AI reasoning models trained on wrong answers can now learn to ignore them

A new training method helps large language models learn to reason correctly even when given deliberately incorrect training examples. The method works by having the model gradually filter out bad data as it gets better, rather than treating all training examples as equally reliable — which means you can use cheaper, messier training data without tanking performance.

Everyone building reasoning AI has the same problem: getting enough correct answers to train on is expensive, so you end up mixing in wrong answers just to have more data. This paper shows a concrete mechanism for that to work anyway. The practical effect is straightforward — it cuts the cost of labeling training data if you can tolerate some noise in what you're feeding the model. The catch is real: the model has to be good enough first to recognize what's wrong, which means early training still needs clean examples.