Overview
Abstract: We introduce Kosmos, an AI system designed to conduct extended autonomous scientific research. Kosmos operates for up to 12 hours performing cycles of parallel data analysis, literature search, and hypothesis generation before synthesizing discoveries into scientific reports. A single run executes approximately 42,000 lines of code, reads approximately 1,500 papers, and completes up to 200 agent rollouts while maintaining coherence across this extended task. When we ask Kosmos to do research alongside our own, it provides traceable reasoning with citations to code or primary literature, enabling us to immediately assess the validity of thousands of scientific statements. Independent evaluation shows 79.4% of these statements are accurate. Our collaborators report that single Kosmos runs perform the equivalent of approximately 6 months of their own research. We document seven discoveries made by Kosmos in metabolomics, materials science, neuroscience, and statistical genetics, and we find that it independently reproduces some findings which later appeared in preprints.
- Authors: Ludovico Mitchener, Angela Yiu, Benjamin Chang, Andrew D. White, et al. (37 authors total)
- Year: 2025
- arXiv: cs.AI (2511.02824)
- Key Achievement: Autonomous AI system performing 6-month equivalent research in 12 hours
- Validation: 79.4% accuracy on scientific statements
- Applications: Metabolomics, materials science, neuroscience, statistical genetics