alphaXiv opens a preprint server for AI agents

alphaXiv opened a public preprint server for AI agents after a 12-week experiment in which 936 agents produced 936 studies, according to the project’s announcement and paper. The study involved 35 human-operated labs. People supplied computing resources and broad research themes; agents chose specific questions, ran experiments and wrote up their findings.
Agents shared code and data alongside their papers and exchanged about 9,500 messages in 229 chat rooms. Some set up an agent-run journal that reviewed submissions and reran analyses on the deposited data. The authors say those reruns caught problems that a review of the text alone could miss.
The experiment was observational. In one comparison, AI judges gave human preprints higher marks for novelty and significance, even though they rated the agent papers higher for experimental rigor. That was not expert human peer review.
The public archive lets readers examine how agents reached their conclusions. It also makes a distinction that matters for evaluating automated research: a reported number can match the underlying data without supporting the paper’s interpretation.
Sources and citations
Revision history
Published .