Gemini Co-Scientist moves from hypotheses to execution-grounded research

Summary

A Google DeepMind-led preprint reports an execution-grounded extension of Co-Scientist across materials science, biology, and computer science. The system generated machine-executable CVD recipes that produced three monolayer semiconductors on the first attempt, built an E. coli phenotype predictor checked against unpublished wet-lab measurements, and autonomously discovered an inference-time medical-response architecture. A double-blind study of generated papers found fewer severe hallucinations and less plagiarism than unconstrained baselines.

AI role

Generated and executed research plans, translated laboratory constraints into machine-control code, predicted wet-lab phenotypes, searched agent architectures, and checked manuscript claims against execution logs.

Narrative role

This is a substantive follow-up to the original hypothesis-generation system: it moves the Co-Scientist timeline into physical execution, laboratory feedback, autonomous computational experimentation, and empirically evaluated research-integrity controls.

Caveat

The evidence is an author-team v1 preprint without independent or cross-laboratory replication. Human researchers directed the studies, handled samples, refined tasks and recipes, and ran more than 70 MXene experiments; the MXene phase remains unconfirmed, and the HealthBench agent initially exploited verbosity and requires 40-80 model calls per query.

Related events

• AI co-scientist frames hypothesis generation as a multi-agent research workflow
• AI co-scientist hypothesis matches bacterial gene-transfer mechanism