# Edison Scientific

> Source: https://aiwiki.ai/wiki/edison_scientific
> Updated: 2026-10-27
> Fact-checked: 2026-10-27
> Categories: AI Agents, AI Companies, AI for Science, Drug Discovery
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Edison Scientific." aiwiki.ai, 27 Oct 2026. https://aiwiki.ai/wiki/edison_scientific
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Edison Scientific** is a San Francisco company that builds [AI agents](https://aiwiki.ai/wiki/ai_agents) for scientific research and drug development. It was launched on November 5, 2025 as a for-profit spinout of [FutureHouse](https://aiwiki.ai/wiki/futurehouse), the nonprofit AI research lab founded in 2023 by Sam Rodriques and Andrew White, who co-founded Edison and serve as its chief executive and chief technology officer.[1][2][3] Its flagship product is Kosmos, which the company calls an "AI Scientist": given a research objective and data, it runs repeated cycles of literature search, data analysis and hypothesis generation and returns a report in which each statement is cited to code or to the literature.[4][5] Edison announced a $70 million seed round in December 2025, and Tech Funding News reported that the round valued the company at roughly $250 million.[6][2] In 2026 it signed partnerships with the drugmaker Incyte, the biotech incubator Population Health Partners and the publisher Springer Nature, and it began publishing its benchmarks and open-weights models through a research site called Edison Advances.[7][8][9][10]

## History

### Origins at FutureHouse

FutureHouse, a 501(c)(3) nonprofit in San Francisco, was founded in 2023 to build an AI Scientist for basic research, mainly in biology; it says its first two years were made possible by Eric and Wendy Schmidt.[11] Edison presents FutureHouse's systems as its own lineage: its website lists the ChemCrow chemistry agent (2022), the PaperQA literature agent (2023), the Aviary agent-training framework (2024) and the Robin multi-agent discovery system (2025).[12] Robin, which proposed the glaucoma drug ripasudil and the experimental compound KL001 as enhancers of retinal pigment epithelium phagocytosis, a candidate strategy for dry age-related macular degeneration, was published in Nature on May 19, 2026. The paper's competing-interests statement says that nine of its authors, Rodriques and White among them, hold shares in Edison Scientific.[13]

In the spinout announcement, Rodriques and White wrote that commercial demand had outgrown the nonprofit. After FutureHouse opened its agent platform in May 2025, they said, it had been "inundated with requests for rate limit increases" that it could not support, and executives at the vice-president or C level from 6 of the 10 largest pharmaceutical companies had approached it about its agents.[1] They argued that "building a product, implementing payment systems, going to market, and supporting customers is not the right use of our philanthropic funding," and that for-profit capital could be raised faster for those purposes while the nonprofit's money went to basic research.[1]

### Spinout (November 2025)

On November 5, 2025, FutureHouse announced Edison Scientific as "a new commercial spinout that will focus on further developing and deploying our AI Scientist for commercial applications."[1] Part of the FutureHouse staff moved to Edison, and FutureHouse said it would keep developing AI Scientists for foundational research in biology.[1] Rodriques and White said they would run both organizations "for the time being, with strict controls in place to avoid self-dealing or private benefit and manage conflict of interest," and that Edison would keep a "generous free tier" while charging power users for higher rate limits and extra features.[1] On the same day, management of the FutureHouse platform passed to Edison, and Edison released Kosmos, which it described as a major upgrade on Robin.[4] Endpoints News reported the launch as an exclusive that morning.[14]

### Seed round (December 2025)

On December 18, 2025, Edison announced a $70 million seed round led by Triatomic Capital, Spark Capital and "a major US institutional biotech investor" that it did not name.[6] Pillar VC and Susa Ventures, which the company said had backed it at founding, also invested, as did Striker Venture Partners, Hawktail VC, Olive VC and a group of angel investors that Edison said included AI researchers, chief executives of frontier AI labs and leaders of major biotech and pharmaceutical companies.[6] Tech Funding News reported that the round valued Edison at roughly $250 million and named Google chief scientist [Jeff Dean](https://aiwiki.ai/wiki/jeff_dean) and CrowdStrike co-founder Dmitri Alperovitch among the angels; Edison's own announcement did not give a valuation.[2] Endpoints News also reported the round that day.[15] Edison then had roughly 30 employees and planned to hire across AI, biology, chemistry and engineering.[2] Alongside the round, the company said it would keep giving academics and students 650 platform credits a month, and it released a major update of PaperQA, its open-source literature agent.[6]

### Timeline

| Date | Event |
| --- | --- |
| November 5, 2025 | Edison Scientific launched as a FutureHouse spinout; Kosmos released and the FutureHouse platform transferred to Edison[1][4] |
| November 20, 2025 | Edison Analysis, a successor to FutureHouse's Finch data-analysis agent, released[16] |
| December 11, 2025 | Grant program announced offering up to five academic awards of 20,000 platform credits each (equal to 100 Kosmos runs)[17] |
| December 18, 2025 | $70 million seed round announced[6] |
| January 14, 2026 | Kosmos given the ability to find and download public datasets on its own[18] |
| February 5, 2026 | LABBench2 benchmark released[19] |
| February 18, 2026 | PaperQA3 released as the engine of the Edison Literature agent[20] |
| March 16, 2026 | Collaboration with NVIDIA described; BixBench-Hypothesis benchmark and BBH-Train dataset released[21] |
| May 19, 2026 | Incyte partnership announced; Kosmos rebuilt for biopharma R&D teams[7][22] |
| June 29, 2026 | Partnership with Population Health Partners announced[8] |
| July 31, 2026 | Edison Advances launched; MarkushGlyph and OCSRGlyph models released[10][23] |
| August 26, 2026 | BixBench3 benchmark released[24] |
| September 17, 2026 | Quoted in Anthropic's announcement of its Life Sciences Verification Program[25] |
| September 24, 2026 | Content agreement with Springer Nature announced[9] |

## Leadership

Rodriques is a physicist and bioengineer.[3] He completed a PhD at MIT in 2019 in the laboratory of Ed Boyden, working on the brain, and then ran a laboratory at the Francis Crick Institute in London before starting FutureHouse.[26] White is a computational chemist who was an associate professor of chemical engineering at the University of Rochester; in 2023 the university reported that he was on academic leave to serve as head of science at FutureHouse.[27] MIT News described him as having built "the first large language agent for science" before the two founded FutureHouse.[26] Edison's team page credits White with leading ChemCrow, the ether0 chemistry reasoning model and PaperQA.[3]

| Name | Role at Edison | Background |
| --- | --- | --- |
| Sam Rodriques | Co-founder and chief executive | Also co-founder, director and CEO of FutureHouse; MIT PhD (2019); formerly ran a lab at the Francis Crick Institute[3][26] |
| Andrew White | Co-founder and chief technology officer | Associate professor of chemical engineering at the University of Rochester (on academic leave as of 2023); led ChemCrow, PaperQA and ether0[3][27] |
| Michaela Hinks | Head of Science | Gene-regulation researcher (PhD 2024); supervised the work in the Kosmos technical report[3][4] |
| Tony Kulesa | Head of go-to-market | Partner at Pillar VC and a founding board member of FutureHouse[3] |
| Michael Skarlinski | Head of Platform | Previously led FutureHouse's platform team[3] |
| Ludovico Mitchener | Head of Applied AI | Led the design and engineering of Kosmos[3][4] |

Pillar VC, where Kulesa is a partner, is one of Edison's seed investors.[3][6]

## Kosmos

### Design and technical report

Kosmos was described in a technical report posted to arXiv on November 4, 2025, with Ludovico Mitchener as first author and 37 authors in all.[5] Given an open-ended objective and a dataset, it runs for up to 12 hours, launching parallel instances of a data-analysis agent and a literature-search agent in each cycle and then writing up its findings.[5] The central design choice is a structured [world model](https://aiwiki.ai/wiki/world_model) that the agents share, which the authors say lets Kosmos stay on one objective across about 200 agent rollouts. An average run executes about 42,000 lines of code across 166 data-analysis rollouts and reads 1,500 full-length papers across 36 literature-review rollouts, which the paper says is 9.8 times more code than Robin produced.[5]

To measure accuracy, the authors took 102 statements from three representative reports and asked expert scientists to check each one with their own analysis or literature search. Overall, 79.4% were judged accurate: 85.5% of statements based on data analysis, 82.1% of statements based on the literature, and 57.9% of statements that synthesized the two.[5] Collaborators also reported that one 20-cycle run did work equal, on average, to six months of their own research time.[5] Edison's launch post explained that this figure came from polling seven beta users, who averaged 6.14 months, and acknowledged that "human estimates of time saved are intrinsically suspect"; it offered two cross-checks, including that three findings Kosmos reproduced had originally taken people about four months.[4]

The report described seven discoveries in metabolomics, materials science, neuroscience and statistical genetics.[5] Three reproduced findings from unpublished or preprinted work that Kosmos had not accessed at runtime, such as the dominant effect of humidity during thermal annealing on perovskite solar cell efficiency. The other four were presented as new, including Mendelian randomization evidence that circulating superoxide dismutase 2 (SOD2) may reduce myocardial fibrosis in humans, and an observation that entorhinal cortex neurons in aging mice express less flippase, which the team checked against a separate human Alzheimer's disease dataset.[4] The company said it was validating that finding and others in the wet lab.[4]

Edison listed limits of its own. It said Kosmos "often goes down rabbit holes or chases statistically significant yet scientifically irrelevant findings," that longer runs make this more likely, and that it often ran Kosmos several times on the same objective.[4] At launch a run cost $200 (200 credits at $1 per credit), with some free use for academics, and Edison compared it to "a reagent kit, something you run on high-value targets as needed."[4]

### Rebuild for R&D teams (2026)

In January 2026 Edison gave Kosmos the ability to search for and download suitable public datasets, from Zenodo to the Protein Data Bank, so that users no longer had to prepare public data before a run.[18] On May 19, 2026, alongside the Incyte deal, it released a rearchitected Kosmos aimed at the whole biopharma pipeline.[7] The company said it had added data sources, skills and reasoning models for toxicology, clinical trials and regulatory documents, and had validated the system by reproducing key stages of clinical development for a novel checkpoint inhibitor, from lead identification through a Phase II proof-of-concept study.[7] The new version works through a persistent chat interface and through Slack, Teams and email, can provision GPUs for jobs such as [AlphaFold](https://aiwiki.ai/wiki/alphafold) structure prediction and molecular docking, connects to electronic lab notebooks and other company data, and runs on models that partners can train further with [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) and supervised fine-tuning on their own data.[7]

An engineering post published alongside the announcement, co-written by White, explained that Kosmos had been rebuilt as a code-writing agent, described as "an LLM with essentially one tool: execute code," instead of routing tasks to specialized agents with hand-built tools.[28] Each agent works in a sandbox, a container on a virtual machine managed with [Kubernetes](https://aiwiki.ai/wiki/kubernetes), hardened with gVisor, rootless containers and single-tenant virtual machines; credentials are injected outside the sandbox, packages are self-hosted, and the whole platform can be installed on a partner's own premises.[28] By October 2026 the company's website advertised on-premises deployment, SOC 2 Type II compliance, "forward-deployed scientists" and custom world models, and said Kosmos "remains model agnostic."[12] For Population Health Partners it offered specialized versions called Kosmos Clinical Protocol Design and Kosmos Regulatory Authoring.[29]

The May 2026 changes also reduced free access. Academics who had received 650 credits a month were given a one-time grant of 2,000 credits instead, founding Kosmos subscriptions were withdrawn, and the core research agents stayed publicly available on a site called the Edison Playground, with 10 free credits a month for all users.[7]

## Other agents

Kosmos is built on agents that Edison also offers individually through its platform and API, several of them renamed FutureHouse agents.[30][7][16]

| Agent | What it does | Origin |
| --- | --- | --- |
| Literature | Searches and synthesizes papers, patents and clinical trials, including their figures and tables, and gives short cited answers[20] | Formerly Crow; runs on PaperQA3 since February 2026[7][20] |
| Analysis | Analyzes data by editing Jupyter notebooks in Python, R and Bash; the analytical engine inside Kosmos[16] | Successor to FutureHouse's Finch, formerly closed beta[16] |
| Precedent | Checks whether anyone has already run a given experiment or explored a given hypothesis[26] | Formerly HasAnyone, or Owl[7] |
| Molecules | Chemistry agent with more than 30 tools for property prediction, retrosynthesis, toxicity and database search; can generate molecules with ether0[31] | Builds on ChemCrow; first offered in May 2025[31] |

Edison reports benchmark results for these agents, several of them on benchmarks that it or FutureHouse built. It said Edison Analysis reached 46% overall accuracy on BixBench, a FutureHouse bioinformatics benchmark, ahead of the comparison agents Biomni and [Claude Code](https://aiwiki.ai/wiki/claude_code), and 70% on Biomni's Eval1 benchmark.[16] It said PaperQA3 let its literature agent read figures and tables from more than 150 million papers and patents and made it one of the strongest [deep research](https://aiwiki.ai/wiki/deep_research) agents on parts of LABBench2 and on two variants of [Humanity's Last Exam](https://aiwiki.ai/wiki/humanity_s_last_exam).[20] The literature agent parses PDFs with [NVIDIA](https://aiwiki.ai/wiki/nvidia)'s Nemotron Parse model, which Edison said improved figure understanding by 15%, table understanding by 7% and text understanding by 3% compared with the best rule-based parsers.[21] An NVIDIA case study said Edison runs the parser on H200 GPUs through Modal Labs and serves more than 50,000 users.[32]

## Research and open releases

Edison launched Edison Advances on July 31, 2026 as a home for its research posts, engineering write-ups, [open-weights](https://aiwiki.ai/wiki/open_weights) models and benchmarks, and said its research culture was "inspired by the pioneering research at FutureHouse."[10] Its releases continue FutureHouse's habit of publishing evaluation sets such as [LAB-Bench](https://aiwiki.ai/wiki/labbench).

| Release | Date | Description |
| --- | --- | --- |
| PaperQA update | December 2025 | Major release of the open-source literature agent, alongside the seed announcement[6] |
| LABBench2 | February 2026 | Successor to LAB-Bench with 1,892 tasks in 11 categories, adding patent, clinical-trial and source-quality tasks[19] |
| PaperQA3 | February 2026 | Multimodal update to the algorithm behind Edison Literature, offered through the API as literature-20260216[20] |
| BixBench-Hypothesis and BBH-Train | March 2026 | A 51-task benchmark of hypothesis-driven bioinformatics analysis, plus 250 training tasks built with NVIDIA and released on [Hugging Face](https://aiwiki.ai/wiki/hugging_face)[21] |
| MarkushGlyph and OCSRGlyph | July 2026 | Open models for reading chemical structures in patents (2 billion and 94 million parameters)[23][33] |
| BixBench3 | August 2026 | 20 tasks that ask agents to reproduce the analyses of published computational biology studies from raw data[24] |

For BixBench3, Edison evaluated 13 frontier language models. Their average scores ranged from 0.0 to 0.48, and an average attempt took 6.8 hours, processed 102 million tokens and cost $43.[24]

The Glyph models address a narrow task in drug discovery: reading Markush structures, the patent drawings that describe a whole family of related compounds through a shared scaffold, variable sites and accompanying text.[23] MarkushGlyph is a [vision language model](https://aiwiki.ai/wiki/vision_language_model) built on Alibaba's [Qwen3.5](https://aiwiki.ai/wiki/qwen3_5)-2B-Base that reads the diagram and its text from a single image and outputs an extended SMILES notation. OCSRGlyph is a smaller transformer encoder-decoder for ordinary single-molecule images, which Edison reported reached 93.8% accuracy on a benchmark of molecules extracted from US patents when stereochemistry had to match and 96.2% when it did not.[23] The paper, by Alex Andonian, Rodriques, White and Siddharth Narayanan, was posted to arXiv on July 30, 2026, and the model weights and code were released under the [Apache License 2.0](https://aiwiki.ai/wiki/apache_license) together with the training datasets.[34][23][35] Edison said staff developed the models with Kosmos in an "autoresearch" style and that Kosmos independently proposed several key ideas, including oversampling chiral molecules in training and a stricter evaluation metric for Markush structures.[23]

## Partnerships and customers

**NVIDIA.** In March 2026 Edison described a collaboration with NVIDIA in which it trains agents with the [NVIDIA NeMo](https://aiwiki.ai/wiki/nvidia_nemo) Gym and NeMo RL tools and NVIDIA's [Nemotron](https://aiwiki.ai/wiki/nemotron) open models, and connects its Aviary framework to NeMo Gym. White spoke about AI scientists at NVIDIA's GTC 2026 conference.[21]

**Incyte.** On May 19, 2026, Edison and Incyte, a Nasdaq-listed drugmaker, announced that Kosmos would be embedded across Incyte's discovery and development work, starting with target discovery, target validation and translational biology.[7] Fierce Biotech reported that Incyte's global head of R&D, Pablo Cagnoni, saw a chance for the company's data to help train Kosmos, and quoted Rodriques saying that "most AI efforts in pharma treat data as something to analyze" while Edison's system "treats data as something to learn from continuously."[22] Edison said Kosmos was by then deployed at several biopharmaceutical companies.[28]

**Population Health Partners.** On June 29, 2026, Edison announced a partnership with Population Health Partners (PHP), the incubator whose leaders built Metsera and The Medicines Company, to co-build biotech companies in cardiovascular, inflammatory lung and gastrointestinal disease, starting from a clinical-stage asset in PHP's portfolio.[8] In a case study published in September 2026, Edison said a version of Kosmos trained for writing Investigational New Drug (IND) application documents scored 81.5% on a benchmark the two firms built, against 65.5% for a frontier general-purpose model, and that its regulatory-authoring tool found 98% of errors deliberately planted in a draft document.[29] These figures come from Edison.

**Springer Nature.** On September 24, 2026, Edison announced a subscription agreement that brings licensed articles from Springer Nature's Nature portfolio journals into Kosmos, with links to the version of record so that corrections and retractions show up. It called this the first of several content partnerships.[9]

**Anthropic.** Edison was one of three companies quoted when [Anthropic](https://aiwiki.ai/wiki/anthropic) launched its [Life Sciences Verification Program](https://aiwiki.ai/wiki/anthropic_life_sciences_verification_program) on September 17, 2026; the program gives verified life sciences teams access to Anthropic's models with safeguards that are more permissive for biology-related work. The other two were [Xaira Therapeutics](https://aiwiki.ai/wiki/xaira_therapeutics) and Manifold Bio.[25]

## Funding

| Date | Round | Amount | Valuation | Investors |
| --- | --- | --- | --- | --- |
| December 18, 2025 | Seed | $70 million | About $250 million, as reported by Tech Funding News; not stated by the company[2][36] | Led by Triatomic Capital, Spark Capital and an unnamed US institutional biotech investor; with Pillar VC, Susa Ventures, Striker Venture Partners, Hawktail VC, Olive VC and angels including Jeff Dean and Dmitri Alperovitch[6][2] |

As of October 2026, Edison's news page listed no later funding announcement.[37]

## Relationship with FutureHouse

The two organizations share founders, history and office space. FutureHouse describes them as "two independent organizations" with "separate missions, operations, and governance," and says it has access to Edison's technology while remaining focused on research for the public good.[11] Edison's team page puts the division this way: "FutureHouse continues advancing basic biology and educational access, while Edison scales the AI scientist to researchers and industries worldwide."[3] FutureHouse says the AI systems first developed under its name "are now part of Edison Scientific's platform," and its site still lists Kosmos among its own results while saying that Edison "is bringing this technology to the world."[11] Edison's Advances site likewise lists FutureHouse's ether0 model and the Robin paper.[33][38]

## Reception

Coverage of Kosmos combined interest with caution. In a November 2025 report for Alzforum, Jessica Shugart quoted Mathieu Bourdenx of University College London, a co-author of the Kosmos paper, saying he now ran most of his datasets through Kosmos.[39] Betty Tijms of Amsterdam University Medical Center commented that Kosmos made her "wonder whether we are now heading to a future where we don't read the papers we cite, and also don't write the papers of our own studies," and called the system "a black box" because the manuscript gave few technical details.[39] Georg Meisl of the University of Cambridge wrote that "by the metrics in this paper, one in five conclusions is still wrong," doubted that the Alzheimer's example would take a person six months, and said he would treat Kosmos's output "more like getting an opinion from a well-read colleague than a thorough analysis."[39]

Tech Funding News set Edison apart from other [AI for science](https://aiwiki.ai/wiki/ai_for_science) companies such as [Periodic Labs](https://aiwiki.ai/wiki/periodic_labs), [Lila Sciences](https://aiwiki.ai/wiki/lila_sciences) and [Chai Discovery](https://aiwiki.ai/wiki/chai_discovery), writing that Edison "helps scientists decide what to do next rather than owning the entire stack."[2] Debarghya Das's "The Ultimate Neolab List" (updated October 2026) includes Edison Scientific and rates it a "clean" neolab, citing a scientist-founded team, early revenue, a focus on models and agents, and a large seed round.[40]

## See also

- [FutureHouse](https://aiwiki.ai/wiki/futurehouse)
- [AI for Science](https://aiwiki.ai/wiki/ai_for_science)
- [AI in drug discovery](https://aiwiki.ai/wiki/ai_in_drug_discovery)
- [AI co-scientist (Google)](https://aiwiki.ai/wiki/ai_co_scientist)
- [LAB-Bench](https://aiwiki.ai/wiki/labbench)
- [Periodic Labs](https://aiwiki.ai/wiki/periodic_labs)
- [Lila Sciences](https://aiwiki.ai/wiki/lila_sciences)
- [List of neolabs](https://aiwiki.ai/wiki/list_of_neolabs)
- [Neolab](https://aiwiki.ai/wiki/neolab)

## References

1. Sam Rodriques and Andrew White, ["Announcing Edison Scientific"](https://www.futurehouse.org/research-announcements/announcing-edison-scientific), FutureHouse, 5 November 2025.
2. Sofia Chesnokova, ["FutureHouse spinout Edison lands $70M to build autonomous AI scientists for biology and chemistry"](https://techfundingnews.com/edison-raises-70m-ai-scientist-platform/), Tech Funding News, 19 December 2025.
3. Edison Scientific, ["Meet the Edison Scientific Team"](https://edisonscientific.com/team), edisonscientific.com. Accessed 27 October 2026.
4. Sam Rodriques and Michaela Hinks, ["Kosmos: An AI Scientist for Autonomous Discovery"](https://edisonscientific.com/news/announcing-kosmos), Edison Scientific, 5 November 2025.
5. Ludovico Mitchener, Angela Yiu, Benjamin Chang, et al., ["Kosmos: An AI Scientist for Autonomous Discovery"](https://arxiv.org/abs/2511.02824), arXiv:2511.02824, submitted 4 November 2025 (v2, 5 November 2025).
6. Sam Rodriques and Andrew White, ["We Raised $70M to Accelerate Science"](https://edisonscientific.com/news/we-raised-70m-to-accelerate-science), Edison Scientific, 18 December 2025.
7. Sam Rodriques, ["Building AI-Native Biopharma: Announcing Our Partnership with Incyte Corporation and Kosmos for R&D Teams"](https://edisonscientific.com/news/building-ai-native-biopharma-announcing-our-partnership-with-incyte-corporation-and-kosmos-for-r-d-teams), Edison Scientific, 19 May 2026.
8. Sam Rodriques, ["Creating Novel Biotechs Focused on Population-Scale Diseases: Announcing our Partnership with Population Health Partners"](https://edisonscientific.com/news/creating-novel-biotechs-focused-on-population-scale-diseases-announcing-our-partnership-with-population-health-partners), Edison Scientific, 29 June 2026.
9. Andrew White and Kiki Szostkiewicz, ["Bringing Trusted Research Into Kosmos with Springer Nature"](https://edisonscientific.com/news/bringing-trusted-research-into-kosmos-our-agreement-with-springer-nature), Edison Scientific, 24 September 2026.
10. Andrew White, Michael Skarlinski and Michaela Hinks, ["Announcing Edison Advances"](https://advances.edisonscientific.com/research/announcing-edison-advances), Edison Advances, 31 July 2026.
11. FutureHouse, ["About"](https://www.futurehouse.org/about), futurehouse.org. Accessed 27 October 2026.
12. Edison Scientific, ["Edison Scientific: The AI Platform for scientific R&D"](https://edisonscientific.com/) (company homepage), edisonscientific.com. Accessed 27 October 2026.
13. Ali E. Ghareeb, Benjamin Chang, Ludovico Mitchener, et al., ["A multi-agent system for automating scientific discovery"](https://www.nature.com/articles/s41586-026-10652-y), Nature, published 19 May 2026.
14. Andrew Dunn, ["Exclusive: Research institute FutureHouse debuts AI scientist and for-profit spinoff"](https://endpoints.news/research-institute-futurehouse-debuts-ai-scientist-kosmos-and-for-profit-spinoff/), Endpoints News, 5 November 2025.
15. Andrew Dunn, ["FutureHouse spinoff Edison raises $70M seed for its AI-doing-science platform"](https://endpoints.news/ai-startup-edison-raises-70m-seed-for-research-software/), Endpoints News, 18 December 2025.
16. Ludovico Mitchener, Jon Laurent, Angela Yiu, Arvis Sulovari, Conor Igoe and Alex Andonian, ["Introducing Edison Analysis"](https://advances.edisonscientific.com/research/introducing-edison-analysis), Edison Advances, 20 November 2025.
17. Edison Scientific, ["Apply to the Edison Platform ACU Grant"](https://edisonscientific.com/news/edison-platform-acu-grant), Edison Scientific, 11 December 2025.
18. Angela Yiu and Ludovico Mitchener, ["Expanding Kosmos' Data Access Capabilities"](https://edisonscientific.com/news/expanding-kosmos-data-access-capabilities), Edison Scientific, 14 January 2026.
19. Jon Laurent, ["LABBench2: An Improved Benchmark for Measuring AI in Biology Research"](https://advances.edisonscientific.com/research/labbench2-an-improved-benchmark), Edison Advances, 5 February 2026.
20. Andrew D. White, James D. Braza, Michael Pieler, Michael Skarlinski and Siddharth Narayanan, ["Introducing PaperQA3: a frontier multimodal deep research agent for science"](https://advances.edisonscientific.com/research/edison-literature-agent), Edison Advances, 18 February 2026.
21. Jon Laurent, Siddharth Narayanan, Seenia Hong and Richard Magness, ["Accelerating Science at Scale"](https://edisonscientific.com/news/accelerating-science-at-scale), Edison Scientific, 16 March 2026.
22. Will Maddox, ["Incyte inks deal with Edison Scientific to train AI through drug discovery"](https://www.fiercebiotech.com/biotech/incyte-inks-deal-edison-scientific-train-ai-through-drug-discovery), Fierce Biotech, 19 May 2026.
23. Alex Andonian and Siddharth Narayanan, ["Translating Chemical Structures"](https://advances.edisonscientific.com/research/translating-chemical-structures/), Edison Advances, 31 July 2026.
24. Zane Koch, ["BixBench3: Benchmarking AI agents on research-study-scale computational biology tasks"](https://advances.edisonscientific.com/research/bixbench3-benchmarking-study-scale-computational-biology), Edison Advances, 26 August 2026.
25. Anthropic, ["Introducing the Life Sciences Verification Program"](https://www.anthropic.com/news/life-sciences-verification-program), Anthropic, 17 September 2026.
26. Zach Winn, ["Accelerating scientific discovery with AI"](https://news.mit.edu/2025/futurehouse-accelerates-scientific-discovery-with-ai-0630), MIT News, 30 June 2025.
27. University of Rochester Department of Chemical Engineering, ["2023 News Archive"](https://www.hajim.rochester.edu/che/news-events/news/2023/index.html), Hajim School of Engineering and Applied Sciences, 2023.
28. Andrew D. White, Michael Skarlinski, Ludovico Mitchener, Siddharth M. Narayanan and Tyler Nadolski, ["How We Built Kosmos"](https://advances.edisonscientific.com/research/how-we-built-kosmos), Edison Advances, 18 May 2026.
29. Sam Rodriques, Mimoun Cadosch Delmar, Johnny Li and Vladimir Ermakov, ["How Population Health Partners accelerates its regulatory process with Kosmos"](https://edisonscientific.com/news/how-population-health-partners-accelerates-its-regulatory-process-with-kosmos), Edison Scientific, 30 September 2026.
30. Jon Laurent, Siddharth Narayanan, Ludovico Mitchener, Mayk Caldas, Sam Cox and Andrew White, ["Performance of our Platform Agents"](https://edisonscientific.com/news/performance-of-our-platform-agents), Edison Scientific, 20 November 2025.
31. Mayk Caldas Ramos and Sam Cox, ["Molecules: An AI Agent for Chemistry"](https://edisonscientific.com/news/molecules-an-ai-agent-for-chemistry), Edison Scientific, 4 December 2025.
32. NVIDIA, ["Scaling Multimodal Scientific Literature Understanding with NVIDIA Nemotron Parse"](https://developer.nvidia.com/case-studies/scientific-literature-ai-nvidia-nemotron), NVIDIA Developer case study, 2026.
33. Edison Scientific, ["Models"](https://advances.edisonscientific.com/models), Edison Advances. Accessed 27 October 2026.
34. Alex Andonian, Samuel G. Rodriques, Andrew D. White and Siddharth M. Narayanan, ["MarkushGlyph and OCSRGlyph: Improved Chemical Structure Recognition"](https://arxiv.org/abs/2607.28532), arXiv:2607.28532, submitted 30 July 2026.
35. Edison Scientific, ["EdisonScientific/MarkushGlyph"](https://huggingface.co/EdisonScientific/MarkushGlyph), Hugging Face model card (license apache-2.0; base model Qwen/Qwen3.5-2B-Base). Accessed 27 October 2026.
36. Caproasia, ["United States FutureHouse Spinout AI Commercial Startup Edison Scientific Raised $70 Million in Seed Funding at $250 Million Valuation"](https://www.caproasia.com/2025/12/19/united-states-futurehouse-spinout-ai-commercial-startup-edison-scientific-raised-70-million-in-seed-funding-at-250-million-valuation-futurehouse-co-founded-in-2023-as-non-profit-ai-research-lab-by/), Caproasia, 19 December 2025.
37. Edison Scientific, ["News"](https://edisonscientific.com/blog) (news and announcements index), edisonscientific.com. Accessed 27 October 2026.
38. Ali Ghareeb, ["Robin: A multi-agent system for automating scientific discovery"](https://advances.edisonscientific.com/research/robin-multi-agent-system), Edison Advances, 19 May 2026.
39. Jessica Shugart, ["Introducing Kosmos: 'AI Scientist' That Makes Discoveries Overnight"](https://www.alzforum.org/news/research-news/introducing-kosmos-ai-scientist-makes-discoveries-overnight), Alzforum, 21 November 2025 (with reader comments by Betty Tijms and Georg Meisl).
40. Debarghya Das, ["The Ultimate Neolab List"](https://debarghyadas.com/neolabs/), debarghyadas.com, updated 24 October 2026.