The SignalToday's digest →

Problems in 22 scientific fields had solutions hiding in plain sight. An AI has found them

Science · original · full text captured 17 November 2026 · WhatsApp · Email

Problems in 22 scientific fields had solutions hiding in plain sight. An AI has found them | Science | AAAS

Opens in a new window

Opens an external website

Opens an external website in a new window

<!---->Close this consent banner<!---->

We and our 3 IAB TCF partners store and access information on your device for the following purposes: store and/or access information on a device, advertising and content measurement, audience research, and services development, personalised advertising, and personalised content.

Personal data may be processed to do the following: use precise geolocation data and identify devices based on information actively requested.

Our third party IAB TCF partners may store and access information on your device such as IP address and device characteristics. Our IAB TCF Partners may process this personal data on the basis of legitimate interest, or with your consent. You may change or withdraw your preferences for this website at any time by clicking on the cookie icon or link; however, as a consequence, you may not see relevant ads or personalized content.

You may change your settings at any time or accept the default settings.

You may close this banner to continue with only essential cookies.

<!----><!---->

science translational medicine

Quick Search in Journals

Enter Search Term

Quick Search in Journals

Enter Search Term

Quick Search in Journals

Enter Search Term

Quick Search in Journals

Enter Search Term

Quick Search in Journals

Enter Search Term

Quick Search in Journals

Enter Search Term

Searching:

Science Translational Medicine

Advanced Search

science.org

science translational medicine

Main content starts here

News

Technology

Problems in 22 scientific fields had solutions hiding in plain sight. An AI has found them

Teams led by physicist used “BootLoops” software to produce 36 papers in weeks on everything from linguistics to health policy

9 Oct 2026

11:35 AM ET

By

Celina Zhao

An artificial intelligence–powered analysis of the Voynich manuscript, a 15th century book full of undecipherable script, concludes it does not hold a hidden message.

Beinecke Rare Book & Manuscript Library/Yale University via Wikimedia Commons

Share:

A version of this story appeared in Science, Vol 394, Issue 6822.

Download PDF

Go ten seconds backward

Skip forwards

Go ten seconds forward

Progress

1.0x

Mute

Unmute

Volume is at 50%

00:00

7:38

1.0x

Audio is AI-generated.

Borrowing an idea from one research discipline to solve a thorny problem in another is an age-old scientific trick. Charles Darwin’s theory of natural selection, for example, famously inspired algorithms in computer science that “evolve” solutions to complex problems.

But such clever leaps are fleetingly rare. Usually, “A physics researcher wouldn’t spend 6 months reading linguistics research in order to figure out how to repurpose linguistics methods,” says Georgia Channing, artificial intelligence for science lead at Hugging Face, an online platform for sharing AI tools.

For AI, however, hopping between fields can be a breeze. Last week, researchers highlighted that ability with the release of

36 preprints produced with a single suite of AI tools called BootLoops

. The so-called harness can be used with any large language model (LLM), such as Anthropic’s Claude or OpenAI’s ChatGPT. And in just 3 months, researchers used Claude outfitted with BootLoops to cross-pollinate at a massive scale, spotting—and precisely solving—problems in seemingly unrelated fields that share the same mathematical structures. The resulting papers range widely, from figuring out the age of sunspots, to analyzing government health policies, to attempting to decipher a mysterious medieval manuscript.

SCIENCE

ADVISER NEWSLETTER

The latest news, commentary, and research, free to your inbox daily

In essence, BootLoops “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,” Matthew Schwartz, a physicist at Harvard University and visiting researcher at Anthropic who developed the tools, wrote in a 1 October

blog post

unveiling the preprints.

The project also helps bring AI to scientific disciplines that have been slower to adopt it, Schwartz and other researchers say. “It’s phenomenal that the technology is being saturated through more fields,” Channing says.

Researchers caution that BootLoops still needs extensive human involvement and judgment to produce useful results. They also add that it suffers from limitations common in AI-powered science. For one, it favors older, highly cited findings over new problems,

potentially shrinking the breadth of scientific exploration

.

Schwartz began to explore AI’s autonomous research capabilities in December 2025 when he took on Claude Opus 4.5 as his virtual “graduate student.” With heavy hand holding, it managed to spin up a reasonable physics research paper in just 2 weeks. Intrigued, Schwartz started to give the LLM bigger challenges, eventually as ambitious as reinventing quantum mechanics. But the process was painfully frustrating. “It just wasn’t ready,” he says.

Then, he thought: Instead of forcing AI to perform like a human researcher, why not play to its strengths?

One obvious strength was coding. Schwartz tasked Claude Fable 5 with writing code to solve a certain type of integral used in theoretical analyses of high energy particle collisions. It combed physics and math papers for other examples of the integrals, then built software to solve 30 of them, including 15 that had never before been solved. The resulting package became BootLoops.

He then unleashed it on the broader scientific literature, where it soon identified many similar unsolved problems in other disciplines. From there, Schwartz says, “The project exploded,” and he began to recruit experts from other fields to collaborate.

One of the resulting papers, co-authored by evolutionary biologist Scott Edwards of Harvard, delves into phylogenetics, in which researchers build evolutionary trees from DNA data. Here, BootLoops was able to calculate that two possible trees depicting the evolutionary relationships of humans, chimpanzees, gorillas, and orangutans were exactly tied in probability—something past estimation strategies could not establish. “This is my first experience with Claude-based science, and I can definitely see the potential,” Edwards says.

Similarly, BootLoops was also able to pin down an optimal way to rank health insurance plans offered under the U.S. government’s Medicare program. The current classification system, based on metrics such as patient experience, uses an approximate clustering algorithm to set the cutoff between five possible ratings. Those cutoffs then determine the bonus the government pays insurance companies. BootLoops, however, was able to calculate precise optimal bins. The accompanying preprint suggests implementing those suggestions would shift roughly $1 billion of taxpayer-funded bonuses to different insurance companies from 2025 to ’27.

Soon, collaborations expanded beyond math problems to other work particularly suited to LLMs. In yet another project, Schwartz and a trio of linguists used AI agents to scour the academic literature and amass the largest ever standardized catalog of languages’ stress and tone patterns, using BootLoops to check the agents hadn’t skipped any additional stress systems a language could presumably use. The resulting database, called

ACCSTACK

, is more than eight times larger than

the largest previous data set of its kind

, documenting the sound systems of more than 6000 languages. ACCSTACK is “of a scale that’s really far, far beyond what we’ve been able to do as just humans doing it,” says preprint co-author Joe Pater of the University of Massachusetts Amherst.

But as BootLoops wandered farther afield, Schwartz found that although the LLM could tell him what it could calculate, it couldn’t identify what scientists would really care about. Put simply, it lacked research taste.

For example, Claude saw an opportunity to test ecology’s neutral biodiversity theory—which holds that many genetic variations between individuals and populations are due to randomness. The AI soon generated a counterexample: an analysis, based on data from a research island in Panama, that showed tree diversity was changing faster than expected by chance alone.

Schwartz then asked James O’Dwyer, an ecologist at the University of Illinois Urbana-Champaign, to examine the case study. “I could see there were some solid, interesting results, but they might not be likely to massively move the field forward,” O’Dwyer recalls. Still, he was impressed at the speed of BootLoops. “I don’t think it’s going to replace ecologists of any kind just yet,” he says. “But I do know that everything moved much faster with this collaboration.”

Not all researchers embraced the tools. One preprint analyzes the Voynich manuscript, a 15th century book full of undecipherable script and illustrations that has perplexed scholars for centuries. The paper claimed the manuscript may not encode a hidden linguistic message after all. But in an online forum dedicated to the manuscript, some contributors complained that the paper was so dense as to be AI “slop.”

Schwartz acknowledges the need for caution: On colleagues’ advice, he withheld some results that were insufficiently validated or risked sending the wrong message about AI. Nevertheless, he believes researchers in many areas will benefit from working with discipline-spanning AI tools—especially when AI can help close the distance between the various “jagged frontiers” of science. “Once everyone catches up,” he says, “this human-AI collaborative era will be very exciting.”

With reporting by

Michael Greshko

,

Rodrigo Pérez Ortega

, and

Erik Stokstad

.

Science

’s AI in Science reporting initiative is supported by Ray Rothrock & family.

doi: 10.1126/science.zwu3e54

Relevant tags:

Technology

Artificial intelligence

Science’s AI in Science Reporting Initiative

Science

’s AI In Science Reporting Initiative, covering how artificial intelligence is affecting science and society at large.

View more

10 Nov 2026

Can researchers bring order to the confusing world of AI ‘benchmarks’?

By

Ananya

9 Nov 2026

To deter bioterrorism, U.S. imposes new rules on ordering synthetic DNA

By

Phie Jacobs

9 Nov 2026

Researchers sue to block NSF from demolishing social science directorate

Got a tip for Science's news department?

Connect

Get

Science

’s award-winning newsletter with the latest news, commentary, and research, free to your inbox daily.

9 Oct 2026

By

Erik Stokstad

U.S. crop genome sequencing gets $180 million boost

9 Oct 2026

By

Adrian Cho

Department of Energy expands ‘Genesis Mission’ to boost AI, awarding larger grants

9 Oct 2026

By

Taylor Mitchell Brown

Major employer of archaeologists shuts down, leaving workers in the lurch

8 Oct 2026

By

Kai Kupferschmidt

Will the world end next month?

Science

said so, in 1960

9 Oct 2026

By

Julia Vaz

Trump administration proposes $100,000 fee for work authorization for international students

6 Oct 2026

By

Taylor Mitchell Brown

Exclusive: Archaeology megafirm abruptly shuts down

4 Sep 2026

By

John Travis

Gene-edited pig kidney transplant sets record—and offers bridge to human organs

18 Aug 2026

By

Phie Jacobs

FDA pick is Heidi Overton

October 2026

Across fields, AI finds solutions hiding in plain sight

Special Issue News

January 2015

Hiding in plain sight

11 Jan 2018

By

Matthew Hutson

Artificial intelligence can ‘evolve' to solve problems

Research Article

January 2018

Superhuman AI for heads-up no-limit poker: Libratus beats top professionals

18 Nov 2026

We didn’t evolve alone: How microbes shaped human adaptation

×

$50

$75

$100

Other

$

Please enter a valid amount

Single

Monthly

Email:

Please enter a valid email

Country:

Bolivia, Plurinational State of

Bonaire, Sint Eustatius and Saba

British Indian Ocean Territory

Congo, The Democratic Republic of the

Cook Islands

Costa Rica

Cote D'Ivoire

Falkland Islands (Malvinas)

French Southern Territories

Heard Island and Mcdonald Islands

Holy See (Vatican City State)

Iran, Islamic Republic of

Korea, Democratic People's Republic of

Korea, Republic of

Kuwait

Kyrgyzstan

Lao People's Democratic Republic

Macedonia, The Former Yugoslav Republic of

Moldova, Republic of

Saint Helena, Ascension and Tristan da Cunha

Saint Kitts and Nevis

Saint Lucia

Saint Martin (French part)

Saint Pierre and Miquelon

Saint Vincent and the Grenadines

Samoa

San Marino

Sao Tome and Principe

Sint Maarten (Dutch part)

South Georgia and the South Sandwich Islands

Svalbard and Jan Mayen

Tanzania, United Republic of

Turks and Caicos Islands

Venezuela, Bolivarian Republic of

Vietnam

Virgin Islands, British

also wish to receive emails from AAAS/Science and Science advertisers,

including information on products, services and special offers which may

include but are not limited to news, careers information & upcoming

events.

Support nonprofit science journalism

Amid the unprecedented upheaval in U.S. science, hard-hitting, trustworthy reporting has never been more important. As part of the AAAS mission,

Science

has built a global award-winning network of reporters and editors that independently cover the most important developments in research and policy.

Your tax-deductible contribution plays a critical role in sustaining this effort.

If, at any time, you would like to stop or make changes to your monthly gift, please contact the Philanthropy Office at

philanthropy@aaas.org

or 202-326-6636.

Charitable Disclosures

Loading...

×

Support nonprofit science journalism

Help News from

Science

publish trustworthy, high-impact stories about research and the people who shape it.

Please make a tax‑deductible gift today.

Thanks for sharing!