New AI Test Unearths the Overlooked Papers Behind Big Ideas

When scientists search the literature, they usually find the same famous, highly-cited papers. But the truly influential work that sparks a breakthrough is often obscure. A new study introduces MUSES, a massive test kit with over a million examples to check whether AI search tools can find these "intellectual roots" of a discovery, not just the popular stuff. Think of it like finding the exact recipe that inspired a chef's signature dish, rather than the most-liked cooking video online.

The researchers didn't just guess which papers mattered. They asked the authors themselves. Over 750 scientists pointed to the specific papers that genuinely inspired their new work. That makes this benchmark unusually trustworthy. The results aren't flattering for today's AI systems. Even the best method finds the correct root paper in its top 100 guesses only about half the time when the paper is well-known, and its performance drops by more than half when the root is less familiar. In a tougher test, half of the true root papers weren't found even after 1,000 guesses.

Why does this matter? Because discovery depends on finding the right foundations. If search tools keep favoring hype, they can miss the hidden gems that could unite ideas and spark the next breakthrough. The team also created CiteRoots, a two-layer database and open-source tool that tracks both how citations are used in text and how authors themselves acknowledge their inspirations. This could help researchers, funding agencies, and even patent offices understand where ideas truly come from.

The catch is that the task is genuinely hard. There are unlimited ways a paper can inspire another, and even humans don't always agree. But building fairer and smarter ways to surface "the quiet papers that started the fire" could speed up science and give credit where it's due. For the rest of us, it means faster discoveries in medicine, energy, and technology—and a fairer system for the minds behind them.