Kosmos: An AI Scientist for Autonomous Discovery

TL;DR

Kosmos is Edison Scientific’s AI scientist for open-ended data-driven discovery: given a dataset and a natural-language research objective, it runs up to 20 iterative cycles, each launching up to ten parallel data-analysis-agent and literature-search-agent rollouts, and coordinates them through a structured world model — a queryable database of entities, relationships, results, and open questions updated after every task. A single run averages 42,000 lines of code generated across ~166 data-analysis rollouts and ~1,500 full papers read across ~36 literature-search rollouts, claimed equivalent to ~6 months of human researcher time. The paper reports seven case-study discoveries spanning metabolomics, materials science, neuroscience and statistical genetics — three reproductions of unpublished/post-cutoff results, four novel contributions — with independent expert evaluators marking 79.4% of report statements as accurate.

Key claims

The system’s core advancement over prior AI scientists (Robin, Sakana, Google co-scientist, Virtual Lab) is a structured world model as long-term memory: a queryable database of entities, relationships, results, and open questions that is updated after every task, decoupling per-task context windows from cross-cycle state and enabling 8× more iterations than existing systems [§2.1].
A typical run executes ~42,000 lines of code across ~166 data-analysis-agent rollouts and reads ~1,500 full-length papers across ~36 literature-search-agent rollouts — a 9.8× increase in code generation compared to the team’s prior Robin system [§1, Fig. 1b].
Independent expert evaluators classified 79.4% of 102 statements drawn from three representative Kosmos reports as “Supported” with the underlying code/literature withheld during evaluation; accuracy splits by statement type: 85.5% data-analysis-based, 82.1% literature-review-based, 57.9% cross-domain synthesis [§2.1, Fig. 1c].
Expert-equivalent research time scales with run length: collaborators estimated a 20-cycle Kosmos run produced findings equivalent to 6.14 months (n=7, σ=2.49) of their own research, and that the count of “valuable findings” scales roughly linearly with cycles up to the tested 20-cycle ceiling [§2.1, Fig. 1e–f].
Discovery 1 (metabolomics): Kosmos independently reproduced the nucleotide-salvage mechanism in cooling-induced neuroprotection, recovering 9 of the top 15 metabolites and the precursor-to-product inversion pattern reported in the unpublished Kamal et al. dataset, with near-indistinguishable fold-change estimates [§2.2.1, Fig. 2g].
Discovery 2 (materials science): on the Liu et al. perovskite-solar-cell dataset, Kosmos identified absolute humidity during thermal annealing as the dominant fabrication factor (matching the source SHAP analysis) and surfaced a previously unreported linear short-circuit-current dependence on DMF solvent partial pressure that was later validated by the original researchers [§2.2.2, Fig. 3].
Discovery 3 (neuroscience): given connectome data from five species and a broad “find universal principles” prompt, Kosmos independently rediscovered the log-normal distribution finding and the power-law cross-metric scaling of Piazza et al., with Pearson r > 0.85 between Kosmos’s fitted log-normal means and the preprint’s published values; reruns with pre-preprint-cutoff model snapshots reproduced the conclusion, ruling out memorization [§2.2.3, Fig. 4e].
Discovery 4 (statistical genetics): using a Mendelian-randomization pipeline over a myocardial T1 GWAS + cis-pQTL data, Kosmos independently identified SOD2 as the primary causal protein for myocardial fibrosis (PP.H4 = 1.00), with 31 of 32 Bonferroni-significant proteins overlapping the human analysis and a near-perfect effect-size correlation (Pearson > 0.99) [§2.3.1, Fig. 5b,e,f].
Failure-mode honesty: the SOD2 run’s colocalization step cascade-failed due to p-values stored as 0 being misread as lead SNPs and merge-time column collisions; Kosmos proceeded to SuSiE fine-mapping anyway, and a miR-222 binding-site hypothesis it proposed turned out not to overlap predicted sites in target databases — both reported in §2.3.1 [§2.3.1].

Method

Kosmos is initialized with a research objective and a dataset. Each cycle launches up to ten parallel agent rollouts split between a general-purpose data-analysis agent (a Jupyter-notebook-executing coding agent built on Edison Scientific’s prior work) and a general-purpose literature-search agent. After all tasks in a cycle complete, the structured world model — a database of entities, relationships, experimental results, and open questions — is updated with summaries of each agent’s output. The world model is then queried to propose the next cycle’s tasks, so the planning loop is mediated entirely through the structured store rather than a stuffed context window. When Kosmos judges the objective complete, it synthesizes three to four scientific reports, with every statement and figure linked to either a literature citation or a specific Jupyter notebook generated by the data-analysis agent.

The world model is the load-bearing design choice and is what the authors credit for the 8× increase in iterations over prior AI scientist systems (Robin, Sakana AI Scientist, Virtual Lab). Information persists across tens of thousands of tokens of total per-run context because it lives in a queryable, structured form rather than in the rollout transcript.

Results

Per-run averages: ~42,000 LOC executed, ~1,500 papers read, ~200 total agent rollouts [§1, Fig. 1b].
Statement-level accuracy on 102 sampled report statements: 79.4% supported overall (85.5% data-analysis / 82.1% literature / 57.9% synthesis) [§2.1, Fig. 1c].
Time-equivalence: ~4.1 expert-months estimated by tallying task counts at 15 min/paper + 2 h/notebook; ~6.14 expert-months (n=7) when collaborators rated against their own research time [§2.1, Fig. 1d–e].
Seven case-study discoveries: three reproductions of unpublished/post-cutoff findings (metabolomics, perovskite SHAP, log-normal connectomes), two novel-support runs (SOD2/myocardial fibrosis, SSR1/T2D), one new analytical-method run, and one novel clinically-relevant discovery [§2.1 categorization].

Why it’s interesting

This is the third filed datapoint on closed-loop AI-for-AI-research scaffolds, but at a granularity the existing cluster doesn’t cover: where Scaling Karpathy's Autoresearch: What Happens When the Agent Gets a GPU Cluster tunes one knob, AlphaGo Moment for Model Architecture Discovery proposes neural architectures, and FARS: Fully Automated Research System produces short ML papers, Kosmos targets scientific discovery on real-world datasets (metabolomics, GWAS, connectomics, materials) with domain-expert validation as the ground truth. That makes it the first filed AI-for-AI-research system whose loss function is “did human experts agree”, not “did val-bpb go down” — and it gives the wiki a concrete number (79.4% statement-level accuracy, 57.9% on cross-domain synthesis) for what that ceiling currently looks like.

The structured world model is also a counterpoint to the parameter-space / context-space split in Tool-Use Agents: instead of fine-tuning a model on agent trajectories (MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling) or maintaining a natural-language experience library on a frozen LLM (Training-Free Group Relative Policy Optimization), Kosmos puts the cross-rollout memory in a structured database, queried as a planning substrate. That’s a third axis on the “where does the agent’s long-horizon memory live” design space the wiki was tracking — neither in weights nor in context.