Machine learning can now learn better policies from old data without throwing away half the information

Researchers developed a math trick that lets AI systems learn decision-making strategies from historical data more efficiently, without the accuracy penalties that usually come with that kind of learning. In practice, this means AI trained on past behavior can now reach useful performance levels with smaller datasets or fewer iterations, which matters for domains where collecting new data is expensive or risky.

For years, the bottleneck in learning from historical data has been a tradeoff: you can either use all your data and risk bias, or throw half of it away to eliminate bias. This paper shows you don't have to choose. The result is narrower in scope than it sounds — it applies specifically to policy learning in offline settings, not all machine learning — but within that scope, it removes a real constraint. The practical effect is that expensive domains like healthcare, robotics, or scientific discovery can extract more value from their logged data.