arXiv began as an overflowing physics mailbox and changed how science gets shared
Around 1990 a physicist's habit of emailing papers to colleagues kept filling their inboxes. Paul Ginsparg's fix, a central mailbox at Los Alamos in 1991, became arXiv, where Grigori Perelman posted his famous proof outline rather than bother with a journal, and which in 2026 became fully independent.
arXiv, pronounced archive, with the X standing for the Greek letter chi, is an open repository of scientific e-prints that are moderated but not peer reviewed. It covers mathematics, physics, astronomy, computer science, quantitative biology, statistics, economics and more. It owes its existence to TeX, a compact format that let papers travel easily over the internet. After Joanne Cohn's emailed preprints swamped mailboxes, Ginsparg set up a repository in August 1991 at the Los Alamos National Laboratory, first known by the domain xxx.lanl.gov.
Los Alamos showed little interest, so in 2001 Ginsparg moved to Cornell and renamed the service arxiv.org, a name he dreamed up with his wife after finding archive already taken. He won a MacArthur Fellowship in 2002 for creating it and later quipped that what he intended as a brief excursion had turned into a lifelong commitment. Cornell University Library took over running it in 2011. In 2021 Nature listed it among ten computer codes that transformed science, and arXiv separated from Cornell on July 1, 2026 to broaden its funding, appointing its first chief executive, Penelope Lewis, soon afterward.
Its influence runs deep. arXiv's success with preprints helped spark the open access movement, and many papers never reach a journal. Perelman's November 2002 outline proving Thurston's geometrization conjecture, which includes the Poincaré conjecture, remains purely an e-print. Crank papers claiming to prove Fermat's Last Theorem with school algebra do appear but were called surprisingly rare in 2002, and tend to be moved to a general mathematics category rather than deleted.
Quality control has tightened. Moderators screen submissions, and since 2004 new authors in some categories need endorsement from established ones, who judge fit rather than correctness. Some 14,000 preprints have been withdrawn, most often for crucial errors. Faced with a surge of AI-generated work, arXiv stopped accepting unvetted computer science reviews and position papers in November 2025, and in May 2026 its computer science chair warned that clear signs of unchecked language model output would bring a one-year ban.
Source: ArXiv
