Kosmos: AI Scientist and Its Experimental Achievements

Edison Scientific reveals that Kosmos AI Scientist has completed multiple research achievements, equivalent to months of human labor. - The technical report states that Kosmos has made 7 genuine scientific discoveries. - Among these, 3 replicate previously undisclosed human results, and 4 are original research. - Some original findings are currently being validated in laboratories. - A single run is equivalent to approximately 6.14 months of human research workload (estimated by beta users). - Kosmos can read and integrate thousands of papers and tens of thousands of lines of analytical code. - The key technology is Structured World Models. - It can maintain logical coherence over a span of more than ten million tokens. - Example: For Alzheimer's disease, it retrieved ~1500 papers and executed 42,000 lines of analytical code, inferring key molecular pathways and generating a report with citations. - The output is traceable, allowing each conclusion to link back to the corresponding papers and code records.
Key insights
Kosmos Claims to Achieve Months of Research Results in a Single Run
Edison Scientific reveals that Kosmos can complete approximately 6.14 months of human research workload in a single run. Comparisons between official and beta user data indicate that similar tasks typically require reading about 1500 papers and executing around 200 analytical pipelines. Kosmos integrates papers and analytical code in a single long-span run to generate research conclusions.
Kosmos Has Produced Multiple Traceable Research Discoveries
The technical report shows that Kosmos has completed 7 scientific discoveries, of which 3 replicate previously undisclosed human results and 4 are original, with some currently being validated in laboratories. The output reports can link back to the corresponding papers and analytical code, ensuring traceability and reproducibility of conclusions.
Structured World Models and Long Contexts as the Technical Foundation of Kosmos
Kosmos employs Structured World Models to construct dynamic, iterative knowledge graphs, integrating hundreds of autonomous reasoning trajectories while maintaining coherence over a context span of more than ten million tokens, extracting and inferring research pathways from thousands of papers and tens of thousands of lines of code.
Kosmos Demonstrates End-to-End Automation Capabilities in Specific Research Scenarios
In the case of Alzheimer's disease, Kosmos retrieved ~1500 papers, extracted proteomics and single-cell data, automatically executed about 42,000 lines of analytical code, and generated a cited research report, showcasing its end-to-end workflow capabilities from retrieval to reasoning to reporting.