Teaching an AI to grade its own reasoning with a checklist beats giving it one right answer

When you train an AI to reason, you usually do one of two things. Either you feed it a worked-out example to copy, or you let it try problems and give it a single thumbs-up or thumbs-down at the end. Both have a flaw.

The worked example is risky because those step-by-step explanations are expensive to write and often sloppy. Even when the final answer is right, a muddled explanation can teach the AI bad habits. The thumbs-up approach is worse in a different way: one number tells the AI "that was good" but never says which part was good. If the answer took twelve steps and three were wrong, the AI gets no map of where it went off course.

This paper's fix is a grading checklist — a rubric. Think of how a teacher marks an essay: not just a final grade, but points for "clear thesis," "evidence," "logic." Here, the AI first writes a checklist of what a strong answer to each problem needs. Then a "teacher" copy of the model, holding that checklist in hand, watches the "student" copy solve the problem and nudges it word by word toward meeting each criterion. The student learns from its own attempts, graded against specific standards, instead of mimicking one possibly-flawed example.

The gains are modest: on science reasoning tests, it beat two standard training methods by about one point each. And the rubrics are generated by the same kind of model, so a bad checklist could still mislead.

Why you should care: The detailed, per-step feedback this method squeezes out — instead of one final score — is what lets an AI fix the exact reasoning step it botched, not just guess.