Shengjia Zhao
Last reviewed
Jun 5, 2026
Sources
15 citations
Review status
Source-backed
Revision
v2 · 1,807 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 5, 2026
Sources
15 citations
Review status
Source-backed
Revision
v2 · 1,807 words
Add missing citations, update stale details, or suggest a clearer explanation.
Shengjia Zhao is a Chinese-born AI researcher who serves as chief scientist of Meta Superintelligence Labs (MSL), the artificial intelligence unit Meta formed in 2025. Mark Zuckerberg announced his appointment on 25 July 2025, describing him as a co-founder of the lab and its lead scientist from the start. [1][2] Before joining Meta, Zhao spent three years at OpenAI, where he was a contributor to ChatGPT and GPT-4 and a foundational contributor to the o1 reasoning model. [1][3] He holds a PhD in computer science from Stanford University and is known in the academic literature for work on generative models, calibration, and uncertainty quantification. [4][5]
Zhao sits at the intersection of two strands of modern AI: the academic study of probabilistic generative models and uncertainty, which he pursued during his doctorate, and the industrial development of frontier large language models, which occupied his time at OpenAI and now at Meta. [4][1] At OpenAI he was associated with several of the lab's most consequential releases, and his recruitment was part of a broader wave of departures from OpenAI to Meta in mid-2025. [3][1] In his current role he sets the scientific agenda for Meta Superintelligence Labs under Alexandr Wang, the former Scale AI chief executive whom Meta hired as its chief AI officer. [1][2]
Zhao studied at Tsinghua University in Beijing, where he completed his undergraduate degree in computer science between 2012 and 2016. [6] During his undergraduate years he spent time in 2014 as an exchange student at Rice University in Houston. [11] He then moved to the United States for graduate study, enrolling at Stanford University in 2016 and earning his PhD in computer science in 2022. [6] His doctoral advisor was Stefano Ermon, a Stanford professor whose group works on machine learning, probabilistic modeling, and AI for sustainability. [5][6] His dissertation centered on the calibration of probabilistic predictions and on deep generative models, the same topics that dominate his published work from the period. [11] Much of Zhao's research dates from this period and was carried out with Ermon and fellow Stanford students. [4][5]
During his doctorate Zhao received several fellowships and scholarships. He and his frequent collaborator Jiaming Song were named recipients of the Qualcomm Innovation Fellowship in 2018, an award given to PhD-student teams for promising research proposals. [12] He has also been reported as a recipient of a Google PhD scholarship and a JP Morgan PhD Fellowship during his time at Stanford. [11]
Zhao's academic work centers on deep generative models and on making the predictions of machine-learning systems trustworthy. His Google Scholar profile lists research interests in uncertainty quantification, generative models, and information theory. [4] A recurring theme is calibration: the question of whether a model's stated confidence in an outcome matches how often that outcome actually occurs, and how to reason about calibration in ways that are useful for downstream decisions rather than merely as an abstract statistical property. [5][7]
In 2017 Zhao, with Jiaming Song and Stefano Ermon, introduced InfoVAE, a family of training objectives for variational autoencoders intended to balance the quality of the learned latent representation against the fidelity of reconstruction; the work was later published at AAAI 2019. [8] InfoVAE became his most cited piece of purely academic research, with on the order of a thousand citations recorded on Google Scholar. [4] He went on to publish a series of papers on calibration and decision-making, including work that reframes calibration around the decisions a forecast is meant to inform. [5][7] In 2022 he and his co-authors received an Outstanding Paper Award at the International Conference on Learning Representations (ICLR) for "Comparing Distributions by Measuring Differences that Affect Decision Making," which proposes a way to compare two probability distributions according to how much their differences would change an optimal decision. [7]
Beyond generative modeling and calibration, Zhao contributed to work on the theory of representation learning and fairness, including "A Theory of Usable Information Under Computational Constraints" and "Learning Controllable Fair Representations," both with Jiaming Song, Aditya Grover, and Stefano Ermon and both presented in 2019. [4][9] His calibration line continued through 2022 with papers such as "Modular Conformal Calibration" at ICML and "Local Calibration: Metrics and Recalibration" at the Conference on Uncertainty in Artificial Intelligence (UAI). [9] Across his academic and industrial publications his Google Scholar profile records tens of thousands of citations and an h-index in the high twenties, although the bulk of that citation total comes from large multi-author OpenAI technical reports rather than his single-investigator academic work. [4]
The table below lists a selection of his peer-reviewed papers.
| Paper | Co-authors (selected) | Venue | Year |
|---|---|---|---|
| InfoVAE: Balancing Learning and Inference in Variational Autoencoders | Jiaming Song, Stefano Ermon | AAAI | 2019 |
| A Theory of Usable Information Under Computational Constraints | Jiaming Song, Stefano Ermon | ICLR | 2019 |
| Learning Controllable Fair Representations | Jiaming Song, Aditya Grover, Stefano Ermon | AISTATS | 2019 |
| Permutation Invariant Graph Generation via Score-Based Generative Modeling | Chenhao Niu, Yang Song, Jiaming Song, Aditya Grover, Stefano Ermon | AISTATS | 2020 |
| Improved Autoregressive Modeling with Distribution Smoothing | Chenlin Meng, Jiaming Song, Yang Song, Stefano Ermon | ICLR | 2021 |
| Calibrating Predictions to Decisions: A Novel Approach to Multi-Class Calibration | Michael P. Kim, Roshni Sahoo, Tengyu Ma, Stefano Ermon | NeurIPS | 2021 |
| Comparing Distributions by Measuring Differences that Affect Decision Making (Outstanding Paper Award) | Abhishek Sinha, Yutong He, Aidan Perreault, Jiaming Song, Stefano Ermon | ICLR | 2022 |
| Modular Conformal Calibration | Charles Marx, Willie Neiswanger, Stefano Ermon | ICML | 2022 |
Sources: ICLR 2022 award announcement, Google Scholar, and DBLP. [7][4][9]
Zhao joined OpenAI in June 2022, shortly after finishing his PhD, and remained there until 2025. [3][6] He was part of the team behind ChatGPT and is credited as a co-author of the research describing the system. [3] He also contributed to GPT-4 and to subsequent models in that line, including GPT-4.1. [1] His Google Scholar record lists him as a co-author on a series of OpenAI technical documents, among them the GPT-4 Technical Report, the GPT-4o System Card, the OpenAI o1 System Card, the gpt-oss model card, and the GPT-5 System Card, which together account for most of his recorded citations. [4]
His most frequently cited industrial contribution is to OpenAI's o1, the company's first model trained to "reason" by spending additional computation at inference time before answering. Zhao is listed among the project's foundational contributors, a roster that also included OpenAI co-founder Ilya Sutskever. [1][3] In announcing Zhao's move to Meta, Zuckerberg credited him with having "pioneered a new scaling paradigm," a phrase generally read as referring to the inference-time, reasoning-oriented approach embodied by o1. [1] At OpenAI he also worked alongside reasoning researcher Trapit Bansal, with whom he is associated in coverage of the o1 effort. [1]
In 2025 Meta assembled a new organization, Meta Superintelligence Labs, to pursue advanced AI, and recruited heavily from rival labs to staff it. [1][2] Zhao was among more than a dozen people who left OpenAI for Meta over that period, and his name first surfaced in a June 2025 report by The Information that listed him alongside three other OpenAI researchers joining the effort. [2][13] On 25 July 2025 Zuckerberg announced, in a post on Threads, that Zhao would be the lab's chief scientist, writing that "Shengjia co-founded the new lab and has been our lead scientist from day one." [1][2] Zuckerberg added that, with recruiting "going well" and the team "coming together," Meta had "decided to formalize his leadership role." [3] Alexandr Wang, who leads MSL as chief AI officer, said Zhao would "lead our scientific direction for our team." [1]
Coverage of the appointment noted that Zhao joined Meta alongside other former OpenAI researchers, including Jiahui Yu, Shuchao Bi, and Hongyu Ren, and that he had previously worked with reasoning researcher Trapit Bansal. [1][13] Reporting framed the hire as part of an aggressive talent campaign through which Meta sought to close the gap with OpenAI and Google. [3][2]
In late August 2025, several outlets, citing reporting by Wired, described a period of turbulence inside Meta's new AI organization in which a number of recently hired researchers left, some of them returning to OpenAI after only a few weeks. [14][15] In Zhao's case the reporting was that he had considered leaving and had gone as far as signing paperwork to return to OpenAI before Meta moved to retain him by formally naming him chief scientist of Meta Superintelligence Labs. [14][15] Unlike colleagues such as Avi Verma and Ethan Knight, who were reported to have actually departed for OpenAI within weeks, Zhao stayed at Meta after the appointment. [14] Meta characterized his position differently, with a spokesperson stating that he had been a co-founder of MSL and had held a leadership role from the start. [14][1]
As chief scientist of Meta Superintelligence Labs, Zhao is responsible for the lab's research agenda and reports to Wang within Meta's AI leadership. [1][2] Reporting from late 2025 and into 2026 continued to describe him in that role, including in connection with the large compute resources Meta has committed to the effort, such as its planned gigawatt-scale "Prometheus" cluster. [10] As of June 2026 no reputable source indicates that he has left the position.