BootLoops helped prepare 36 scientific preprints — Science News

The BootLoops artificial intelligence toolkit, combined with the Claude language model, helped researchers prepare 36 preprints in various fields of science within three months — from physics and linguistics to ecology and health policy. As Science News reports, the system searched for unsolved problems in different disciplines that have similar mathematical structures and applied methods from mathematics, physics or computer science to them.

From integrals to evolutionary trees

BootLoops was developed by Harvard University physicist and Anthropic visiting researcher Matthew Schwartz. He initially tasked the Claude model with writing code to solve a certain type of integral used in the theoretical analysis of high-energy particle collisions. The system analyzed scientific papers and created software to solve 30 such integrals, 15 of which had not previously been solved.

In one of the preprints, prepared together with evolutionary biologist Scott Edwards, the tool was applied to phylogenetics. It calculated that two possible patterns of evolutionary relationships among humans, chimpanzees, gorillas and orangutans have exactly the same probability. Previous assessment methods did not make it possible to establish this precisely.

More current news is available on the UA.News Telegram channel Telegram.

Medicare and linguistics

Another preprint proposes a more precise way to rank health insurance plans in the US Medicare program. The authors state that applying the optimal thresholds calculated by the system for five ratings could potentially have saved taxpayers approximately $1 billion in 2025–2027.

In a linguistics project, AI agents compiled a standardized catalog of stress and tone systems in languages. The ACCSTACK database covers more than 6,000 languages and is more than eight times larger than the previous largest dataset of this type.

At the same time, scientists warn that BootLoops requires substantial human oversight and expert evaluation. According to Schwartz, a language model can determine what it is able to calculate, but does not always understand which questions are truly important for a particular scientific field. The tool also tends to rely on older and frequently cited research, which may narrow the range of topics it searches for.