Teaching AI to solve math by finding problems it can copy from

When you're stuck on a hard math problem, you don't look for one that sounds similar. You look for one you can solve the same way. That's the core idea here.

AI models often pull in outside examples before answering, a trick called "retrieval." But they usually grab examples that look alike on the surface — same words, same topic. The problem: two questions can read almost identically but need totally different tricks to crack, while two that look nothing alike share the same hidden method.

The authors built a system that hunts for examples by usefulness, not surface resemblance. It trains a search tool to rank past problems by how much their solution steps actually help solve the new one. Then it lets the model practice with those genuinely useful worked examples, rewarding it only when it reaches the correct, checkable answer.

The payoff: on a hard high-school math competition test, the upgraded smaller model scored about 7 points higher than the standard training method, and the larger model about 3 points higher. The team also found that good "by-usefulness" search digs up several different solution strategies for the same problem, giving the model more than one way in.

The honest catch: the gains were measured only on math problems, where answers can be automatically checked as right or wrong. It's unclear whether the same approach helps with messier tasks — writing, planning, or anything without a clean correct answer. The improvement was also a few points, not a leap.

Why you should care: The AI tools you'll lean on for tutoring, coding, or planning still flub hard reasoning. Teaching them to learn from genuinely relevant examples — not just look-alikes — is a quiet path to assistants that actually think problems through instead of bluffing.