Meet Kosmos: An AI scientist for data discovery

Kosmos, developed by Edison Science, is an independent discovery program that conducts long-term research missions on a single mission. Given a dataset and an open-source language intent, it performs iterative cycles of data analysis, literature search, and hypothesis generation, and compiles the results into a fully articulated scientific report. A typical run takes up to 12 hours, includes 200 agents, generates about 42,000 lines of code, and reads about 1,500 papers.

Architecture, world model, and agent roles

The basic choice of Core Design is a structured world model that acts as a long-term memory of the program. The World Model is a database of structures, relationships, test results, and open questions that is updated after every job. Unlike the context window, it is dynamic and editable, so information from the first steps is still available after tens of thousands of tokens.

Kosmos uses two main agents, a data analysis agent and a literature search agent. Each cycle, the program proposes 10 concrete activities based on the research objective and the current country model. Examples include applying differential analysis to metabolomics data, or searching for pathways that link a selection genotype to a disease phenotype. Agents write code, run it in a notebook environment, or find and read documents, and write back structured outputs and measurements on world images.

This loop repeats for many cycles. At the end of the run, a different part of the synthesis takes the state report and outputs a report in which all the statements are linked either in a jupyterbook cell or in a specific paragraph in special books. This clear view is important in scientific settings because it allows people to collaborate to test individual claims instead of treating the system like a black box.

Time accuracy and research time

This group examines reporting quality by sampling 102 statements from 3 reports representing kosmos and asking domain experts to rate each statement as it is supported or given. In total, 79.4 percent of the statements were judged correctly. Data analysis statements are most reliable about 85.5 percent of the time, literature statements are correct about 82.1 percent of the time, and synthesis statements that include evidence are correct about 57.9 percent of the time.

To measure the effort of the same person, the authors took 2 hours to find a typical trajectory for data analysis and 15 minutes for reading the paper, then calculated the trajectories and papers for each run. This burns out to about 4.1 months for a typical runner, assuming a 40-hour work week. In a separate survey, working with scientists who cooperate with 7 cosmos is equal to about 6.14 months of their work for the same purpose, and this considered effort has been directly measured by the number of cycles up to 20.

Findings

Kosmos is tested on 7 subjects studied by metabolomics, materials science, neuroscience, statistical genetics and neurodegenation. In 3 cases, independently multiplying the previous results of people without reaching the actual reasprints during the run. In 4 cases, it proposes methods that the authors describe as a variety of literature.

In a first discovery, kosmos analyzes metabolomics data from mouse hypothermia experiments. It points to the metabolism of nucleotides as a prominently altered way in hypothermic brain, with a reduction of previous bases and nucleosides and increased products of monophosphate. The program concludes that salvage nucleotide pathways dominate over de Novooting syntherthesis during hypothermia protection, which is consistent with independent human analysis that was published during the run.

In the second discovery, kosmos is analyzing natural materials from the perovskite system of the solar system. It rediscovers the result that absolute humidity during thermal induction is the main attribute of the efficiency of the device and points to the critical humidity threshold defined as a fatal filter, where the devices fail. What you find is like a mover in material science that could not be found in the cosmos at the last time due to the discovery of the models of the last obstacles and the obstacles of retrieval.

In the third discovery, the kosmos are given a reconstruction of the levels of the neuron in several species and can be distributed by the length of the neurite, degree, and the number of oscillations. It concludes that the distribution of degree and synapse is better as a log normal than scale free and rests to scale the power law between neurite length and several food counts in many datasets. These results are consistent with the rules of communication reported in Neuroscience Preprint.

The remaining four discoveries are described as novel. It includes a randomized Mendelian analysis that emphasizes the circulation of superoxide dismutase 2 as a protective fibrosic factor for myocardial fibrosis, proteomic processing that includes the events behind the aste TranscripTomic analysis that links the age-related loss of Flippase Expression and the expression of phosphatidylserine markers to enter the vulnerability of the cortex neuron.

Key acquisition

1. KOSMOS is an autonomous AI scientist that runs for 12 hours for each purpose, generating 42,000 lines of code and reading nearly 1,500 papers for each world run, integrated into a structured world model.
2. The system uses the same data analysis and literature search agents that share a centralized country model, which allows Kosmos to continue meeting for a long time with about 200 agents.
3. Expert testers found 79.4 percent of the sampling culture to be accurate, with data analysis and literature statements of more than 80 percent accuracy, while interpretation statements remain reliable.
4. The Clizcle Kosmos Run is shared as 6 months of technical research efforts, and the number of important discoveries measures almost exactly the cycle count by 20.
5. In the course of 7 metabolomics, synthetic sciences, neuroscience, mathematical genetics and neurodegeneration, both cosmos produce novel methods, while seeking human scientists for data selection and validation.

Kosmos shows what happens when a structured world model and Domain Agnostic Edison Agents are pressed by the current boundaries of LLM, re-emergence, and interpretation of synnonthetis statements that remain reliable by data and literature statements. Overall, Kosmos is a solid template for accelerated science, not a replacement for human investigators.

Look Paper and Technical details. Feel free to take a look at ours GitHub page for tutorials, code and notebooks. Also, feel free to follow us Kind of stubborn and don't forget to join ours 100K + ML Subreddit and sign up Our newsletter. Wait! Do you telegraph? Now you can join us by telegraph.

Michal Sutter is a data scientist with a Master of Science in Data Science from the University of PADOVA. With a strong foundation in statistical analysis, machine learning, and data engineering, Mikhali excels at turning complex data into actionable findings.