中文

Direction analysis

Lila Sciences

Turns multi-domain automated laboratories into training environments where physical experiments produce verifiable scientific tokens.

Last verified AI4S.fyi's own reading of the public record; its sources are listed at the end

← Lila Sciences organization page

How the loop runs

Operating sequenceResearch objective and model stateIris proposes candidates and experimentsAutomated experiment executionInstrument results become the verifierUpdate model and select the next round[S1][S2][S3]

System architecture [S1][S2][S3]
Reasoning / orchestrationIris, hypothesis generation, experiment planning and scheduling
ExecutionInstruments, robots, planar motors and human operators below the API
EvidenceRaw measurements, provenance, failure states and verifier signals
Learning loopReward signals, model updates and next-experiment selection

Current public boundary. An architecture from hypothesis to physical experiment to feedback; cross-instrument autonomy, transfer and unit economics remain undisclosed.

  • Humans may sit below the API; automation does not necessarily mean fully unattended operation.
  • Successful execution is not the same as a successful scientific discovery.
  • The reliability of the verifier itself still needs independent validation.

The directions that can currently be confirmed are: a disclosed architecture in which a physical laboratory acts as the verifier for a reinforcement-learning loop, an mRNA stability result reported with specific numbers, in-vivo CAR-T pharmacodynamics in non-human primates presented at a conference, three DOE Genesis Mission Phase I awards with named laboratory partners, an iridium-free oxygen-evolution catalyst reported in the company's first preprint, and $550 million raised. The preprint (arXiv, September 24, 2026) is the one Lila result written up in full with its methods; it has not been peer reviewed, and a Europe PMC search for Lila Sciences as an affiliation still returned zero results on October 7, 2026. The mRNA and primate results are public only as a company post and a conference abstract.

Scope note: this page tracks the public record only: the company's own releases, the peer-reviewed literature, and third-party reporting. No inside information. It records how far public material goes, not how well the company is doing.

Overview

DirectionWhat can be confirmedLatest evidence
1. The thesisThe architecture is disclosed: experiments supply the reward signalCompany pages, 2026
2. mRNA stability950,000 sequences built and measured; EPO expression on day 15 matching benchmark day 2Company post, 2026-08-28
3. In-vivo CAR-TIn non-human primates, B cells near zero to day 7 and about half of baseline at day 14; the benchmark back at baselineASGCT 2026 abstract; company post, 2026-08-28
4. Genesis MissionThree Phase I awards with Caltech, LBNL, Northwestern, Argonne and NLRAnnounced 2026-07-22
5. Oxygen-evolution catalystInMnPdOx held below 0.5 V overpotential for 1,000+ hours in acid; company-run, not peer reviewedarXiv preprint, 2026-09-24
6. Capital and build-out$550M total; an Alewife lab given as 244,000 or 270,000 sq ft depending on the sourceSeries A closed 2025-10-10

1. The thesis: a laboratory as the verifier

What can be confirmed: the company states its bet plainly. "Science is the ultimate token generator. Every experiment produces new data that has never existed before. The internet data that powered the current generation of AI has been exhausted, but scientific data generation never runs out."

The mechanism follows from that. "Verifiable rewards have driven AI breakthroughs in math and coding, but most of science is hard to verify. Lila's AI Science Factories change that. By running real experiments at scale, they act as the world's largest verifier for science, providing the reward signal the model needs to learn." The reasoning model is called Lila Iris; the instrument network is the AI Science Factory.

Why this is the right frame to judge the rest by. The argument is not that automation makes experiments cheaper. That claim is old and many companies make it. It is that the experiment is the reward function, which is the missing piece that made RL work for mathematics and code and has never been available for chemistry or biology. If that holds, the interesting number is not how fast the robots move but how much the model improves per experiment. That number has never been published.

The fullest public account of the architecture is not on the company's website but in a 2026 Latent Space interview with Lila's Andy Beam and Rafael Gómez-Bombarelli, co-founder and Chief Scientific Officer of Physical Sciences (Lila team page), which this site keeps a source note on. It is where the honest edges are: instruments take real work to onboard, humans sit below the API for some steps, and a verifier can be reward-hacked like any other.

Boundary: an architecture, described by the company that built it. Confidence: high that the architecture is as described; low for the scaling-law claim built on it, which is a thesis and has no published measurement behind it.

Open leads: any curve showing model performance against experiments run; the cost of one usable scientific token; how much of what the factories produce is reusable rather than program-specific.

2. mRNA stability: the most specific claim they have made

What can be confirmed: the company reports that its platform "has tested more than 950,000 mRNA sequences, designing, building, and measuring each one in living cells." For erythropoietin, Lila reports sequences "that produced as much protein on day 15 as benchmark mRNA did on day 2." Follow-on half-life experiments in human cells are reported across five more payloads (Factor IX, IL-10, IL-22, eGFP and firefly luciferase), with "substantially longer half-lives than published sequences from major pharmaceutical companies and published benchmarks."

The company states its own limit: "It should be noted that these comparisons are against the best published sequences; proprietary industry formulations aren't disclosed and couldn't be compared." So the comparison is against what has been published, not against what the industry actually uses. The post also says all its results are preclinical, reflect Lila's internal testing and have not been independently peer-reviewed.

Boundary: expression and half-life in cells. A durable mRNA is an input to a therapy, not a therapy. Confidence: medium, numerically specific, self-limited, and entirely company-reported; no sequences, no dataset and no publication accompany it.

Open leads: release of any of the 950,000 sequence–measurement pairs; a head-to-head run by an outside laboratory; a peer-reviewed write-up of the EPO result.

3. In-vivo CAR-T: non-human-primate data in a conference abstract

What can be confirmed: Lila's August 2026 post says that in work "presented at the 2026 ASGCT meeting in Boston," three Lila scientists ("none with prior expertise in CAR-T biology") designed anti-CD20 CAR mRNAs, formulated them into CD8-targeted lipid nanoparticles, and tested them through cell culture and mouse models into non-human primates dosed intravenously. The post's word "presented" links to the meeting abstract, "Accelerated Development of High-Performance CAR-T Therapies Through Autonomous Science" (ASGCT 2026, abstract 3251), which lists twelve authors, all at Lila Sciences, with Yue Yang first.

The post's summary: "Lila's mRNA payloads drove peripheral B cells to near-zero depletion by day 4 and maintained that depletion through day 14, as compared to benchmark sequences that rebounded past baseline B cell levels by day 14." The programme is described as taking "under six months" at "an estimated ten-fold reduction in R&D costs."

The abstract gives doses and words the primate result differently. At a single 0.5 mg/kg dose, treated animals showed "near-complete peripheral B cell depletion through Day 7, with levels remaining suppressed (~50% of baseline) at Day 14," while "benchmark-treated animals returned to baseline by Day 14." In mice, a single 0.25 mg/kg dose of Lila's sequence depleted B cells more than the benchmark at 1.0 mg/kg, and "No significant toxicity events were observed." The abstract puts the work at "within three months" of applying the platform; the post says under six.

What is public. Non-human-primate pharmacodynamics is a real, expensive, hard-to-fake stage. Its public record is the company post and a one-page abstract, both by Lila. The abstract gives no animal count, error bars or statistics, and no paper, preprint or protocol has followed. B-cell depletion is also a pharmacodynamic marker: evidence that the construct did what it was designed to do in an animal, not evidence of benefit to a patient.

Boundary: animal pharmacodynamics. Confidence: medium. The claim is specific and dated, made by named scientists in a meeting abstract that gives doses; nothing beyond Lila's own reporting has been published to check it against.

Open leads: animal numbers and statistics; a paper or preprint; whether a partner or acquirer independently reproduces the construct.

4. Genesis Mission: three Phase I awards

What can be confirmed: announced July 22, 2026. Three proposals into the DOE Genesis Mission's Lighthouse Challenges were selected for Phase I awards, each with a named partner.

  • AI-Enabled Electrochemical Refinery with Caltech and Lawrence Berkeley National Laboratory: a closed-loop electrocatalytic platform navigating "more than 10 billion possible combinations of feed gases, electrolytes, and operating conditions" to convert waste carbon into chemical products.
  • Agentic AI-Driven Manufacturing of Polyethylene from Waste Carbon with Northwestern University and Argonne National Laboratory: reasoning across an electrochemical and a thermochemical stage at once.
  • Autonomous Identification of Semiconductor Charge Transport Physics with the National Laboratory of the Rockies: separating "true physical phenomena from measurement artifacts across more than 1,000 samples" in ZnO and AlN, and generating "new, testable equations for leakage current and breakdown."

What a Phase I award is, stated by Lila itself: several months of concept development, after which "successful Phase I teams then have the opportunity to apply for larger Phase II awards to build and deploy." These are funded intentions with named partners and an explicit "AI advantage" test to meet later.

Boundary: selection, not result. Confidence: high that the awards and partners exist; by construction there is no scientific output yet.

5. An oxygen-evolution catalyst: the first preprint

What can be confirmed: on September 24, 2026 Lila posted its first preprint, "AI-guided high-throughput discovery of iridium- and ruthenium-free palladium-oxide catalysts for durable acidic oxygen evolution" (arXiv 2609.30133). All 22 authors are at Lila Sciences, and the paper states that they are its employees. The corresponding authors are John Gregoire, Lila's Chief Autonomous Science Officer (Lila team page), and Rafael Gómez-Bombarelli. The paper describes an "AI-guided, human-supervised closed-loop platform (>90% automation)" that evaluated 2,942 catalysts across 53 material systems and 26 elements for acidic oxygen evolution, the anode reaction of PEM water electrolysis, which today relies almost entirely on iridium or ruthenium. Its most stable material, InMnPdOx, kept its overpotential below 0.5 V over 1,000 hours in 1 M sulfuric acid at 10 mA cm⁻²; two other palladium oxides in the paper, PdOx and NiTaPdOx, passed 0.5 V at about 200 and 470 hours. In a retrospective benchmark, the paper reports that its agent advanced the activity–stability frontier faster than fixed-policy Bayesian optimisation or a language model choosing in context. It says code and data to reproduce the main figures are on GitHub.

Boundary: a physical test, run and reported by the company, posted as a preprint and not peer reviewed; no outside lab has tested the catalyst. The timeline records it at rung 4, tested in the physical world. Confidence: high that the preprint reports these results; medium for the catalyst's performance, which only Lila has measured.

Open leads: the peer-reviewed version; an outside lab's long-run test of InMnPdOx.

6. Capital and the physical build-out

What can be confirmed: the Series A closed October 10, 2025 at $350 million, bringing total funding to $550 million including the $200 million seed. Named investors include Flagship Pioneering (where the company originated), NVentures (NVIDIA's venture arm), Analog Devices, IQT, Braidwell, Collective Global, General Catalyst, ARK Venture Fund, a wholly owned subsidiary of the Abu Dhabi Investment Authority and the State of Michigan Retirement System. Reuters reported on October 14, 2025 that the valuation topped $1.3 billion.

The facility figures differ by source. Lila's August 2026 RNA post describes its newest AI Science Factory as "a 270,000-square-foot purpose-built facility in Alewife, Massachusetts, where its AI scientific reasoning model and an autonomous laboratory will operate as a single closed-loop system," to run experiments "continuously, 24 hours a day." The landlord, IQHQ, announced a 244,000-square-foot lease at Alewife Park in October 2025, covering move-in-ready space and a new laboratory building. Andrew Beam, Lila's CTO, said at Ray Summit 2026 that Lila runs a 15,000-square-foot lab in Cambridge now and plans 100,000 square feet at the end of the Red Line in early 2027 (his account).

Kenneth Stanley, Lila's SVP of Open-endedness (Lila team page), has built Lila's open-endedness team since 2025; Lila's September 2026 profile places the team in San Francisco.

Boundary: financing and facilities. Confidence: high for the financing (company statements, named investors, dated); medium for the facility's size, which the sources give differently. The Series A post says the company is "welcoming our first cohort of customers now"; no customer is named.

What could not be found

  • No peer-reviewed publication under a Lila Sciences affiliation. A Europe PMC search for the affiliation returned zero results on September 17 and again on October 7, 2026. The one preprint is the catalyst paper in section 5.
  • No paper, preprint or protocol for the ASGCT CAR-T work. The meeting abstract gives doses but no animal count or statistics.
  • No sequences, model, weights or runnable benchmark for the biology work. Nothing from the 950,000-sequence dataset is public.
  • No operating metrics for the AI Science Factories: throughput, hit rate, unattended fraction, cost per experiment, failure recovery. These are what would make "the world's largest verifier for science" a measurable claim rather than a description.
  • No named commercial customer, despite a stated first cohort.
  • No benchmark behind the comparative claim on the company's own homepage that "LILA consistently outperforms other models across scientific domains." No model, task or competitor is named.

Method and scope

Built from the company's own technology, solutions and newsroom pages, its Series A and Genesis Mission announcements, its August 2026 RNA write-up, the ASGCT 2026 abstract it links, the September 2026 arXiv preprint, Lila's team pages, IQHQ's lease announcement, Andrew Beam's Ray Summit 2026 talk, and affiliation searches of Europe PMC for any published output. Every figure quoted above was read at its source.

Confidence is assigned claim by claim, not to a whole direction. The labels:

  • High: a public artifact, execution by an outside institution, or review by a named expert;
  • Medium: an official release or company-reported figure (specific and inspectable, unverified outside);
  • Low: a single-source claim, a thesis, or a claim with no published measurement behind it.

Two operating rules: revisions are dated by version and are not treated as new projects; evidence that a project exists is not evidence that the technology has succeeded.

One comparison is worth drawing, because it is about the shape of evidence rather than the quality of science. Microsoft publishes its models, and an outside crystallographer was therefore able to contest one of its results in a peer-reviewed journal. Lila's mRNA and primate results are further along the path toward a medicine than anything in Microsoft's portfolio, and they are public only as a company post and a conference abstract, with no sequences or full methods for anybody to contest. Its catalyst preprint is the first Lila result written up with its methods. Neither of those facts settles which company is doing better work. They do settle which claims a reader is in a position to check.

Primary sources

Evidence timeline

DateEventNew evidenceWhat it does not proveSources
First preprint: an AI-guided lab finds an iridium-free oxygen-evolution catalystThe closed-loop platform screened 2,942 oxides across 53 systems; the best, InMnPdOx, ran 1,000+ hours in acid. It is Lila's first public experimental result that can be checked line by line.All 22 authors are Lila employees and it is not peer reviewed; iridium was not tested, and on the replay benchmark plain active search did about as well as the agent at reaching palladium.PREPRINT · COMPANY AUTHORS ↗
The physical lab as an RL verifierA public interview clarifies how experimental successes and failures become reward signals and next-round data, while adding detail on instrument onboarding, human-executed steps, concrete programs and reward-hacking boundaries.Program performance, ten trillion tokens, cross-domain advantage and business economics are primarily company-reported and lack independent validation.INTERVIEW · COMPANY-REPORTED ↗Source note
Series A closes at $350M, $550M raised in totalThe company says the money builds out AI Science Factories and opens the platform to a first cohort of customers.A round shows where capital is betting, not that the platform has produced discoveries.COMPANY ↗
Emerges from stealth with $200M in seed capitalFlagship Pioneering unveils Lila, aiming at 'scientific superintelligence' and fully autonomous labs.The aim is the company's statement, not a demonstrated capability.COMPANY ↗

On evidence. A CEO claim is not a benchmark; a benchmark is not a physical experiment; a physical experiment is not manufacturing scale; a third-party test is not qualification; qualification is not deployment. Every line keeps its source and says what it supports and what it cannot show.

Sources

  1. Lila Technology · Company
  2. Lila Solutions · Company
  3. RL with Verifiable Rewards, but the Verifier is a Lab · Interview
  4. AI-guided discovery of iridium- and ruthenium-free palladium-oxide catalysts (arXiv 2609.30133) · Preprint

← Lila Sciences organization page