Google DeepMind says Gemini Co-Scientist operated a CVD reactor

• Co-Scientist has moved into execution-grounded research: omarsar0's paper summary says it interfaced with a semi-automated CVD reactor while human experts executed the selected materials recipe.
• The materials team used a non-hazardous C2Cl6 precursor route for a Ti3C2Tx MXene target, but iScienceLuvr's paper thread notes that the resulting 2D material still needs atomic-structure confirmation.
• In biology, sparse images were enough for Co-Scientist to forecast engineered E. coli swarming across IPTG concentrations, with predictions that iScienceLuvr's paper thread says largely matched unpublished wet-lab measurements.
• In computer science, the system autonomously designed an inference-time scaling architecture that beat six frontier models on two HealthBench subsets, according to omarsar0's paper summary, while blinded physicians also assessed lower potential clinical harm.
• The paper attributes lower hallucination and plagiarism to reliability modules tested through 450 expert reviews, as omarsar0's paper summary reports.

Google's May Co-Scientist announcement described a Gemini multi-agent system for generating, debating, and evolving hypotheses. The new preprint shifts the reported validation into experiment design, physical execution, and generated research papers.

The materials result used a split workflow: Co-Scientist designed the route and interfaced with the semi-automated reactor, while humans physically ran and optimized the top-ranked recipe.

• Co-Scientist selected C2Cl6 as a non-hazardous precursor for a Ti3C2Tx MXene growth target, according to iScienceLuvr's paper thread.
• Microscopy and diffraction indicated that the product was a lamellar 2D material with structural similarities to the Ti3C2Tx lattice; omarsar0's paper summary says further experiments are required to establish its atomic structure.
• In separate transition-metal dichalcogenide trials, Gemini 3 Deep Think tailored growth recipes to lab constraints in minutes and produced monolayer MoS2, MoSe2, and WS2 in one attempt, as omarsar0's paper summary describes.

The biology configuration predicted emergent swarming phenotypes of engineered E. coli from sparse imaging data across an IPTG concentration gradient.

Those predictions largely matched unpublished wet-lab morphological measurements, with iScienceLuvr's paper thread characterizing the validation as quantitative. The preprint frames the result as a way to reduce experimental screening cycles.

For the computer-science experiment, Co-Scientist operated autonomously to design an inference-time scaling architecture. It outperformed six frontier models on HealthBench Hard and HealthBench Professional, iScienceLuvr's paper thread reports.

The evaluation also used blinded physician review of potential clinical harm, where the authors reported a significant but modest reduction, according to omarsar0's paper summary. The public summaries identify neither the architecture's mechanism nor score deltas.

The paper also evaluated end-to-end generated papers in a double-blind study: 30 domain experts supplied 450 independent reviews, as described in omarsar0's paper summary.

The authors report that Co-Scientist's reliability modules reduced hallucination and plagiarism while improving research safety in that study.