Abstract
Data-driven scientific discovery requires iterative cycles of literaturesearch, hypothesis generation, and data analysis. Substantial progress has beenmade towards AI agents that can automate scientific research, but all suchagents remain limited in the number of actions they can take before losingcoherence, thus limiting the depth of their findings. Here we present Kosmos,an AI scientist that automates data-driven discovery. Given an open-endedobjective and a dataset, Kosmos runs for up to 12 hours performing cycles ofparallel data analysis, literature search, and hypothesis generation beforesynthesizing discoveries into scientific reports. Unlike prior systems, Kosmosuses a structured world model to share information between a data analysisagent and a literature search agent. The world model enables Kosmos tocoherently pursue the specified objective over 200 agent rollouts, collectivelyexecuting an average of 42,000 lines of code and reading 1,500 papers per run.Kosmos cites all statements in its reports with code or primary literature,ensuring its reasoning is traceable. Independent scientists found 79.4% ofstatements in Kosmos reports to be accurate, and collaborators reported that asingle 20-cycle Kosmos run performed the equivalent of 6 months of their ownresearch time on average. Furthermore, collaborators reported that the numberof valuable scientific findings generated scales linearly with Kosmos cycles(tested up to 20 cycles). We highlight seven discoveries made by Kosmos thatspan metabolomics, materials science, neuroscience, and statistical genetics.Three discoveries independently reproduce findings from preprinted orunpublished manuscripts that were not accessed by Kosmos at runtime, while fourmake novel contributions to the scientific literature.