A popular technique for teaching AI models to reason better turns out to be grading on a test it cannot actually read.
Researchers studied on-policy self-distillation, a method where a teacher model that already knows the correct answer scores a student model's own reasoning attempt, then feeds that score back into training. Using reasoning traces with known error locations, they checked two signals the method relies on, the full-distribution KL and the token log-ratio, to see whether either one actually flagged where the student went wrong. Neither signal localized the errors, neither pushed the model away from the specific wrong tokens, and neither could reliably tell a correct rollout from an incorrect one on the same problem. The same teacher model did much better at finding errors once it was first asked to write a critique of the rollout before scoring it.
That gap matters because self-distillation was pitched as a fix for the weak, all-or-nothing feedback of standard reinforcement learning on verifiable rewards. If the scoring signal cannot locate mistakes, it is not delivering the dense, corrective credit it promises, just a fuzzier version of the same pass-or-fail judgment. The critique-conditioned version changes that: the extra critique step shifted probability toward words like Wait, But, and Actually, the kind of self-correcting language that signals a model catching its own mistake, and it helped the model solve math problems it could not otherwise solve by sampling alone.
It is a reminder that asking a model to check its own work is different from asking it to explain how it checked, and only one of those actually teaches it anything.