Google's Co-Scientist can generate and refine scientific hypotheses without a human in the loop for each iteration.
Co-Scientist is a multi-agent system running on Gemini that takes a researcher's stated objectives and existing evidence, then produces candidate hypotheses for experimental testing. Agents generate, critique, and revise hypotheses in a continuous loop — a process the paper calls "tournament evolution" — while the system scales compute at inference time to improve output quality over time. The architecture uses asynchronous task execution, meaning the pipeline can flex across varying workloads without requiring a fixed compute budget. Three biomedical use cases anchored the validation: drug repurposing, novel target discovery, and explaining mechanisms of anti-microbial resistance.
The headline result is that Co-Scientist flagged drug repurposing candidates and combination therapies for acute myeloid leukemia that held up in in vitro experiments — not a simulation, but actual lab work. That matters because the recurring failure mode of AI in science has been systems that look convincing on benchmarks and collapse when a pipette gets involved. Clearing the in vitro bar is modest but real.
The broader pitch is that AI can compress the slowest part of research — the speculative, iterative grind of hypothesis generation — rather than just automating literature search. Whether that compression survives contact with more complex disease models, or with research domains outside the relatively structured world of drug repurposing, is the question the paper does not answer.