Biology Atlas

Map / Concepts

concept · intermediate · partly written

Kosmos and its 'structured world model'

Kosmos is Edison Scientific's AI scientist; it says it keeps a structured world model between cycles and does not publish what is in it.

Also called: Kosmos AI scientist

What the paper says

Each cycle runs up to ten literature or analysis tasks, then updates the world model with summaries of the task outputs and queries it to plan the next tasks. The paper's methods are per-figure; no section describes the world model's fields or update rule. 1

Reported, not yet replicated

Scale and accuracy, as the builders report them more

About 1,500 papers and 42,000 lines of code per run, around 200 agent rollouts and up to 12 hours. Evaluators judged 79.4% of 102 report statements accurate: 85.5% for data analysis, 82.1% for literature and 57.9% for interpretation. There is no ablation with the world model switched off, and the paper does not say how its evaluators were recruited. These are the builders' own figures. 1

Reported, not yet replicated

What an outside audit found more

A two-author audit (abstract read, preprint, three claims) found that the DDR-p53 claim was not distinguishable from random gene sets, while the CDO1 claim was a real outlier (r 0.70, empirical p 0.0039). Gene names as checked against HGNC: TP53 is tumor protein p53; CDO1 is cysteine dioxygenase type 1; SOD2 is superoxide dismutase 2; SSR1 is signal sequence receptor subunit 1. 23

Reported, not yet replicated

Company description deep

Edison's own announcement treats custom world models as a product and gives no structure. Edison later said its PaperQA3 reader powers Kosmos's literature reading. Both are company pages about the company's own product. 45

Reported, not yet replicated

Plain answer

Kosmos keeps a 'world model' updated with summaries of task outputs. Its format, fields and update rule are not published in the paper read, so what it holds is unknown. 1

Reported, not yet replicated

Why it matters

It is the clearest case where a system's central claim about representation cannot be inspected. 14

Our inference

Still open here

Not yet written: Equation, Real numbers.

Linked from

Sources

  1. Edison Scientific and collaborators (37 authors). Kosmos: an AI scientist for autonomous discovery. arXiv 2511.02824 (2025). preprint, not peer reviewed; opened 2026-11-18 all records citing this
  2. Nusrat, Nusrat. Null-model audit of three Kosmos discovery claims. arXiv 2511.13825 (2025). preprint, not peer reviewed; opened 2026-11-18 all records citing this
  3. HGNC REST service (genenames.org): gene symbol lookups. HUGO Gene Nomenclature Committee (2026). database, not peer reviewed; opened 2026-11-17 all records citing this
  4. Announcing Kosmos. Edison Scientific (2025). organisation-page, not peer reviewed; opened 2026-11-18 all records citing this
  5. Edison literature agent (PaperQA3). Edison Scientific (2026). organisation-page, not peer reviewed; opened 2026-11-18 all records citing this
Where it sits on the map
History of this record (5)

This record has been changed 5 times, on the dates below. The note kept against each change is written for the people who maintain the atlas, in their own shorthand, so it is not shown here; it is in the record itself.

Status
partly written
Last updated
2026-11-19
Record
content/concepts/kosmos-world-model.json