Run a 12-hour AI scientist on your research data
Edison Scientific's Kosmos reads about 1,500 papers per run; independent scientists judged 79.4% of its report statements accurate, and it highlights seven discoveries.
Done with Custom agent
What they did
Edison Scientific's team built Kosmos, an AI scientist. It takes an open-ended goal and a dataset and runs for up to 12 hours. One agent analyzes the data and another searches published research. They share findings in a common notebook, then repeat cycles of analysis, reading and forming new ideas. Kosmos writes reports that link every statement to its code or to a source paper. Scientists supplied data and goals and checked the results.
How it went
The paper reports about 200 agent runs, 42,000 lines of code and 1,500 papers read per run on average. Independent scientists found 79.4% of report statements accurate. Collaborators said a 20-cycle run equaled about 6 months of their own research time. It highlights seven discoveries: three match unpublished findings and four are new.
Worth knowing
About one in five report statements was not judged accurate. The links to code and papers matter because they let you trace and check each claim.
Try it yourself
I have a dataset at [file or folder] and this research objective: [objective]. Work in cycles. In each cycle, run an analysis on the data, search the published literature for related findings, and update a running notes file that lists hypotheses, evidence for and against, and open questions. Save all analysis code. After [number] cycles, write a short report where every claim cites either the specific code output or a paper with a link. Flag which findings look new and which repeat known results. Show me your plan for the first cycle before starting.
Swap the highlighted [parts] for your own details.
Source: arXiv · Edison Scientific team · Nov 2025
Written with AI assistance from the original post and checked against our editorial guidelines. Results are as reported by the source, not independently verified.
