Direct Revision (critique-free ablation)
Constitutional AI: Harmlessness from AI Feedback — introduced
An ablation of the SL-CAI pipeline that skips the critique step and asks the model to generate a revision directly, compared against critiqued revisions to test whether the critique step is necessary.
Is the critique step actually necessary, or would a model do just as well revising its harmful response directly, with no critique in between? This page covers the ablation Constitutional AI ran to find out.
It walks through the test itself and the paper's reasoning for keeping the critique step regardless of what the numbers showed, then connects back to critique and revision, the process it is testing.
Testing whether the critique step earns its keep
Direct revision is an ablation of Critique and Revision that deletes the middle step: instead of critiquing a harmful response and then revising it, the model is instructed to produce a revision directly, with no critique in between. Comparing preference-model harmlessness scores between critiqued and direct revisions shows a scale-dependent picture: at smaller model sizes, critiqued revisions score meaningfully more harmless, while at the largest scale tested (52B parameters) the gap "made no noticeable different," and inspection of the critiques themselves found they "often made inaccurate or overstated criticisms" (constitutional-ai, §"3.5 Are Critiques Necessary?", p. 10).
Why the paper keeps the step anyway
Despite the weak effect at scale, the paper uses critiqued revisions throughout its main results, "as it may provide more transparency into the model's reasoning process," adding that such reasoning "may also be useful to help models uncover more subtle harms" (constitutional-ai, §"3.5 Are Critiques Necessary?", p. 10). This makes direct-revision-ablation, captured in the Direct Revision (critique-free ablation) — tests the necessity of critique in → Critique and Revision edge, one instance of a broader pattern in the corpus: a design choice can fail its most obvious quantitative benchmark and still be retained, because its value lies in legibility rather than score. The The same technique, cited or deployed twice, doing different work each time connective theme groups this alongside Principle Ensembling, another case where a technique's payoff shows up somewhere other than where it was expected.