BootLoops AI System Produces 36 Research Papers Across 22 Fields

BootLoops offers a striking example of how AI might help researchers find useful connections between fields. In a three-month effort, the system supported work that led to 36 manuscripts spanning subjects from ecology and linguistics to health insurance and sunspots. The result suggests AI can help surface methods that specialists might otherwise miss, while also showing why human expertise remains central to judging whether a promising connection matters.

The project was led by Harvard physicist Matthew Schwartz, working with 19 coauthors. BootLoops is a collection of tools designed to search scientific literature, write code and solve mathematical problems. It is model agnostic, meaning its tools are not limited to one AI model.

Schwartz began by using Claude, an AI model made by Anthropic, as a general research assistant. He found that approach required frequent guidance. Instead, he focused on problems that matched the model’s strengths—a set of tasks he called “Claude-shaped problems.” He had Claude build tools to combine mathematical methods that were scattered across software written in different programming languages or described only in papers.

From particle physics to other fields

The first work centered on equations used to interpret particle collisions at the Large Hadron Collider. Using the tools it had assembled, Claude reproduced results from one of Schwartz’s papers in about 20 minutes. Schwartz then applied the approach to a more difficult group of equations. Within a few weeks, the model had solved 30, including 15 that had not previously been calculated.

Some of the methods proved useful in nearby areas of physics, including cosmology and string theory. A larger jump came when Claude suggested that similar equations appeared in population genetics and evolutionary biology. That was a potentially valuable lead, but Schwartz had less expertise in those fields and could not easily tell when the model was overstating a result.

He brought in specialists to assess candidate problems and decide which findings merited further work. Of roughly 400 candidates, 36 developed into manuscripts. That selection process is a key part of the story: generating possibilities was only one stage. Researchers with subject knowledge helped distinguish technically plausible answers from results that could contribute something meaningful.

Schwartz described the recurring pattern in an Anthropic blog post: methods from mathematics, physics or computer science can sometimes solve problems in other disciplines, but researchers may not know those methods exist. AI can search across a wider range of material and propose a connection. It cannot, by that alone, establish that the connection is scientifically important.

Two examples of cross-field methods

One project examined the star ratings used to assess private health plans under Medicare. Those ratings affect bonus payments to insurers. The rating boundaries are intended to place each plan’s score as close as possible to the average for its group, but the algorithm used in practice is a shortcut and does not guarantee that result.

BootLoops applied a statistical method dating to 1958 to calculate the boundaries more precisely. The analysis found that 34 plan ratings would cross the bonus threshold under the exact calculation: 16 would move up and 18 down. The researchers estimated that this could shift between $0.9 billion and $1.4 billion in bonus payments over 2025 to 2027. The finding describes what the alternative calculation would imply; it does not establish that the current ratings will change.

Another project addressed a long-running question about sunspots: how long they last and what causes them to disappear. Their movement across the Sun means they are out of view for part of their lifetimes, making it harder to track their full development. The team borrowed a method doctors use to measure how the risk of death changes with age, then applied it to every sunspot logged since 1874.

The analysis indicated that sunspots age. For spots of a given size, the chance of disappearing rose by about 20 percent per day, a pattern the team said existing published explanations did not account for. The method offers a way to examine the data; further scientific work would be needed to explain the pattern.

What the results do and don’t show

BootLoops has grown as Claude builds tools for each new problem. The project’s website says the toolkit now spans 22 fields. Other topics it has addressed include how Earth’s atmosphere first accumulated oxygen, how tree species turn over 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 these kinds of cross-disciplinary connections are rare. A physicist may not spend months reading linguistics research to find a method that could be adapted, she said. A system that searches across fields could make such leads easier to spot.

Still, the paper count alone doesn’t show how much each finding will change its field, and the source account does not establish that the manuscripts have been validated through later research. Schwartz has also warned that many early results were technically correct but scientifically unremarkable until experts helped refine them. Claude, he said, can be quick to declare success.

My read is that BootLoops is most useful as a research assistant for finding and testing possibilities, with specialists guiding the work and assessing the claims. The next measure to watch is whether the connections lead to results that other researchers can reproduce and build on.