AI can now learn reasoning from past mistakes without retraining — faster and cheaper than before

Researchers built a method that lets AI systems learn mathematical reasoning from saved problem-solving attempts instead of learning through trial-and-error. The payoff is speed and cost: the system reaches the accuracy of methods that require constant retraining, but uses a fraction of the computing power.

For three years, the only way to make large language models better at reasoning was expensive: run them live, watch them fail, adjust weights in real time. That's like teaching someone by having them solve problems while you're actively correcting them. This method changes the setup: the system learns from a saved record of what worked and what didn't, which means you can train cheaper models to reason at the level that previously required expensive live training. The structural effect is cost reduction at scale. If this method sticks, it collapses the computational overhead of reasoning-capable AI by an order of magnitude.