New AI Tool Could Help Scientists Make ‘Rare’ Cross-Disciplinary Leaps More Common
Using a system called BootLoops, scientists created 36 research papers in three months. When used with experts, the tool could help spread ideas from one field of science to another.
Image Credit
Anthropic
Some important scientific discoveries happen when researchers take a method developed in one field and use it in another. But research that connects different fields can be difficult and take a lot of time. Now, a new AI research system has automated parts of this work and produced 36 papers in three months.
Most researchers focus on a small area of science. Few have time to search through unrelated fields for tools that might help them. This means they may struggle with a problem that scientists in another field have already solved.
Harvard physicist Matthew Schwartz found a possible solution while working as a visiting researcher at the AI company Anthropic. He had asked the company's AI, Claude, to build a set of tools for mathematical physics. He then discovered that these tools could also help solve number-based problems in many other fields.
The system is called BootLoops. Its tools can work with different AI models, rather than only one. They help AI search scientific papers, write computer code, and solve math problems. With 19 coauthors, Schwartz used BootLoops to create 36 research papers on subjects ranging from ecology and linguistics to health insurance and sunspots.
"Once Claude had BootLoops, it kept noticing the same pattern: many fields have problems that a technique from mathematics, physics, or computer science would solve outright if anyone knew it existed," Schwartz wrote in a blog post by Anthropic about the system.
Schwartz first began using Claude as a general research assistant in December last year. Although it helped him with some useful work, he says he had to guide it constantly. So he decided to focus instead on problems that matched the AI's strengths. He calls these "Claude-shaped problems."
He began with the equations physicists use to understand particle collisions at the Large Hadron Collider. The best methods were spread across software written in several programming languages, or were available only in written papers. Schwartz asked Claude to bring these methods together into one system. Using the tools it had built itself, the AI reproduced the results of one of his papers in about 20 minutes.
Schwartz then moved on to a more difficult group of equations. Within a few weeks, the AI had solved 30 of them, including 15 that had never been worked out before. Schwartz soon realized that many of the methods the AI had developed could also be used in nearby areas of physics, such as cosmology and string theory.
But Claude made an even bigger leap on its own. It pointed out that similar equations also appear in population genetics and evolutionary biology. Schwartz found it harder to judge whether Claude was exaggerating its results in fields he did not know well. So he brought in experts from each field to decide which findings were worth exploring. Of about 400 possible problems, 36 became research papers.
One project studied the star ratings used to judge private health plans under Medicare. These ratings help determine the bonuses insurance companies receive. The rules say the dividing lines between rating levels should be set so that each plan's score is as close as possible to the average score for its group. But the method Medicare actually uses is a shortcut and does not guarantee this result.
BootLoops used a statistical method from 1958 to calculate the exact answer. The results showed that more precise dividing lines would move 34 plans across the boundary that determines whether they receive a bonus. Sixteen would move up and 18 would move down. This could shift an estimated $0.9 billion to $1.4 billion in bonus payments between 2025 and 2027.
Another project tackled a mystery that has puzzled scientists for 150 years: what makes sunspots disappear? Because the Sun's rotation hides sunspots from view for much of their lives, astronomers have not agreed on how long they last or why they vanish.
The team borrowed a tool that doctors use to measure how the risk of death changes with age. They applied it to every sunspot recorded since 1874. The results suggested that sunspots age. When two sunspots are the same size, the risk that one will disappear rises by about 20 percent each day as it gets older. None of the published theories could explain this finding.
Claude built new tools for every problem, so the BootLoops toolkit continued to grow. According to the project's website, it now covers 22 fields. Other problems it has tackled include how Earth's atmosphere first gained oxygen, how quickly tree species appear and disappear in a Panamanian rainforest, and how natural selection removes harmful mutations in humans.
Georgia Channing, who leads AI for science at Hugging Face, told Science that the kinds of discoveries BootLoops makes by connecting different fields are very rare. "A physics researcher wouldn't spend six months reading linguistics research in order to figure out how to repurpose linguistics methods," Channing said.
However, the system still has weaknesses. Schwartz says many of Claude's early findings were technically correct but did not offer anything scientifically important until experts helped guide the work. He also warns that the model "loves to declare victory" too soon.
Still, the results suggest that AI could be a powerful way to spread ideas between different areas of research. Human experts are still needed to decide which ideas are truly worth pursuing.
