Can AI predict what scientists will discover next?

Researchers built CUSP, a benchmark of 4,760 real scientific events across disciplines, to test whether AI can predict future discoveries. Current frontier models can sometimes identify plausible research directions but systematically fail at predicting whether advances will actually happen and when. Interestingly, performance doesn't improve much with additional training data—suggesting the problem isn't missing knowledge but rather a fundamental limitation in how models reason about scientific possibility and timing.