Adaption Labs, which markets itself as Adaption (legal name Adaption, Inc.), is an artificial intelligence company based in San Francisco that develops AI systems intended to keep learning and adapting after deployment, as an alternative to building ever-larger models and retraining them at high cost.[1][4] It was founded in 2025 by Sara Hooker, formerly vice president of research at Cohere and head of Cohere's research lab, and Sudip Roy, formerly Cohere's senior director of inference. Hooker is chief executive and Roy is chief technology officer.[2][4][7][20] The company announced itself in October 2025. On February 4, 2026, it disclosed a $50 million seed round led by Emergence Capital Partners, and it declined to disclose a valuation.[2][3][4][5]

Adaption organizes its work around three "pillars": Adaptive Data, Adaptive Intelligence and Adaptive Interfaces.[4][8] Its first product, Adaptive Data, is a platform for improving, localizing and generating training datasets and opened in early access in February 2026.[8] AutoScientist followed in May 2026. It automates fine-tuning by optimizing a model's training data and training recipe together.[19][20] In October 2026 the company released Visualize, a data-exploration tool and the first product under Adaptive Interfaces.[35] Fortune and TechCrunch have both described Adaption as one of the "neolabs", a label for a new generation of AI labs that Fortune glosses as "frontier AI labs" and TechCrunch as "research-driven AI labs".[4][20]

Overview

Field	Detail
Legal name	Adaption, Inc.[1]
Other name	Adaption Labs (used at launch, in press coverage and in its web domain, adaptionlabs.ai)[2][4]
Founded	2025[7]; publicly announced October 2025[2][3]
Founders	Sara Hooker (CEO) and Sudip Roy (CTO)[4][20]
Headquarters	San Francisco, California, United States[4]
Focus	Continual learning, efficient adaptation of AI systems, training-data optimization, automated model training[4][7]
Products	Adaptive Data (February 2026), AutoScientist (May 2026), Visualize (October 2026)[8][19][35]
Funding	$50 million seed round led by Emergence Capital Partners, announced February 4, 2026[4][5][6]
Valuation	Not disclosed[4]
Scientific lead	Hugo Larochelle, in an advisory role, from July 2026[26]

History

Founding

Hooker joined Cohere in 2022 to lead its research lab, Cohere For AI (known by 2025 as Cohere Labs), after about five years as a research scientist at Google, where she worked in Google Brain.[2][3][5][39] She announced in the summer of 2025 that she was leaving Cohere. TechCrunch reported that she left in August 2025, while BetaKit reported that she stayed through September to oversee applications to Cohere's research scholars program.[2][3] Roy had joined Cohere in 2021 after seven years as a research scientist at Google, and was Cohere's director of engineering before becoming senior director of inference.[2]

The company became public in October 2025 through posts by its founders. On October 7, 2025, Hooker wrote on X that she was "starting a new project" and was "working on what I consider to be the most important problem: building thinking machines that adapt and continuously learn". The post linked to adaptionlabs.ai and said the company was hiring for engineering, operations and design.[3][50] Roy wrote on LinkedIn that the next AI breakthrough "won't just be about larger models", and would instead come from "systems that can efficiently adapt without the computational overhead of traditional gradient-based methods."[2] BetaKit noted that some of the first job postings referred to "severe compute budgets", and that the California-based company was also hiring for hybrid roles in Canada and the United Kingdom.[2] Hooker told TechCrunch the company would open a San Francisco office but planned to hire worldwide.[3]

Seed funding

In October 2025, TechCrunch reported that the startup had been in talks earlier that fall to raise a seed round of $20 million to $40 million, according to three investors who had reviewed its pitch decks. The investors said the round had since closed, but the final amount was unclear, and Hooker declined to comment.[3] Emergence Capital's portfolio page gives 2025 as the year the firm partnered with the company.[7]

On February 4, 2026, Adaption announced a $50 million seed round led by Emergence Capital Partners, with participation from Mozilla Ventures, Fifty Years, Threshold Ventures, Alpha Intelligence Capital, E14 Fund and Neo.[4][5][6] Fortune reported that the company "declined to provide any information about its valuation following the fundraise."[4] The law firm Wilson Sonsini Goodrich & Rosati advised Adaption on the transaction.[6] Hooker told Fortune that the money would pay for more AI researchers and engineers, and for designers working on interfaces beyond the standard "chat bar".[4] In Emergence's announcement, Yaz El-Baba, the investor the firm lists for the company, wrote that most AI founders Emergence had met in the previous year had pitched labs following "scale-at-all-costs approaches", and that Adaption was founded on the conviction that "scale alone is no longer enough."[5][7]

Date	Round	Amount	Lead investor	Other participants	Valuation
Fall 2025 (reported October 2025)	Seed, sought and reported closed	$20 million to $40 million sought; final amount unclear	Not reported	Not reported	Not reported[3]
February 4, 2026	Seed (announced)	$50 million	Emergence Capital Partners	Mozilla Ventures, Fifty Years, Threshold Ventures, Alpha Intelligence Capital, E14 Fund, Neo	Not disclosed[4][5][6]

Public reporting does not make clear whether the round described by TechCrunch in October 2025 was part of the $50 million announced in February 2026.[3][4]

Releases and announcements in 2026

Date (2026)	Announcement
February 24	Adaptive Data opens in early access[8]
March 10	Research Grant Program for academic researchers[10]
March 17	Blueprint, a specification layer for Adaptive Data[9]
March 31	Adaption for Startups program[11]
April 2	Uncharted Data Challenge, an open dataset competition[12]
April 9	Adaptive Data API and Python SDK[13]
April 14	Expand Your World, a multilingual dataset feature[14]
April 21	Adaption Ambassadors program[15]
April 23	Forge, for turning unstructured documents into datasets[16]
April 30	Together AI fine-tuning integrated into Adaptive Data[17][18]
May 13	AutoScientist launched, free for its first 30 days[19][20]
May 18	Public Sector Grant[21]
June 2	Adaptive Data extended to images[22]
June 5	AutoScientist Challenge with a $60,000 prize pool[23]
June 10	Tiny AutoScientist, for models under 10 billion parameters[24]
June 22	Partnership with NCS for the Asia-Pacific region[25]
July 2	Hugo Larochelle joins as Scientific Lead[26]
July 16	Multimodal AutoScientist (text and images)[27]
July 28	AutoScientist API and SDK[28]
August 4	Results of a pilot with AI Singapore[29]
September 3	Invent a Dataset[31]
October 15	Visualize, the first Adaptive Interfaces release[35]

Founders and team

Person	Role at Adaption	Background
Sara Hooker	Co-founder and CEO[20]	Research scientist at Google, including Google Brain; vice president of research at Cohere and head of Cohere For AI (Cohere Labs)[2][3][39]
Sudip Roy	Co-founder and CTO[4]	PhD, Cornell University (2014); research scientist at Google, where he worked on Pathways; director of engineering and senior director of inference at Cohere[2][5][42]
Hugo Larochelle	Scientific Lead (advisory), from July 2026[26]	Scientific Director of Mila, which he continues to lead[26]

Sara Hooker

Fortune wrote that since her time at Google, Hooker had built a reputation as an opponent of the "scale is all you need" view held by many AI researchers. Her 2020 essay "The Hardware Lottery" argued that ideas in AI often succeed or fail according to whether they happen to fit existing hardware, rather than on their inherent merit.[4][38] Fortune also cited a more recent essay, "On the Slow Death of Scaling", which argued that smaller models trained with better techniques can outperform much larger ones; Adaption's research page says the essay informs the company's work.[4][36]

At Cohere, the lab she led worked on model efficiency and training-data quality. One of its best-known projects was Aya, a multilingual set of models and datasets built with more than 3,000 researchers worldwide.[39][41] TIME named her to its 2024 TIME100 AI list in the "Shapers" category.[39] In 2025 she co-authored "The Leaderboard Illusion", a paper that alleged the rankings of the LM Arena chatbot leaderboard were unreliable. BetaKit reported that it drew largely applause but some backlash from AI researchers.[2][40]

Sudip Roy

Roy graduated from the Indian Institute of Technology Kharagpur with a bachelor's degree in computer science and engineering in 2009.[43] He completed a PhD in computer science at Cornell University in August 2014, advised by Johannes Gehrke and Christoph Koch, with research on transaction processing and distributed systems. His first job afterward was at Google.[42][43] As a doctoral student he co-authored "Entangled Queries: Enabling Declarative Data-Driven Coordination", which won the SIGMOD Best Paper Award in 2011.[46] At Google he worked on core infrastructure including Pathways and on systems that serve Gemini models in production, according to Emergence.[5] He is one of the authors of the paper describing Pathways, Google's orchestration layer for accelerators, which received an Outstanding Paper Award at MLSys 2022.[44][45] At Cohere he led teams responsible for fine-tuning and serving the company's large language models.[47] Hooker told Fortune: "My cofounder makes GPUs go extremely fast, which is important for us because of the real-time component."[4]

Team

Authors of Adaption's research posts include Shivalika Singh, Wei-Yin Ko and Gbemileke Onilude, all three of whom were co-authors with Hooker of the 2024 Aya model paper.[33][34][41] The company says its team has collectively authored more than 150 publications.[36] In October 2026 its careers page listed seven open roles. Research positions were remote, open in several countries in the Americas, Europe and India; platform engineering roles were hybrid in San Francisco.[37]

Hugo Larochelle joined on July 2, 2026 as Scientific Lead. Adaption said he would serve in an advisory capacity while continuing as Scientific Director of Mila. Larochelle said that "how AI systems learn is one of the most consequential questions facing the AI field."[26]

Research approach

Hooker told Fortune that she wants to build models that learn continuously, without expensive retraining or fine-tuning and without the extensive prompt and context engineering that most enterprises use to adapt models to their needs. She said her goal was "to eliminate prompt engineering", and dismissed the practice of writing highly specific prompts as "prompt acrobatics".[4] Fortune noted that creating models that learn continuously, the problem studied in continual learning, "is considered one of the big outstanding challenges in AI."[4] In October 2025, Hooker told TechCrunch that "scaling-pilled approaches, which are attractive but extremely boring" had not produced intelligence able to interact with the world. She declined to say what methods the company used or whether it relied on large language models.[3]

In February 2026, Hooker described the three pillars to Fortune as follows.[4]

Pillar	Description given to Fortune
Adaptive data	AI systems generate and manipulate the data they need to answer a problem on the fly, rather than being trained on a large static dataset
Adaptive intelligence	Automatically adjusting how much compute to spend based on the difficulty of a problem
Adaptive interfaces	Learning from how users interact with the system

Fortune reported that one idea under investigation was "gradient-free learning", which changes a model's behavior at inference time while leaving its core weights untouched, in contrast to training with gradient descent.[4] Hooker gave two examples. "On-the-fly merging" selects from a repertoire of adapters, often small models trained separately on small datasets, which then shape the main model's response; the system picks the adapter according to the question asked. "Dynamic decoding" adjusts the probabilities a model uses to pick its output according to the task, without changing its weights (see decoding strategies).[4] Hooker argued this would change the economics of AI because "with inference compute, you get way more bang" than with pretraining compute.[4] The company's research page lists its algorithmic research areas as adaptive compute, on-the-fly alignment, dynamic learning strategies and generative interfaces. It treats latency, memory, cost and energy limits as design inputs.[36]

The products the company had released by October 2026 work mainly on training data and on automated fine-tuning, which changes model weights. TechCrunch described AutoScientist as "an automated approach to conventional fine-tuning".[20] In its AutoScientist announcement, Adaption said it was working on "learning techniques that enable real-time adaptation without requiring training at all."[19]

Products

Adaptive Data

Adaptive Data treats a dataset as something to be continually reshaped rather than a fixed input. It rewrites, expands and rebalances existing data toward a target behavior, and fills gaps in coverage.[8][17] The company says early deployments saw an average 82% increase in data quality and that the platform supports 242 languages; these are Adaption's own figures.[8][18]

Feature	Released	Description
Adaptive Data	February 24, 2026	Early-access platform for optimizing and adapting datasets[8]
Blueprint	March 17, 2026	Users define goals such as length, tone, safety thresholds and custom content policies, which become objectives applied to every dataset the platform produces[9]
API and Python SDK	April 9, 2026	Programmatic access to upload, preprocessing, adaptation, evaluation and export. Accepts JSONL, CSV and Parquet files or imports from Hugging Face and Kaggle. Recipes include deduplication, prompt rephrasing and reasoning-trace generation[13]
Expand Your World	April 14, 2026	Starting from as few as 10 examples in one language, generates up to 2,420 examples across 242 languages and localizations[14]
Forge	April 23, 2026	Extracts and structures data from documents such as PDFs, scanned documents, spreadsheets, slide decks and email threads[16]
Together AI fine-tuning	April 30, 2026	Users can launch fine-tuning jobs on Together AI infrastructure from inside Adaptive Data[17][18]
Vision support	June 2, 2026	Visual question answering, image captioning, visual reasoning, image classification and document question answering. Adaption reported an average win rate of 67% for adapted data across five vision datasets[22]
Invent a Dataset	September 3, 2026	Generates a training-ready dataset from a written description of the behavior a model should learn, with no seed data[31]

AutoScientist

AutoScientist launched on May 13, 2026. It runs the research loop behind fine-tuning: it searches training recipes, optimizes the data, trains and evaluates, and repeats until the model reaches the user's objective.[19][28] Adaption said the system addresses common reasons fine-tuning fails, including catastrophic forgetting, overfitting on small datasets and conflicting training signals.[19] Hooker told TechCrunch that it "co-optimizes both the data and the model" and suggested "we can finally allow for successful frontier AI trainings outside of these labs."[20]

In its launch post, Adaption reported that AutoScientist outperformed training configured by its own research staff by an average of 35%, and that aggregate win rates rose from 48% to 64%. The tests covered eight verticals, dataset sizes from 5,000 to 100,000 examples, and models offered for fine-tuning by Together AI, with win rates computed on in-house evaluations for each vertical.[19] TechCrunch reported that the launch materials claimed AutoScientist had "more than doubled win rates across different models", and called these "impressive numbers, but difficult to put into context". Because the system adapts models to specific tasks, it noted, conventional benchmarks such as SWE-Bench or ARC-AGI do not apply.[20] AutoScientist was free to use for its first 30 days.[19][20]

The company extended AutoScientist several times:

Tiny AutoScientist (June 10, 2026) applies the same loop to models in the 0.8 billion to 8 billion parameter range, aimed at edge deployment, on-device inference and regulated industries.[24] (See small language model.)
Multimodal AutoScientist (July 16, 2026) trains models on text and images together.[27]
AutoScientist API and SDK (July 28, 2026) lets developers start runs in code. A user gives a target win rate, and the run stops when it reaches that target or exhausts its iteration budget. Users can let the system choose a base model or pick one from a registry that the company says ranges from sub-billion-parameter to frontier-scale models. Trained weights can be downloaded.[28]

Adaption's enterprise page lists Qwen, Gemma, GPT, Nemotron, Mistral and Llama among the base model families AutoScientist can adapt.[48] By October 2026 the company said AutoScientist supported 44 expert domains.[34]

Visualize

Visualize, released on October 15, 2026, is the first product under Adaptive Interfaces. Users ask questions about a text or tabular dataset in plain language and receive interactive charts. They can click through to the rows behind a pattern, select part of a chart for follow-up questions, and edit charts by describing the change. The company presents it as a way to decide what data to change before reshaping it with Adaptive Data.[35]

Partnerships

Partner	Announced	Details
Together AI	April 30, 2026	Together Fine-Tuning became available natively inside Adaptive Data. Together said its service supports LoRA and full fine-tuning, including open models over 100 billion parameters such as Kimi K2.5, GLM 5.1 and Qwen 3.5-397B.[17][18]
NCS	June 22, 2026	Adaption announced a partnership with NCS, a Singtel Group subsidiary, to offer adaptive AI in Singapore, Southeast Asia, Hong Kong and Australia. Adaption supplies AutoScientist and Adaptive Data; NCS handles customer engagement, advisory, integration and managed services.[25]
AI Singapore	August 4, 2026	Adaption reported results of a pilot on training data for AI Singapore's SEA-LION language models[29]

In the AI Singapore pilot, Adaptive Data processed a sample of 1,050,000 rows from AI Singapore's SEA-Instruct-2602 dataset in seven Southeast Asian languages. Adaption said that within one month it enhanced 1 million samples and generated about 750,000 new samples by localizing existing ones into Singapore Tamil, Singapore Malay, Burmese, Khmer and Lao, which are low-resource languages for AI. It also said AutoScientist saved AI Singapore 56 engineering hours of manual hyperparameter tuning.[29] Adaption presents such projects as examples of sovereign AI, meaning models built on an organization's own data and kept under its control.[25][29]

Community programs

Program	Announced	Details
Research Grant Program	March 10, 2026	Free platform access for academic researchers worldwide, reviewed on a rolling basis, with a planned annual symposium for grant recipients[10]
Adaption for Startups	March 31, 2026	Access to the Adaptive Data Plus plan and free credits for startups[11]
Uncharted Data Challenge	April 2, 2026	Challenge (April 10 to May 1, 2026) to build and open-source datasets for underserved domains and under-resourced languages, with a $20,000 prize pool[12]
Adaption Ambassadors	April 21, 2026	Six-month program giving community organizers credits, early access and funding for local events[15]
Public Sector Grant	May 18, 2026	Platform access for public sector agencies, academic partners and civic organizations[21]
AutoScientist Challenge	June 5, 2026	Hackathon in two parts (June 8 to August 10, 2026) with a $60,000 prize pool. Entrants had to publish their adapted datasets and trained weights on Hugging Face and Kaggle[23]

The company also maintains a public leaderboard that ranks the best AutoScientist runs submitted by users in each domain.[49]

Published research

Adaption publishes research results on its blog. The figures below are the company's own measurements.

Date	Title	Reported findings
August 13, 2026	Better Agent Memory Starts Before Retrieval	An agent memory system that stores both "narrative" and "atomic" memories. Adaption reported 90.6% on LongMemEval, against 60.6% for full-history prompting with the same model and 71.6% for its own reproduction of Mem0 OSS[30]
September 9, 2026	A Better Harness Can Unlock Smaller Models	On 140 held-out tasks from the Harvey Legal Agent Benchmark, a redesigned agent harness raised Qwen3.6-27B's criterion pass rate from 67.10% to 85.92%, and post-training raised it to 88.03%[32]
October 1, 2026	Invent a Dataset: Measuring Dataset Generation Abilities with Zero Seed Data	Compared Adaption's Invent API with APIs from Anthropic, Google, OpenAI, DeepSeek and Zai on dataset requests of 200 to 20,000 samples. Adaption reported datasets 19% to 55% more diverse than every model it tested, and quality about 17% higher than the strongest baselines, Claude Opus 5 and GLM-5.3[33]
October 8, 2026	Automated Data Quality Control With Agentic Checklists	An agent writes data-quality checklists from critiques of past AutoScientist runs. Models trained on 10,000 checklist-filtered examples had relative win rates 9.4% to 13.2% higher than models trained on 10,000 random examples[34]

The company's research page also lists Hooker's essay on the slow death of scaling, "The AI Observatory", a public dataset of real AI conversations drawn from seven sources, and a paper arguing for long, open-world tasks to evaluate AI agents.[36]

Reception

Fortune described Adaption as "a significant bet against the prevailing AI industry wisdom" that the best way to more capable AI is bigger models trained on more data.[4] It named other new labs pursuing approaches that could lead to continuously learning AI, including Core Automation, founded by former OpenAI researcher Jerry Tworek, and Ineffable Intelligence, founded by former Google DeepMind researcher David Silver.[4] In October 2025, TechCrunch called the company "the latest sign that the industry's faith in scaling LLMs is wavering", and cited recent skeptical comments on current approaches from researchers including Richard Sutton and Andrej Karpathy.[3]

TechCrunch wrote that AutoScientist's win-rate figures were difficult to put into context, because the system is evaluated on specific tasks rather than on standard benchmarks.[20] Adaption's other research results, including those on agent memory, agent harnesses and dataset generation, have been published as posts on its own website.[30][32][33]

See also

References

Adaption. "Adaption: Adaptive AI That Continuously Learns". Company website, accessed October 27, 2026.
Madison McLauchlan. "Ex-Cohere execs Sara Hooker and Sudip Roy unveil new AI startup". BetaKit, October 9, 2025.
Maxwell Zeff. "Why Cohere's ex-AI research lead is betting against the scaling race". TechCrunch, October 22, 2025.
Jeremy Kahn. "Adaption Labs secures $50 million seed round to build AI models that can change on the fly". Fortune, February 4, 2026.
Yaz El-Baba. "Adaptable AI for an Adapting World". Emergence Capital, February 4, 2026.
Wilson Sonsini Goodrich & Rosati. "Wilson Sonsini Advises Adaption Labs on $50 Million Seed Funding". 2026, accessed October 27, 2026.
Emergence Capital. "Adaption". Portfolio page, accessed October 27, 2026.
Sara Hooker and Sudip Roy. "Adaptive Data: Designed for Change". Adaption, February 24, 2026.
Sara Hooker and Sudip Roy. "Blueprint: A Specification Layer for Adaptive Data". Adaption, March 17, 2026.
Sara Hooker. "Adaption Research Grant Program: Where Curiosity Pushes Breakthroughs". Adaption, March 10, 2026.
Sudip Roy. "Introducing Adaption for Startups: AI That's Yours to Shape". Adaption, March 31, 2026.
Sara Hooker and Sudip Roy. "The Uncharted Data Challenge". Adaption, April 2, 2026.
Sudip Roy. "The Adaptive Data API and Python SDK". Adaption, April 9, 2026.
Sara Hooker and Sudip Roy. "Expand Your World". Adaption, April 14, 2026.
Sara Hooker and Sudip Roy. "The Adaption Ambassador Program". Adaption, April 21, 2026.
Sudip Roy. "Forge: Turn Unstructured Documents Into AI Ready Datasets". Adaption, April 23, 2026.
Sara Hooker, Sudip Roy and Max Ryabinin. "Build Models Shaped By Adaptive Data and Together AI". Adaption, April 30, 2026.
Max Ryabinin, Jennifer Wu, Will Van Eaton and Sonny Khan. "Announcing Together AI and Adaption Partnership". Together AI, April 30, 2026.
Adaption Research Staff. "AutoScientist: Automating the Science of Model Training". Adaption, May 13, 2026.
Russell Brandom. "Adaption aims big with AutoScientist, an AI tool that helps models train themselves". TechCrunch, May 13, 2026.
Sara Hooker and Sudip Roy. "Introducing the Public Sector Grant". Adaption, May 18, 2026.
Sara Hooker and Sudip Roy. "Beyond Text: Adaptive Data for the Multimodal Era". Adaption, June 2, 2026.
Sara Hooker and Sudip Roy. "AutoScientist Challenge: Bring Frontier AI to the World". Adaption, June 5, 2026.
Sara Hooker and Sudip Roy. "Tiny AutoScientist: Supersized Intelligence for Small Models". Adaption, June 10, 2026.
Sudip Roy. "Adaption and NCS Partner to Bring Sovereign AI to Enterprise Teams Across Asia-Pacific". Adaption, June 22, 2026.
Sara Hooker and Sudip Roy. "Hugo Larochelle Joins Adaption as Scientific Lead". Adaption, July 2, 2026.
Sudip Roy. "AutoScientist Goes Multimodal: Automate AI Training Across Text and Images". Adaption, July 16, 2026.
Sudip Roy. "AutoScientist API: Automated Model Training". Adaption, July 28, 2026.
Sudip Roy and Darius Liu. "Adaption and AI Singapore Pilot to Improve Model Training Efficiency in Southeast Asia". Adaption, August 4, 2026.
Bargav Jagatha and Sudip Roy. "Better Agent Memory Starts Before Retrieval". Adaption, August 13, 2026.
Sara Hooker and Sudip Roy. "Invent a Dataset: Data Generation for Custom Models". Adaption, September 3, 2026.
Dhruv Naik and Sudip Roy. "A Better Harness Can Unlock Smaller Models". Adaption, September 9, 2026.
Shivalika Singh, Andrija Djurisic, Sara Hooker, Gbemileke Onilude and Sudip Roy. "Invent a Dataset: Measuring Dataset Generation Abilities with Zero Seed Data". Adaption, October 1, 2026.
Wei-Yin Ko and Sara Hooker. "Automated Data Quality Control With Agentic Checklists". Adaption, October 8, 2026.
Simra Shahid and Sudip Roy. "Visualize: The Interface for Understanding Your Data". Adaption, October 15, 2026.
Adaption. "Research: Building Intelligence That Evolves". Accessed October 27, 2026.
Adaption. "Careers at Adaption". Accessed October 27, 2026.
Sara Hooker. "The Hardware Lottery". arXiv:2009.06489, September 14, 2020.
Carleton College. "Sara Hooker '13 named to TIME100 list of most influential people in AI". September 10, 2024.
Shivalika Singh et al. "The Leaderboard Illusion". arXiv:2504.20879, April 29, 2025.
Ahmet Üstün et al. "Aya Model: An Instruction Finetuned Open-Access Multilingual Language Model". arXiv:2402.07827, February 12, 2024.
Johannes Gehrke. "Johannes Gehrke". Cornell University homepage, list of former students, accessed October 27, 2026.
Sudip Roy. "Sudip Roy's Homepage". Cornell University, accessed October 27, 2026.
Paul Barham et al. "Pathways: Asynchronous Distributed Dataflow for ML". arXiv:2203.12533, March 23, 2022.
MLSys. "Pathways: Asynchronous Distributed Dataflow for ML". MLSys 2022, Outstanding Paper Award.
ACM SIGMOD. "SIGMOD Best Paper Award". Accessed October 27, 2026.
AI & Big Data Expo North America. "Sudip Roy". Speaker profile, accessed October 27, 2026.
Adaption. "Enterprise". Accessed October 27, 2026.
Adaption. "The AutoScientist Leaderboard". Accessed October 27, 2026.
Sara Hooker (@sarahookr). Post announcing the new company. X, October 7, 2025.

Editing: Adaption Labs