Accelerating Scientific Research with Gemini in the Real-World
Samuel Schmidgall, Xiaokai Zhu, Marian Shaw, Lin Yang, Valentin Liévin, Jingyun Yang, Yuchen Zhuang, Tim Strother, Alex Bijamov, Min Woo Sun, Anil Palepu, Justin Chen, David Steiner, Jacqueline Shreibati, Wei-Hung Weng, Yilin Zhao, Xingjian Hu, Nicholas Zahn, Sadhya Garg, Julia Kirby, Yuxiang Gan, Jiaoli Li, Divy Thakkar, Shekoofeh Azizi, David Racz, Juraj Gottweis, Vivek Natarajan, Chenglin Wu, Tal Danino, Keran Rong, Haozhe Wang, Benoit Schillings, Yong Cheng, Quoc V. Le, Tao Tu
cs.AI
2026-08-27
DeepMind grounds Gemini Co-Scientist in real CVD and wet-lab work, grows monolayer TMDs first-try, and cuts invalidating paper hallucination from 90% to 4%.
Agents that can "finish a paper" inside a computer and systems that can send a recipe into a lab still sit far apart. The first group optimizes automated reviewer scores and readily fabricates results. The second group is usually locked to a narrow workflow, such as one synthesis route. What has been missing is a general research partner that can shift human-AI labor across materials, biology, and computer science, and that can check claims against experimental records.
Google DeepMind takes the Gemini multi-agent system Co-Scientist, previously used mainly for in silico hypothesis generation, and grounds it in executable work. Given a research directive, the system runs ideation, experimentation, and manuscript writing, and it varies how much a human steps in according to each domain's constraints.
All three stages use evolutionary search. Ideation samples hypotheses at temperature 1.6, has an LLM score novelty, testability, and safety, then ranks parents with a TrueSkill-style Bayesian rating plus UCB (crossover 0.7, mutation 0.3, default 10 generations). Experimentation first scaffolds on a small data slice, then switches to full-scale runs. Failed programs get structured error feedback; a 0.97 multiplicative decay on the best-program buffer keeps search from stalling. Paper writing builds a section scaffold, then parallel solvers revise it under a reviewer score that subtracts plagiarism and hallucination penalties.
Reliability is two-layered. A soft objective writes the manuscript score as reviewer minus 0.5 times plagiarism minus 1.0 times hallucination, with the hallucination term conditioned on execution logs. A fabricated result can cost a full point, enough to cancel the typical reviewer bump of at most 0.3. A hard module extracts numeric claims and rewrites mismatches to log values. If experimentation produces no valid logs, writing is aborted. Safety is also two-layered: an entry filter refuses harmful directions, then an LLM binary harm check steers ideation and planning.
Human involvement is domain-dependent. For materials, the system proposes recipes and people load the furnace. For E. coli phenotypes, the system builds the vision pipeline while experts reframe the task and run wet lab assays. For the medical agent and for 50 AI-topic papers, no one intervenes after the directive.
On a home-built 1-inch quartz-tube CVD reactor, the system searched for a safer substitute for toxic TiCl4 and ranked hexachloroethane (C2Cl6) among 272 candidate recipes. Human-in-the-loop refinement over 25 runs produced a layered 2D crystal: a strong XRD peak at 2θ = 7.8° (interlayer spacing 1.13 nm) and an in-plane STEM d-spacing of 2.51 Å, consistent with wet-etched Ti3C2Tx MXene. Yield is low, the product oxidizes, and the paper states that atomic phase assignment needs cross-sectional STEM that was not obtained. For 2D semiconductors, recipes generated from hardware constraints alone grew monolayer MoS2, MoSe2, and WS2 on the first attempt. MoS2 triangles exceeded 50 μm, with Raman E2g–A1g splitting of 21 cm⁻¹. Full evolutionary ideation took about a day of test-time compute. Gemini 3 Deep Think produced machine-executable recipes in minutes and finished first growth of all three materials in about an hour, at the cost of smaller, less regular crystals.
In biology, the task was to interpolate unpublished swarming colony images of engineered E. coli across held-out IPTG concentrations. Co-Scientist implemented leave-one-out generation with Gemini 3 Pro Image (N=16 candidates, Gemini 2.5 Pro as selector). Mean radius, polar eccentricity, and circumferential intensity CV tracked experimental dose curves (interaction p = 0.593, 0.451, 0.712). Circularity diverged (p = 0.002), with generated colonies slightly more regular. The pLac-gfp control was correctly predicted to have no dose response. This is interpolation along a known gradient, not extrapolation to new circuits.
In computer science, search over 1,282 synthetic health queries produced AgentH, an eight-phase inference-time pipeline: triage, decomposition, 28–48 parallel candidates, a knockout tournament plus three-judge vote, up to five critique rounds, citation audit, and compression toward 2,000 characters. Cost is 40–80 LLM calls per query. Under a Gemini 3.5 Flash judge with length adjustment, AgentH scores 0.377 on HealthBench Hard versus 0.334 for GPT-5, and 0.643 on Professional, first among reported systems. Three board-certified physicians compared AgentH with unscaffolded Gemini 3.1 Pro on 106 questions. Harm likelihood dropped (p = 0.0486); the other eight dimensions were not significant. Autorater-physician agreement on absolute preference was low.
The cleanest reliability test is paper generation: 50 AI topics, three matched conditions, 30 experts, 450 reviews.
| Metric | Co-Scientist | Ablated | Agent Laboratory |
| Invalidating result hallucination (≥5) | 4% | 46% | 90% |
| Complete fabrication (≥8) | 0% | 40% | 44% |
| Severe method-code mismatch (≥5) | 24% | 52% | 100% |
| High-severity derivative content (≥3) | 16% | 50% | 60% |
The system refused 98.7% of harmful research directions (691/700) and incorrectly refused 3.1% of benign ones. With the entry filter off, ethics modules still yielded 96.3% safe ideas and 96.7% safe plans.
The work moves an "AI scientist" onto three real experimental surfaces, with very different autonomy. CVD still needs humans to load samples and pick recipes; biology needs experts to reframe the task; only code-and-paper loops run fully unsupervised. The number that should change how people build these systems is the reliability ablation: the same Gemini stack, without penalties and log clipping, jumps from 4% to 46% invalidating result hallucination. AgentH shows that inference-time architecture search can win length-sensitive medical benchmarks. Blinded physicians only backed a safety gain. Beating an autorater is not the same as being clinically better.
The MXene-like phase is not atomically confirmed, and recipes were not tested in other labs. Colony prediction stays on IPTG interpolation; Gemini 3 Pro Image sometimes painted colonies with an unnatural green glow, filtered by rejection sampling. AgentH has no compute budget, so 40–80 calls per query rule out interactive use. Without a length penalty the search inflated scores by writing longer answers. Log verification assumes deterministic stdout, which wet-lab noise does not provide. The 1.3% safety misses were all dual-use requests written in neutral language. Compositional harm, several "safe" sub-tasks combining into a dangerous trajectory, is outside the per-hypothesis filter.