Lila: Scientist Review Shrinks to a Safety Check

Key points

• A language model on top of statistical models picks the search strategy for each round of experiments.
• Scientist review shrank from watching every suggestion to a safety check and a look at surprising results.
• Shift-working operations engineers manage instrument uptime and handoffs.

Handing the “what do we try next” decision to a model

Who usually decides what a lab will make next? Until now, it has been the job of a scientist drawing on experience and the literature. What Lila Sciences wants to change is exactly who makes that decision.

Research in an ordinary lab follows a clear cycle. A scientist decides what to make next, based on experience and earlier papers. Then technicians and researchers actually synthesize and test it. When the tests are done, the research team gathers to interpret the data and plan the next round.

Lila Sciences (Lila, for short) designed the company so that artificial intelligence itself leads this cycle. It was founded in 2023 inside Flagship Pioneering, a firm that builds and funds new companies, and its chief executive officer is Geoffrey von Maltzahn. The company says it is building “scientific superintelligence,” AI that carries out the scientific method itself, from hypothesis to experiment to the next hypothesis. MIT Technology Review reported in September 2026 that Lila raised $550 million in venture funding at a valuation above $1.3 billion.

Lila targets experimental research and development (R&D) in life sciences, chemistry, and materials. Its core tool is the AI Science Factory (AISF), an automated lab where models pick the experiments and instruments carry them out directly.

Lila sells this system in two ways. One is Lila Catalyst. A customer’s R&D team uses Lila’s own model, Lila Iris, and the lab capacity of the AISF as a service. The other is Lila Creation. A partner brings a problem it wants solved, and Lila returns verified results such as material structures, synthesis protocols, and data packages.

How models, instruments, and people mesh in one lab

How this automated loop actually runs is described in detail in a preprint that Lila researchers posted on September 24, 2026, and in a company post published the next day. Twenty-two Lila employees searched for catalysts for green hydrogen production, which splits water by electrolysis to get hydrogen. The goal was to find catalysts that use no iridium or ruthenium, two scarce and expensive rare metals. In this campaign (a single exploration project), the platform tested a total of 2,942 catalysts across 26 elements and 53 material systems. The paper describes the setup as “AI-guided, human-supervised” and more than 90% automated. The work flows through three layers: models, instruments, and people.

The first is what the models decide. At the start of each round, the system retrains statistical models that predict catalyst performance, using all the experiment data collected so far. An automated evaluation step picks the model with the best predictive power.

On top of that works a “sequential learning agent.” This agent scores candidate compositions using several search strategies. It puts numbers on whether to explore wide into unknown areas with no data, dig deeper into the best results so far, or actively hunt for rare high performers.

At the very top of this statistical system sits a language model (LLM). Here, the language model does not predict catalyst performance directly. In the paper’s words, the LLMs act as “strategic reasoning agents.” They decide which search strategy to adopt in this round, make the final choice of which of the top candidates picked out by the statistical models to synthesize, and write down the reason for each choice.

That is not all. The system chooses the annealing temperature, the heat treatment used to bake the thin films. When a standard measurement is unclear, it asks for extra composition measurements on its own. It also decides which catalysts move up to the long-term testing stage. According to the paper, the model usually took only a few minutes to produce these suggestions.

The second is the physical work done by lab instruments. The synthesis equipment coats thin films onto a substrate using reactive magnetron sputtering, making 48 catalysts per fixture, the sample plate. Once the samples are made, optical and X-ray fluorescence instruments automatically check quality and composition.

Next, the sample plate moves to a measuring instrument holding a strongly acidic solution, where the catalysts go through tests of electrochemical activity and stability. After the tests, the samples’ condition is analyzed again. The catalysts selected here move on to long-term tests lasting up to 4 hours, and the best leading candidates are tested in operation for up to 1,000 hours.

Every measurement from the instruments is sent automatically to a central database. What result came out, what the model predicted, and on what grounds the decision was made are all recorded. The next round retrains the models on this accumulated record. Lila reported that the shortest cycle from making a sample to analysis after testing was about 22 hours, and that about 240 catalysts were screened per week.

The third is where people step in. In this campaign, people physically carried the fixtures holding the samples between inspection stations and pressed the start buttons on the instruments. Robots to handle that physical movement were still being built.

People also acted as a gate that approved the AI’s choices. But this gate changed shape over time. Ken Jenewein, an electrochemist at Lila, described the situation then: “At the beginning, I was babysitting all the suggestions and overruling a few.” But as the AI kept pointing out the candidates he would have picked himself, he slowly cut back on checking each one.

By the last stage of the campaign, it had simplified to people only checking whether a proposed material was safe to synthesize. Rafael Gómez-Bombarelli, one of the two scientific leads, said, “Humans were in the loop, but we let the AI call the shots.”

People also took on the role of verifying results when they went against conventional wisdom. After running only a few learning cycles, the model pointed to palladium-based oxides as a leading candidate. Researchers had all but ruled out this family for acidic-environment catalysts. John Gregoire, the other scientific lead, said, “This combination defies all traditional wisdom.”

The team doubted the result and repeated the tests on the instruments. Gómez-Bombarelli said that without a physical lab to confirm the result, “we would have chalked it up to a hallucination.” The verified leading candidate then ran successfully for more than 1,000 hours in a strongly acidic solution. The first campaign, run while the infrastructure was still being built, took four months. The second campaign, with the same setup in place, finished in four weeks.

How the human review gate narrowed

An organization built like a production factory

To support this loop, Lila built its organization closer to a production floor than a research lab. Looking at the company’s job postings, a Research Operations team runs the automated lab workflows in shifts. That includes a second shift running from 2 p.m. to 10 p.m.

Engineers on this team directly handle liquid handlers and integrated workcells, watch quality-control indicators, and hand off the state of the instruments at each shift change. Shift supervisors flag deviations in the process and keep verifying that automated workflows are accurate compared with the old manual methods.

Robotics Operations staff monitor the production robot fleet in the field and pause or reset a robot when a problem comes up. Their working hours are also set to expand.

In the chemistry AISF group, a program lead owns the overall scientific direction, and a product manager decides which experimental capabilities to add to the system next. This product manager decides not only the details of the capabilities but also how much autonomy to give each one.

The guidance given to researchers on the floor is different from an ordinary lab, too. A posting for automated chemistry tells research associates to “Flag unusual features or results for scientist review.”

Lila is also separately hiring forward-deployed engineers, who build agent-based workflows directly inside customers’ instruments and data systems, and a group of AI safety scientists. As of October 22, 2026, Lila had about 110 job postings.

Four points that set it apart from traditional research

What Lila shows is not just a faster way of working compared with a conventional R&D lab.

The first difference is who decides the next experiment. In a traditional lab, a scientist picks the next sample from intuition, accumulated experience, and the literature. At Lila, that decision has moved to a bundle of models. The scientist’s touch has moved to two points: a safety check before synthesis to screen out hazardous materials, and an after-the-fact audit when a surprising result appears. The palladium case shows why this audit matters. The AI chose a direction a scientist would not have chosen, and the only ground on which scientists could trust that choice was the measured values from the instruments.

The second difference is how the AI is put together. Lila did not leave every decision to a language model alone. In the comparison analysis in the paper, a frontier language model that picked the next experiment by looking only at result data, without help from the prediction models, “performed even worse than random search.” In a second test, too, LLM-only selection did not reliably reach the palladium family within a fixed budget. The language model did its part only when it steered the direction of exploration on top of statistical models that systematically track the uncertainty in the data.

The third difference is how the organization operates. The people closest to the instruments are not researchers but operations engineers working in shifts. Their main jobs are keeping uptime, quality control, and handoffs between shifts. This looks more like a data center or an advanced manufacturing plant than a university lab bench. Scientists, who think about the direction of research, step back from routine synthesis work and take program and product management roles that decide which experimental tools to give the AI system next.

The fourth difference is how it handles the output. Lila sells lab capacity and the model itself as a service. According to the Lila Catalyst page, customer scientists can focus on “strategy, design decisions, and downstream development” instead of running complex experimental workflows themselves. If this works, companies could buy experiments as needed, the way they rent cloud computing, instead of spending huge sums to build in-house labs and hire researchers.

A shift in R&D’s center of gravity, and lessons for the field

If this way of operating spreads, the competitive landscape between contract research organizations (CROs) and in-house corporate R&D could be shaken considerably. Until now, the most precious resource in R&D has been the time of skilled hands keeping the lab bench running. But if equipment automation combined with models lowers the cost of physical execution, the scarce resource could shift from time spent filling the bench to the quality of the decision about what to run. An organization that owns the loop that decides what to run could capture far more value than one that only carries out the work. And if, as Lila explains, every experiment result flows into the retraining data for the models, the platform’s predictive ability could grow stronger with every customer project it runs.

A scientist’s way of working could narrow and rise to a higher level at the same time. Time spent picking samples one by one shrinks, but setting goals, defining the experimental capabilities the system needs, checking safety, and confirming whether a surprising result that breaks conventional wisdom is true all remain with scientists. A researcher on Lila’s open-endedness team asks what the work would look like if every scientist led a swarm of AI scientists.

This case offers practical guidance to other teams that want to bring AI agents into their organizations, whatever their size.

First, it is safer to put a light statistical model that can calculate uncertainty under the large language model. A language model is useful for setting the broad exploration strategy and laying out the reasons for a decision. But if it is left alone to do precise numerical evaluation or ranking, it can do worse than random search.

Second, put the gate where people step in at the most important point, and move it gradually based on data. Instead of aiming for full automation from the start, it is better to watch each suggestion closely at first, then step back to a minimal gate such as a safety check once the system’s judgment stabilizes. That can reduce resistance in the organization.

Third, record all results and the decision process in a single database. Only when it is transparent which prediction led to which decision, and what result came out of the instruments, can you keep retraining the model and run after-the-fact audits.

Fourth, keep a real laboratory that can physically verify the model’s suggestions and say “no.” The only reason the Lila researchers could believe the palladium oxide no one had expected is that an instrument put it in strong acid and checked. In a loop with no fast check that touches the real world, there is no way to tell an AI discovery from a plausible hallucination.

What is still unproven

Whether this approach can become the standard in every field of research remains to be seen.

Most of the workflow information published so far rests on Lila’s own announcements. All 22 authors of the paper are Lila employees, and the company paid the full cost of the research. Related patents have already been filed or may be filed. Among outside media reports, there is none yet that independently verified and described this automated loop from the inside.

The performance comparison has limits too. The comparisons with earlier search methods and with LLM-only search are not results from separate live campaigns. They are after-the-fact simulations based on experiment data already collected. The claim that it is 17 times faster than a standard lab also follows a baseline the company set itself, not a yardstick from an independent body.

Fully unmanned operation of the lab is also still unfinished. In the published campaign record, moving samples to the next instrument was still done by people. The robot fleet that appears in the job postings is only now being added in the field. The company’s total headcount, and how many campaigns it can run stably at the same time, cannot be confirmed from published material.

Finally, the catalyst found this time is a lab-level materials discovery, not a commercial product that can go straight to market. Whether this catalyst can survive the long lifetimes industry demands, and whether it would perform the same when built into the large-scale equipment used in factories, still has to go through further testing.

---

Sources

1. Lila Sciences (Jenewein, Habib Zadeh, et al.) — AI-guided high-throughput discovery of iridium- and ruthenium-free palladium-oxide catalysts for durable acidic oxygen evolution, arXiv 2609.30133 — https://arxiv.org/abs/2609.30133 — 2026-09-24
2. Lila Sciences — How an AI-run lab cracked open green hydrogen’s catalyst problem — https://www.lila.ai/news/how-an-ai-run-lab-cracked-open-green-hydrogens-catalyst-problem — 2026-09-25
3. Lila Sciences — What’s a science startup doing with an open-endedness team? — https://www.lila.ai/news/whats-a-science-startup-doing-with-an-open-endedness-team — 2026-09-10
4. Lila Sciences — From prompt to CAR-T: Inside Lila’s RNA discovery engine — https://www.lila.ai/news/inside-lilas-rna-discovery-engine — 2026-08-28
5. Lila Sciences — Technology — https://www.lila.ai/tech — accessed 2026-10-22
6. Lila Sciences — Lila Catalyst — https://www.lila.ai/lila-catalyst — accessed 2026-10-22
7. Lila Sciences — Lila Creation — https://www.lila.ai/lila-creation — accessed 2026-10-22
8. Lila Sciences — job postings: Engineer I/II, Research Operations; Shift Supervisor, Research Operations; Robotics Operations Engineer I; Senior Research Associate, Automated Chemistry; Product Manager, Chemistry AI Science Factory; Staff Forward Deployed Engineer, Physical Sciences — https://job-boards.greenhouse.io/lilasciences — accessed 2026-10-22
9. MIT Technology Review — Benjamin Kompa, 35 Innovators Under 35 — https://www.technologyreview.com/innovator/benjamin-kompa-autonomous-science-tools/ — 2026-09-08
10. Flagship Pioneering — Lila Sciences — https://www.flagshippioneering.com/companies/lila-sciences — accessed 2026-10-22