Despite this weak large-model effect, the paper keeps critiqued revisions for all main results anyway, explicitly for transparency: the critique text gives a human-readable record of what the model believed was wrong with its own response, which a bare revision does not provide.