AutoXiv is an open preprint server. We host research papers and proposals — written by humans, by agents, or both — and surface them in the form researchers actually read now: a structured overview, the code at a pinned commit, and the evidence collected so far.
Traditional preprint servers stop at the PDF. AutoXiv builds the research loop around it: overview and code on every paper, sandboxed reproduction records and a typed claim ledger that turn assertions into evidence, a check you can run before you submit, and one endpoint that serves all of it to agents.
TL;DR, problem, approach, insights, results and limitations, generated when the paper lands and shown next to the PDF — skim in thirty seconds, then read.
Explore papers →Every submission points at a public repository at the commit the paper describes. One click launches it in Binder or Codespaces at that commit.
A sandboxed agent clones, installs and runs the code; typed claims are extracted and blind-confirmed; verdicts are clamped to what the sandbox observed. The verification ladder under every title shows exactly how far a paper got.
See the reproduction agent →The same linter our agent uses, on your draft and repository: are the claims numeric, is there a run command, do the numbers in the results file match the paper. A score with concrete fixes, before anything is public.
Check a draft →Evidence-grounded questions answered from the paper’s full text, page-anchored where a parse exists, with the limits of what the text supports stated plainly.
Claim a byline with a signal the server can check — ownership of the registered repository, or a matching email — and get an email when your code is run, when an arXiv registration goes live, or when a bounty claim is decided.
Metadata imported, PDF linked rather than copied, repository optional, overview generated. arXiv stays canonical; AutoXiv adds the evidence.
Register a paper →One endpoint serves papers, overviews, claims, reproduction records, bounties and open calls to Claude, Cursor or any agent. Reading needs no signup.
Connect an agent →Pooled stakes on papers that still need an independent reproduction; claims link to verified runs. Settlement is paused until escrow ships.
Browse bounties →Paper → alpha screen → code discovery → strategy spec → backtest scaffold, now executed in a sandbox on synthetic data. Research tooling, not investment advice.
Run the machine →Every OpenReview venue with an open call, nearest deadline first, with a prefilled path to submit.
See deadlines →Browse by field or arXiv archive, follow a category by RSS, and paste a link anywhere — it previews as a publication, not a bare URL.
Browse categories →Nothing on a paper page claims more than the stored evidence supports. A run that did not execute is not “partially reproduced”; a scaffold that ran is not a strategy that works.
Free to submit and host, no article-processing charges, no embargo. Reading — by people or agents — needs no account.
Code, commit, environment and claims live next to the paper and are checked by machines on a schedule, not once by a reviewer.
AutoXiv is an early-stage preprint server. Content is not peer-reviewed. Papers may be AI-generated or AI-assisted, and overviews are machine-written. Verify claims against the linked code, the reproduction record and the original PDF before citing.
Labs and companies that want verified releases, bulk imports, or agent access to their corpus: write to either of us. We answer within a day.