Science in plain language, with the paper one click away Science in plain language
In a preregistered experiment, an AI that supplied starting ideas narrowed the range of what a whole group of writers produced, while an AI that only polished writers' own ideas left that range intact.
Researchers at Deezer tested whether the digital fingerprints of AI-generated audio survive being mixed with human recordings, and report early success at spotting which individual part came from a machine.
Engineers at Marvell report that swapping electrical switches for a passive fiber shuffle gave 16 host machines a shared 32-terabyte memory pool, more than halving access latency versus conventional CXL pooling in their emulation.
In a new arXiv position paper, Chen Shani maps out how concept structure gets into large language models and argues the field has been recovering it after the fact when it could be designing it in from the start.
A Japanese team tested a gel made from Mongolian Lycium ruthenicum on lab-grown skin cells and two mouse models of atopic dermatitis, and found it calmed inflammation, tightened the skin barrier, and quieted itch-sensing nerves.
In a small cross-sectional comparison of 62 traffic officers and 120 hospital-based controls, the officers blew out about half a litre less air and reported more respiratory symptoms, though the study cannot prove the traffic caused it.
Two physicists show that if you know how a turbulent flow carries particles, you can run that transport backward from a whiff of chemical to a map of places the source might be, and the resulting search beats classic strategies in simulated wind.
In a University of Arizona study, people trained to reproduce short time intervals and line lengths improved along similar curves, yet line drawing became accurate sooner and stayed that way, and feedback helped both skills carry over to lengths and durations never practiced.
Mice given oral glycyrrhizin three times a week for a year gained less weight, avoided fatty liver entirely, and held on to more of the anti-aging protein Klotho than untreated littermates, though the study was small.
At a two-day national workshop in Kolkata, policymakers, industrial physicians and meteorologists agreed that heat, dust and chemical exposures are reshaping worker health faster than India's monitoring systems can track them.
Two election security researchers catalogue the ordinary, documented ways optical-scan voting systems go wrong, and point to a category of failure that no recount of the paper ballots can catch.
A survey of 1,223 psychotherapy randomized trials published between 2019 and 2023 found that internet-delivered treatments became the single largest category, and that seven in ten of them were still cognitive behavioural therapy.
Astronomers hunting the faint radio signature of the first stars found that one step in their calibration software was so numerically fragile that it gave different answers on different machines, and they show a way to fix it.
Four theorists have built an algorithm that learns a quantum state's eigenvalues using slightly fewer copies of the state than the long-standing Keyl-Werner method requires, settling a question open since 2001 and refuting a 2016 conjecture.
A study of 10 introductory STEM courses finds that students in sections with learning assistants did better across the board, with the largest pass-rate gains among historically marginalized students.
An Italian team measured how ordinary data traffic in one wavelength band scatters light into the band reserved for quantum signals, and found the resulting noise profile fits a formula that does not care which transceiver produced it.
Xiao-Tian Wu constructs a family of solutions to an optimal transport equation and singles out one whose smoothness stops precisely at the exponent 1/(2n-1), showing that a certain kind of interior estimate cannot exist.
A new preprint proves mathematically that once an orbiting body is deformable enough to be stretched by tides, the elegant clockwork that governs black hole orbits breaks down, and numerical tests show chaos taking its place.
A pair of physicists report 3D structures that machine-learning models of matter would see as identical even when the models look at clusters of up to seven neighboring atoms, and they credit a large language model with pointing them to the decades-old mathematics that made the construction possible.
Two mathematicians prove that a standard numerical method for equations mixing heat flow with random noise hits its best possible accuracy at any time after the start, provided you accept a manageable blow-up right at time zero.