AI training breakthrough: models can now learn from problems they can't yet solve

Researchers found a way to teach AI language models to solve hard reasoning problems by temporarily reformatting them into easier versions, then gradually increasing difficulty as the model learns. This removes a major bottleneck in AI training: previously, if a problem was too hard for the current model, it produced zero learning signal and the model couldn't improve on it.

For years, training language models on complex reasoning hit a wall: hard problems the model couldn't solve provided no gradient to learn from, creating a dead zone where the model got stuck. This method breaks that wall by disguising hard problems in simpler formats first, letting the model build up capability incrementally. The practical effect is measurable: on difficult math and logic tasks, both Qwen and Llama models improved by 8-10% on problems that previously generated no learning signal at all.