If an AI's correction examples come from another model, not its own mistakes, does practicing on them actually teach it to self-correct?

This asks whether a model can learn to fix its own mistakes from a fixed set of model-written correction examples, rather than by practicing on mistakes it makes as it trains. The short answer from the collection is: mostly no, at least not with ordinary supervised fine-tuning. The reason is more interesting than "the data is bad." Even when the correction examples look good, the mistakes in that fixed dataset aren't the mistakes the model will make once it's deployed. The model learns to fix someone else's errors, and it tends to fall into one habitual way of correcting instead of responding to what actually went wrong. What did work was online reinforcement learning over several turns. There, the model makes its own mistakes and is rewarded for actually fixing them Why does self-correction training on offline data fail?. Self-correction looks less like a skill you can copy from examples and more like one you have to practice.

That sits inside a bigger surprise in the collection: training traces may not teach what they appear to teach. Models trained on deliberately corrupted or irrelevant reasoning traces do about as well as models trained on correct ones, and sometimes generalize better. That suggests the traces act more like scaffolding for computation than like lessons in reasoning Do reasoning traces need to be semantically correct?. A related line of work finds that the steps in a reasoning trace don't reliably cause the final answer. Invalid traces often still end in correct answers, which points to learned formatting rather than working logic Do reasoning traces actually cause correct answers?. If so, a correction trace may mostly teach the model to look like it is correcting itself, which helps explain why offline correction data collapses into one fixed style.

A correct trace isn't automatically good training material either. One study found that traces which keep exploring after the answer is already settled make fine-tuning worse, even though everything in them is correct. Cutting just that unnecessary tail helped more than cutting a random chunk of the same length Does every correct chain-of-thought trace improve fine-tuning?. So what shapes learning is the structure of a trace, not only whether it is right. Choosing which traces to keep matters too: checking the model's confidence step by step catches reasoning breakdowns that an average over the whole trace hides Does step-level confidence outperform global averaging for trace filtering?.

Learning from mistakes still works in some forms. One method has the model deliberately make errors on a handful of examples in its prompt, reflect on them, and write down general principles. That improves reasoning with no fine-tuning at all Does learning from mistakes improve in-context learning?. What changes is the target: the model distills a lesson from its own errors instead of copying someone else's fixes. Self-generated training data also works when something trustworthy checks it. LLM judges given reference answers, for example, can supervise self-improvement about as well as a trained reward model Can reference examples make LLM judges reliable enough for self-improvement?.

The takeaway you might not have expected: whether correction training works depends less on how good the examples are and more on whose mistakes they are. Teaching a model to fix errors works best when the errors are its own, encountered while it practices. Online training has its own failure modes, though. Problems that are nearly impossible can reward lucky shortcuts instead of real reasoning Do overly hard RLVR samples actually harm model capabilities?. And reinforcement learning tends to narrow a model toward one dominant output style Does RL training collapse format diversity in pretrained models?, so collapsing into a single correction style isn't a problem only offline training has.