Mark Chen (OpenAI)
Last reviewed
Sources
29 citations
Review status
Source-backed
Revision
v3 · 2,678 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
29 citations
Review status
Source-backed
Revision
v3 · 2,678 words
Add missing citations, update stale details, or suggest a clearer explanation.
Mark Chen is an American artificial intelligence researcher and research executive at OpenAI, where he serves as Chief Research Officer [1][2]. He joined the company in 2018 and led several of its model efforts, including the team that built DALL-E, the team that created Codex, and the work that added visual perception to GPT-4 [3][4]. He was also among the leaders of OpenAI's reasoning models, the o-series that began with o1 in 2024 [5][6]. He became Chief Research Officer in March 2025 and leads research alongside Jakub Pachocki, who is Chief Scientist [1][2].
This article concerns Mark Chen the OpenAI research executive. It does not concern other people who share the name.
Chen earned a bachelor's degree in mathematics with computer science from the Massachusetts Institute of Technology, graduating in 2012 [3][7]. He spent the summer of 2011 as a visiting scholar at Harvard University before completing his degree [16][17]. After leaving MIT he worked in quantitative finance for several years. He held research and trading roles at proprietary trading firms, including Jane Street Capital, where he built machine learning models for trading in equities and futures [4][3]. Accounts of his work history place him in quantitative research between 2012 and 2018, across firms that included Tech Square Trading and Integral Technology, before his move to AI research [7][16].
Chen has a long association with competitive programming. He has served as a coach and team leader for the United States delegation to the International Olympiad in Informatics, a global programming competition for secondary school students [3][8]. Records from the competition list him as a deputy leader of the United States team in 2019 and as team leader in 2022 and 2024 [8]. He has competed under the handle chenmark, with results recorded in online programming contests, and his profile is indexed by the Competitive Programming Hall of Fame [18]. He has linked this work directly to his research, describing a long term aim of building models that can perform at the level of the strongest human competitors in mathematics and programming [11][19].
Chen joined OpenAI in 2018 [4][3]. He worked as a research scientist and went on to lead the organization's frontiers research, an area centered on multimodal modeling and reasoning [3][9]. Over several years he had a hand in a series of the company's model releases. His Google Scholar profile, which lists his affiliation as a research scientist at OpenAI, records more than 168,000 citations, an h-index of 39, and an i10-index of 50, figures that place his published work among the most cited in machine learning [20]. He is a listed author of several of OpenAI's foundational papers, including the 2020 paper "Language Models Are Few-Shot Learners," which introduced GPT-3, and the 2023 "GPT-4 Technical Report" [20].
Among Chen's earliest first author work at OpenAI is "Generative Pretraining from Pixels," presented at the International Conference on Machine Learning in 2020 and commonly known as Image GPT [21]. The paper trained a Transformer to predict pixels one at a time, applying the autoregressive approach used for language directly to images without building in any knowledge of their two dimensional structure [21]. Trained on unlabeled, low resolution ImageNet, the model learned strong visual representations, reaching competitive accuracy on benchmarks such as CIFAR-10 through linear probing and fine tuning [21]. The work supported a recurring theme in Chen's research, the idea that techniques for predicting the next token in text could carry over to other kinds of data, including pixels.
Chen led the team that created DALL-E, the text to image model that OpenAI introduced in 2021 [3][9]. The system generated images from natural language descriptions and became one of the early demonstrations that a single model could map between text and pictures. The underlying method was described in the 2021 paper "Zero-Shot Text-to-Image Generation," which Chen co-authored and which was presented at the International Conference on Machine Learning [20]. He was also a co-author of the 2022 paper "Hierarchical Text-Conditional Image Generation with CLIP Latents," the work behind DALL-E 2 [20]. Both projects fit a broader theme in his research, the carryover of language modeling techniques to images.
Chen led the development of Codex, a model that wrote computer code [3][4]. He is the first listed author of the 2021 paper that described the system, titled Evaluating Large Language Models Trained on Code [10]. The paper introduced Codex as a version of GPT-3 adapted for programming, and it released HumanEval, a set of programming problems used to measure how often a model could produce correct code [10]. A production version of Codex powered the GitHub Copilot coding assistant [4][10].
Chen worked on the visual capabilities of GPT-4, leading the effort that gave the model the ability to interpret images as well as text [3][9]. This work extended the frontiers research direction that combined different kinds of input within one model. He also contributed to GPT-3, the large language model that preceded GPT-4 [3].
Chen was among the leaders of OpenAI's work on reasoning, the line of models that the company released under the o-series name [5][6]. The first of these, o1, appeared in September 2024 and was built to spend more computation working through a problem before giving an answer, which improved its performance on tasks in mathematics, science, and coding [5]. The approach marked a shift away from improving models only by making them larger and toward letting them deliberate at the point of use. Chen and Pachocki have been described in reporting as central figures behind these models, including o1 and the later o3 [6]. Chen has said publicly that he does not regard reasoning as a solved problem, stating that the field has "definitely not solved it" [11]. His interest in the area connects to his longer history with competitive programming and mathematics, domains in which the correctness of a chain of steps can be checked directly [11][8].
In 2025 OpenAI used reasoning systems built on this line of work to reach top tier results in elite student competitions, achievements that Chen has tied to his own motivations. In July 2025 the company reported that a general purpose reasoning model had achieved gold medal level performance at the International Mathematical Olympiad, solving five of six problems [19][22]. The following month it reported a gold medal level score at the 2025 International Olympiad in Informatics, where the system placed sixth when ranked against human contestants [22]. The same competitions that Chen has coached human teams toward thus became benchmarks for the models his organization builds.
In September 2024 OpenAI promoted Chen to Senior Vice President of Research [5][12]. The change came as several senior leaders left the company, among them Chief Research Officer Bob McGrew and a research vice president, Barret Zoph, and shortly after the departure of Chief Technology Officer Mira Murati [5][12]. In the reorganization Chen took charge of the research organization in partnership with Pachocki, who held the title of Chief Scientist [5][12]. The same month, OpenAI released the o1 reasoning models [5].
In March 2025 OpenAI named Chen Chief Research Officer, an expanded role [1][2]. In the company's description of the change, Chen would drive scientific progress and keep the organization at the frontier of capability and safety, while bringing research and product development closer together so that research results reached products more quickly [1]. The announcement also expanded the role of Brad Lightcap, the Chief Operating Officer, who took on broader responsibility for business and day to day operations [1]. The arrangement left OpenAI with two heads of research, Chen as Chief Research Officer and Pachocki as Chief Scientist [2][11].
Reporting on the dual structure describes a division of labor in which Chen manages the research teams and the practical work of turning research into products, while Pachocki sets the longer term technical direction and research roadmap [11]. Chen has characterized the boundary as flexible, saying of himself and Pachocki that "there's fluidity in the roles. We're both researchers, we pull on technical threads" [11]. Part of his work involves deciding how to divide computing resources between projects aimed at near term products and longer term research meant to advance the next set of capabilities [11]. Coverage of the role has also noted the people management side of the job, including competition with other companies to recruit and keep researchers during a period of intense demand for AI talent [15].
In August 2025 OpenAI released GPT-5, a model that Chen helped oversee as Chief Research Officer and that the company launched during a livestreamed event on 7 August 2025 [23][24]. In interviews around the release, Chen described GPT-5 as a convergence of traditional pretraining and post training with deeper reasoning, and he said the team had made factual accuracy a priority, noting that "language models historically have been plagued by hallucinations and factual errors" [23][25]. He also discussed the growing role of synthetic data, output generated by earlier models rather than by people, which he said was proving effective in domains such as coding [25].
Chen has continued to speak for OpenAI's research direction through later leadership changes. In early January 2026, Jerry Tworek, a longtime vice president of research who had helped lead the reasoning work, announced that he was leaving the company to pursue research that he said was harder to do at OpenAI [26]. Chen praised Tworek's contribution, saying his impact would "be felt across OpenAI and our models for years to come," and he pointed to the company's 2026 roadmap and its work toward what he called an automated scientist [27]. As public scrutiny of the company intensified in early 2026, Chen pushed back on the suggestion that OpenAI had shifted toward a product focused agenda at the expense of research, stating that the majority of the company's computing resources remained allocated to foundational research rather than to product milestones [28]. In March 2026 he appeared in a public discussion with the mathematician Terence Tao on artificial intelligence and mathematical discovery, hosted through the OpenAI Forum [29].
Chen has spoken about the methods behind OpenAI's models in public talks and interviews. After the release of GPT-4.5 in February 2025, a large model that the company had developed under the code name Orion, Chen said the result showed that scaling could continue, stating that "GPT-4.5 really is proof that we can continue the scaling paradigm" [13][14]. He has also framed reasoning as a frontier that remains open, describing a gap in which models "know a lot of things but can't chain that knowledge together," and he has pointed to mathematics and coding as areas where progress in reasoning can be measured [11]. He has said that he continues to view scaling laws, the observed relationship between model size, data, and performance, as a workable basis for further progress [11].
Chen's research output is widely cited within machine learning. He is the first author of the Codex paper, one of the early and frequently referenced works on code generation by large language models, and the first author of the Image GPT paper on generative pretraining for images [10][21]. His published work, taken together, has been cited well over one hundred thousand times [20]. He has appeared as a speaker at technology and research events, where his talks have addressed multimodal models and reasoning [9]. His rise from research scientist to Chief Research Officer over roughly seven years at OpenAI has drawn attention in coverage of the company's leadership [6][12].
| Year | Paper | Role | Contribution |
|---|---|---|---|
| 2020 | Generative Pretraining from Pixels (Image GPT) | First author | Applied autoregressive sequence modeling to image pixels [21] |
| 2020 | Language Models Are Few-Shot Learners (GPT-3) | Co-author | Introduced GPT-3 and in-context learning [20] |
| 2021 | Zero-Shot Text-to-Image Generation (DALL-E) | Co-author | Text to image generation from a single model [20] |
| 2021 | Evaluating Large Language Models Trained on Code (Codex) | First author | Codex model and the HumanEval benchmark [10] |
| 2022 | Hierarchical Text-Conditional Image Generation with CLIP Latents (DALL-E 2) | Co-author | CLIP guided image generation [20] |
| 2023 | GPT-4 Technical Report | Co-author | Reported GPT-4, including multimodal input [20] |
| Field | Detail |
|---|---|
| Name | Mark Chen |
| Nationality | American |
| Field | Artificial intelligence, machine learning |
| Education | Massachusetts Institute of Technology, mathematics with computer science, 2012 |
| Before OpenAI | Quantitative trading, including Jane Street Capital |
| Joined OpenAI | 2018 |
| Known for | Leading DALL-E, Codex, GPT-4 vision, and o-series reasoning model work |
| Senior Vice President of Research | September 2024 |
| Chief Research Officer | March 2025 |
| Research partner | Jakub Pachocki, Chief Scientist |
| Other roles | Coach and team leader, United States International Olympiad in Informatics |