Edison's Kosmos AI scientist: independent scientists rated 79.4% of its report statements accurate
Kosmos runs for up to 12 hours on a dataset and a goal. It loops through data analysis, literature search and hypotheses, then writes a cited report. Its makers say it often goes down rabbit holes.
- Organization
- Edison Scientific
- Industry
- AI Research
- When
- Nov 2025
- Evidence
- Research paper
- Agents and tools
- Kosmos
Reported outcome
Independent scientists found 79.4% of statements in Kosmos reports accurate. Collaborators said one 20-cycle run did the equivalent of 6 months of their own research time on average.
- 79.4%Statements rated accurate
- 6 months of their research time, on averageCollaborators' estimate for one 20-cycle run
- 42,000 lines of code, 1,500 papers readPer run, on average
- 7 (3 reproduce unpublished or preprinted findings, 4 are new)Discoveries highlighted
- $200 per runLaunch price
Caveats
Described in a paper with a stated method. Peer review status varies.
- The paper is an arXiv preprint written by the team that built and sells Kosmos. The page does not show peer review.
- About one statement in five was not rated accurate.
- The 6 months figure is an estimate from collaborators and beta users. It is not a measurement.
- Edison says Kosmos often goes down rabbit holes or chases statistically significant yet scientifically irrelevant findings, and more so on longer runs.
- Edison says it often runs Kosmos several times on the same goal.
What the agent does
A user gives Kosmos an open-ended goal and a dataset. Kosmos runs for up to 12 hours. It repeats cycles of data analysis, literature search and hypothesis generation, and then writes its findings as scientific reports.
A structured world model shares information between a data analysis agent and a literature search agent. The paper says this keeps Kosmos on its goal over 200 agent rollouts. Every statement in a report cites code or primary literature, so a reader can trace it.
Edison Scientific is a commercial spinout of FutureHouse, which built the earlier Robin system.
What happened
Independent scientists rated 79.4% of the report statements accurate. The paper highlights seven discoveries in metabolomics, materials science, neuroscience and statistical genetics. Three reproduce findings from preprinted or unpublished manuscripts that Kosmos did not access at runtime. Four are described as new contributions.
Edison launched Kosmos at $200 per run. It describes the tool as closer to a reagent kit than a chatbot: something to run on high-value targets.
What to copy
- Make every claim cite the code or paper behind it.
- Keep shared state outside the model's context for long runs.
- Publish your own accuracy rate, including the part that fails.
Starter prompt
Written by us, not by Edison Scientific. Check every claim against its cited code or paper before you use it.
Goal: {research_goal}
Dataset: {dataset_description}
Work in cycles. In each cycle: run one analysis, search the literature for it, and write one hypothesis.
After each cycle, update a short list of what you know and what is still open.
In the final report, attach the code or the paper behind every statement.
Mark each statement as data analysis, literature, or interpretation.
Sources
- Kosmos: An AI Scientist for Autonomous Discovery — arXiv, Nov 4, 2025
- Announcing Kosmos — Edison Scientific, Nov 5, 2025
Added Nov 16, 2026 · Sources last checked Nov 16, 2026.
Get new use cases weekly
One short email. Sources included.
The newsletter is paused. Past issues stay in the archive.
Related use cases
Stanford's Virtual Lab of AI agents designed 92 nanobodies, and lab tests found two strong candidates
The agents designed 92 new nanobodies. In lab tests, two showed improved binding to the JN.1 or KP.3 variants while keeping strong binding to the ancestral spike protein.
FutureHouse's Robin agents proposed ripasudil as a dry AMD candidate; humans ran the lab tests
Robin's analysis agent reported ripasudil raised retinal cell phagocytosis 7.5-fold over controls in a lab assay; a human analysis of the same data found 1.75-fold.
Sakana's AI Scientist-v2 wrote a paper that passed peer review at an ICLR workshop
One of three fully AI-generated submissions cleared the acceptance bar at the ICBINB workshop (scores 6, 7, 6).