Vague AI feedback is a design choice. The CriticGen paper shows a loop to end it.

AI feedback has a uselessness problem. A grader that hands back "this is wrong" with no hint of how to fix it teaches nothing, and the field of LLM-as-judge has spent years producing exactly that kind of verdict. Evaluation has been decoupled from generation: the critic scores, then walks away.

The CriticGen paper reframes the loop. Instead of a score and a one-line critique, the authors generate a custom rubric for each answer and pair every critique with an executable fix-it plan and a revised draft. The rubric and the rewrite come from the same generator, so the feedback is concrete enough to act on.

On the paper's own benchmark, 73.17% of answers came out the other side measurably better, with a 93.28% non-degradation rate: almost no rewrites made things worse. Two things this evidence does not yet settle: it is a single arXiv preprint, scored on the authors' own benchmark, with no head-to-head against RLAIF or reward models, and no independent replication. The mechanism is the reusable category, not the 73%. That number lives or dies on the next paper. Vague AI feedback is a design choice, and a rubric-conditioned rewrite loop is what "actionable" actually looks like.