AI training needs fewer human preferences — if you can ask smarter questions

A new algorithm cuts the number of human preference judgments needed to train AI systems by asking strategically better questions instead of randomly sampling feedback. In practice, this means companies building AI assistants that learn from human feedback could dramatically reduce the cost and time to get usable systems working.

Training AI by asking humans which of two outputs is better is expensive — a human has to sit and make thousands of pairwise judgments. This research shows that by choosing which comparisons to ask about (targeting the uncertain or informative ones) and tuning how much you trust your learned reward model, you can get the same performance with maybe half the feedback. The structural implication is simple: preference learning systems become cheaper to deploy, which means more teams can afford to build AI that's tuned to their specific human values instead of whatever large language model defaults exist.