Accelerating scientific discovery with Co-Scientist

Summary

Co-Scientist is a multi-agent system built on Gemini for structured scientific thinking and hypothesis generation. Conditioned on a scientist's research objective and prior evidence, it produces research hypotheses intended for experimental verification. Agents continuously generate, critique, and refine hypotheses, and the process is accelerated by scaling test-time compute through an asynchronous task-execution framework and a tournament evolution process. Automated evaluations show hypothesis quality improving with additional compute. Although general-purpose, validation is focused on three biomedical applications — drug repurposing, novel target discovery, and mechanisms of antimicrobial resistance. In the acute myeloid leukemia case, drug-repurposing candidates and synergistic combination therapies proposed by the system were validated in vitro. This is the Nature (2026) version of the February 2025 arXiv preprint "Towards an AI co-scientist"; arXiv v2 (June 2026) carries the Nature title.

Contribution

Claimed: (1) a multi-agent architecture with asynchronous task execution for flexible compute scaling, (2) a tournament evolution process for self-improving hypothesis generation, "demonstrably novel" hypotheses, and real-world validation. What the abstract supports: in-vitro confirmation of AML candidates and an automated-evaluation trend that quality rises with test-time compute. The abstract contains no quantitative results, no comparison with human experts or other systems, and no operational definition of "demonstrably novel", so the size of the effect and the hit rate of proposed versus validated hypotheses cannot be judged from it.

Method

System paper with biomedical validation. Design: multi-agent generate-critique-refine loop, tournament evolution, asynchronous execution, test-time compute scaling. Evaluation: automated evaluations of hypothesis quality as compute scales (metric and judge unspecified in the abstract), plus three biomedical applications with in-vitro experiments for AML. The abstract does not name the agent roles (the catalog page, drawing on the paper body, lists Generation, Reflection, Ranking, Proximity, Evolution, and Meta-review agents under a Supervisor with Elo-based tournaments), the number of hypotheses generated or tested, or how human collaborators selected candidates for the wet lab.

Relevance to RISE

Informs literature-synthesis, hypothesis-generation, and research-design; it does not execute experiments, analyze data, or draft papers. Two catalog entries depend on it: google-co-scientist (the closed first-party system, scored 3 on internal evaluation for this Nature paper and 0 on cross-family policy because every agent runs on Gemini) and open-coscientist (an MIT-licensed reimplementation of the arXiv v1 design). Compared with ghareeb2026robin, the other Nature 2026 paper in the KB with wet-lab validation, Co-Scientist stops at ranked hypotheses whereas Robin also proposes experiments and interprets results. For the multi-agent structure/aggregation question this is the clearest case in the batch: the abstract attributes self-improvement directly to a tournament evolution process in which agents generate, critique, and refine hypotheses, and reports that hypothesis quality keeps improving as test-time compute is scaled through that structure — an aggregation rule presented as the source of epistemic gains, though the abstract does not separate the tournament's contribution from that of more sampling or the base model.

Critique / open questions

The abstract is entirely qualitative on outcomes; whether hypothesis "quality" was judged by humans or by models is not stated, so circularity of the automated evaluations cannot be ruled out. The in-vitro validations are the strongest evidence, but without the number of proposed candidates the base rate is unknown. Single model family throughout means the critics share the generator's blind spots (the catalog page makes the same point). Generality beyond biomedicine is claimed but not validated in the abstract. The catalog project page lists the authors in a different order (Palepu fifth) from the arXiv v2 record (Palepu eleventh); the note follows arXiv.

Key quotes

"The system's design involves agents continuously generating, critiquing and refining hypotheses accelerated by scaling test-time compute." (abstract)

"Key contributions include: (1) a multi-agent architecture with an asynchronous task execution framework for flexible compute scaling; (2) a tournament evolution process for self-improving hypotheses generation." (abstract)

"Specifically, Co-Scientist helped identify new drug repurposing candidates and synergistic combination therapies for acute myeloid leukemia, which were validated through in vitro experiments." (abstract)