AI training data bottleneck solved by having AI agents write their own training problems

Instead of humans writing training problems by hand, researchers built a system where AI agents automatically generate new problem families and test them for quality. This means AI models can learn from vastly more diverse and harder problems than hand-written datasets allow, producing measurable improvements in logic and math reasoning tasks.

For years, training AI systems on logic and reasoning has been limited by how many problems humans can write. This system removes that ceiling by letting AI agents generate problem families autonomously, then validate them with other agents playing adversary. The practical effect is immediate: with just two rounds of evolution, they went from 400 problem families to 953, and the resulting training data produced measurable gains in reasoning accuracy. What's happening under the hood is that the hard part of AI training—collecting and curating good examples—just became automatable.