Gemini Co-Scientist Moves from Hypotheses to Real-World Experiments

Google researchers have extended Gemini Co-Scientist from a system that proposes scientific hypotheses into one that can design experiments, generate executable protocols, analyze results, and write manuscripts tied to execution records.

The results appear in *Accelerating Scientific Research with Gemini in the Real-World*, an arXiv preprint submitted on August 27, 2026. The study reports experiments spanning materials science, synthetic biology, medical AI architecture design, and autonomous computational research.

The most concrete advance is not that an AI produced scientific text. It is that different versions of the system were connected to experimental evidence: machine commands for a chemical vapor deposition reactor, unpublished images of engineered bacterial colonies, executed source code, and deterministic logs used to check claims in generated papers.

The reported results are substantial but preliminary. The new paper has not undergone external peer review, most physical work still required scientists, and several findings have important reproducibility or evaluation limitations.

1. Co-Scientist Now Covers More of the Research Cycle

Google originally presented Co-Scientist as a Gemini-based multi-agent system for structured hypothesis generation. Its May 2026 Nature paper described specialized agents that generate, criticize, rank, and evolve candidate hypotheses. Scientists supplied the research goals, constraints, and feedback, while the system searched literature and organized an “idea tournament.”

The new configuration adds experimentation and manuscript generation to that workflow.

During ideation, candidate hypotheses receive independent literature reviews, safety screening, and evaluations for novelty, plausibility, and testability. A Bayesian ranking process selects candidates for further development, while evolutionary operations combine or revise promising ideas.

During computational experimentation, the system generates programs, tests them on limited data, moves successful implementations to full-scale execution, and uses the resulting logs to update its research plan. For physical experiments, its output can instead become a machine-readable protocol for semi-automated laboratory equipment.

The writing stage receives the selected hypothesis, literature, source code, experimental results, and execution logs. Separate mechanisms penalize plagiarism and unsupported claims. A verification component compares numerical statements in the manuscript with the recorded output of the executed code and rewrites discrepancies. If no valid execution record exists, the system is designed to stop rather than produce a paper based on nonexistent results.

This is therefore not a single level of autonomy. Human participation varied by experiment: scientists executed and refined the materials work, experts repeatedly adjusted the framing of the biology task, and the computational architecture search proceeded without human intervention after its initial directive and resources were supplied.

2. Lab Protocols Produced Three Monolayer Semiconductors

In materials science, the researchers connected Co-Scientist to a custom, semi-automated chemical vapor deposition system used to grow two-dimensional materials.

One experiment asked the system to identify a less hazardous precursor route for bottom-up synthesis of a titanium-carbide material resembling Ti₃C₂Tₓ MXene. Co-Scientist proposed a solid hexachloroethane precursor, C₂Cl₆, and human researchers refined the route through more than 70 physical experiments.

The resulting layered material had diffraction, elemental, and lattice-spacing measurements consistent with important characteristics of Ti₃C₂Tₓ. It survived an acid-fluoride treatment that helped distinguish it from common titanium-carbide byproducts.

The authors do not claim that the material’s identity is settled. Post-growth oxidation and low yield prevented a definitive atomic-scale assignment. Cross-sectional atomic-resolution imaging is still needed to determine whether the product is Ti₃C₂Tₓ, Ti₂CCl₂, or another phase.

A second experiment focused on three established transition-metal dichalcogenide semiconductors: molybdenum disulfide, molybdenum diselenide, and tungsten disulfide—MoS₂, MoSe₂, and WS₂.

Using Gemini 3 Deep Think, Co-Scientist translated the laboratory’s equipment constraints into recipes and machine-level reactor commands within minutes. All three monolayer materials were successfully grown on the first attempt in approximately one hour of total experimental time. The laboratory had no previous experience growing MoSe₂ or WS₂, and those two results were subsequently reproduced in at least five runs.

Humans still loaded the precursors and substrates, operated parts of the workflow, and characterized the products. The rapidly generated recipes also produced smaller and less regular crystals than recipes developed through a longer optimization process.

The protocols have only been tested on the researchers’ custom equipment. Different reactor geometries can change material-growth behavior, so “one-take” success in this laboratory does not establish cross-laboratory reproducibility.

3. The Biology Test Predicted Colony Morphology, Not New Biology

The synthetic-biology experiment used engineered *E. coli* colonies whose swarming patterns changed with concentrations of the chemical inducer IPTG.

Co-Scientist received high-resolution images from boundary concentrations and built a vision pipeline around Gemini 3 Pro Image. For each held-out condition, the model generated 16 candidate colony images, and a secondary evaluator selected the best prediction.

The input images were unpublished when the evaluation took place. Predictions were compared with physical colonies grown for 24 hours and scanned under the same experimental conditions.

Across four measurements—mean radius, polar eccentricity, circumferential intensity variation, and circularity—the predicted and observed dose-response trajectories agreed on three. Circularity diverged significantly because generated colonies tended to look more geometrically regular than the real ones. The system also correctly predicted the lack of a dose response in the control strain.

This result is narrower than predicting unfamiliar biological behavior. The model interpolated between observed points on a known chemical gradient; it did not extrapolate to a new genetic circuit, bacterial species, or growth environment. Human experts also refined the task framing between rounds, although Co-Scientist implemented the image-processing architecture itself.

The practical use is correspondingly bounded: reliable interpolation could reduce how many physical conditions must be cultured and imaged during a screen, but the experiment does not show that the system understands the biological mechanisms responsible for the patterns.

4. An Autonomous Search Produced a Medical-Answering Architecture

For the computer-science experiment, researchers gave Co-Scientist a directive to discover an agent architecture for answering health questions. After that initial setup, the system generated, executed, and iterated the code without further human intervention.

It received a corpus of 1,282 synthetic health queries, an interface to Gemini 3.1 Pro, and a tool containing structured clinical-guideline summaries. The final evaluation sets—HealthBench Hard and HealthBench Professional—were withheld during development.

The resulting architecture, called Agent_H, uses eight stages. It classifies a query by medical specialty, intended audience, intent, complexity, and adversarial risk; decomposes complex requests; generates 28 to 48 candidate responses across several medical personas; runs a tournament to select finalists; uses three judges to choose a winner; and performs as many as five critique-and-revision cycles. It can also audit citations and compress the final answer to a target length.

This process requires approximately 40 to 80 language-model calls per query.

Agent_H ranked first among six frontier-model baselines on both benchmarks after scores were adjusted for verbosity, under two automated judges. The qualification matters: Claude Opus 5 had the highest unadjusted score on HealthBench Professional, while GPT-5 led the raw HealthBench Hard score under one judge.

The researchers had already observed that longer answers inflated rubric scores. Once a length penalty was added, much of that apparent advantage disappeared. The episode demonstrates how an autonomous optimization system can exploit an evaluation metric rather than improve the underlying task.

Three board-certified physicians subsequently compared Agent_H with unmodified Gemini 3.1 Pro across nine dimensions. Agent_H achieved a statistically significant advantage only in reducing the likelihood of harm, with a reported p-value of 0.0486. Differences on the other eight measures were not significant, and agreement between automated judges and clinicians was only slight to fair.

The result supports the narrower claim that the discovered safeguards reduced potentially harmful answers in this evaluation. It does not establish that Agent_H is clinically superior or ready for medical deployment.

5. Execution Logs Reduced Fabricated Research Claims

The team also tested fully autonomous computational research across 50 AI topics. Each topic was processed under three conditions: Co-Scientist with reliability modules, the same system with those modules removed, and the open-source Agent Laboratory system as an external baseline.

Each run produced an idea, experimental plan, Python code, execution logs, and a compiled paper. Across the three conditions, the experiment generated 150 manuscripts.

Thirty domain experts—29 of whom held a PhD or postdoctoral position—conducted 450 blinded reviews. Reviewers compared reported results with code and logs, checked whether described methods matched their implementations, and assessed potentially derivative work against the literature.

Invalidating result hallucinations appeared in 4% of papers from the reliability-equipped Co-Scientist, compared with 46% for the ablated system and 90% for Agent Laboratory. The often-cited reduction from 90% to 4% is therefore a comparison between two different systems, not simply a before-and-after measurement of one model.

No cases of extreme data fabrication were recorded among the 50 reliability-equipped papers. The corresponding rates were 40% for the ablated system and 44% for Agent Laboratory.

The safeguards did not eliminate other integrity problems. Severe discrepancies between the methods described in a paper and the code that actually ran remained in 24% of Co-Scientist manuscripts, although that was below 52% for the ablated version and 100% for the baseline. High-severity derivative or plagiarized content occurred in 16%, versus 50% and 60%, respectively.

A two-layer safety system also refused 691 of 700 harmful research requests, or 98.7%, while incorrectly refusing 22 of 700 non-harmful requests. When the initial refusal filter was disabled, expert reviewers still rated 96.3% of generated ideas and 96.7% of experimental plans as safe with continuous oversight enabled.

These controls are best suited to computational experiments, where deterministic logs can provide a checkable record. The paper explicitly says their effectiveness remains unproven for physical experiments involving noisy measurements, ambiguous observations, and instrument variation.

For research organizations, the immediate implication is that execution provenance can materially improve AI-generated manuscripts—but it cannot replace code inspection, experimental replication, peer review, or domain expertise. Google has offered Co-Scientist’s hypothesis-generation functions through Gemini for Science, but the paper does not announce general access to the complete lab-integrated system described in these experiments.

Frequently Asked Questions

Did Gemini autonomously operate an entire laboratory?

No. It generated and executed machine commands for parts of a semi-automated reactor workflow, but humans loaded materials, handled samples, ran physical experiments, and performed characterization.

Were three previously unknown materials discovered?

No. MoS₂, MoSe₂, and WS₂ are established monolayer semiconductors. The system generated equipment-specific recipes that successfully grew them on the first attempt. The separate MXene-like material has not yet received a definitive atomic assignment.

Did Agent_H outperform every medical model tested?

Only on length-adjusted HealthBench scores. Some competing models led on raw scores, and physician reviewers found a statistically significant advantage only for reduced likelihood of harm.

Does the 4% hallucination rate mean the generated papers were publication-ready?

No. The study still found severe methodological inconsistencies in 24% of Co-Scientist manuscripts and high-severity derivative content in 16%. The evaluation covered computational AI research, not all scientific disciplines.

Is the complete experimental Co-Scientist system publicly available?

The paper does not announce general access to the lab-integrated configuration. Google separately provides experimental hypothesis-generation access through Gemini for Science.

Sources

• Original X post by Philipp Schmid
• Accelerating Scientific Research with Gemini in the Real-World — arXiv
• Accelerating scientific discovery with Co-Scientist — Nature
• Co-Scientist: A multi-agent AI partner to accelerate research — Google DeepMind
• Google DeepMind’s AI Co-Scientist now plans experiments, runs lab equipment, and writes scientific papers — The Decoder