Over the last decade, Professor Bin Yu and her research group have pioneered and refined a new data science paradigm called Veridical Data Science (VDS), beginning with a 2020 PNAS paper co-authored by Yu and her former Ph.D. student K. Kumbir. VDS is designed to set up operational guardrails for the entire data science cycle towards "truthful" results by relying on three core principles: Predictivity, Computability, and Stability (PCS).

The PCS principles govern the framework: predictability, now standing for general reality check beyond supervised learning, ensures models reflect reality by checking results generalize to unseen data and corroborate domain knowledge; computability focuses on algorithmic efficiency and feasibility, and data-inspired simulations; and stability expands upon traditional uncertainty to include reasonable variations caused by human choices, such as data cleaning and model choices.

Unlike traditional statistical and ML methods that focus on sample variability, VDS provides a holistic framework. By mandating a rigorous "reality check" and reasonable stability checks and Pred-checked aggregations (defined in context) at every stage of the data science life cycle, VDS moves beyond reproducibility issues toward truly responsible, trustworthy, and actionable decision-making, and in-context.

“It is really exciting to see the surging interest in VDS across academia and industry. AI is bringing both great promises and serious perils,” said Yu. “VDS is well positioned to systematically make the promises trustworthy and mitigate the perils (or to increase AI safety).”

VDS has already demonstrated a significant real-world impact, specifically in healthcare. In collaboration with the Ashley Group at Stanford Medical School, and published in a 2025 paper in Nature Cardiovascular Research, VDS was used to develop a hypothesis prioritization method, lo-siRF, at the frontier of human functional genomics. 80% of the method’s gene or epistatic gene interactions were successfully confirmed by experimental validation for a heart disease hypertrophic myopathy. She recently appeared on the Sano Genetics Podcast to discuss this work and how her early life shaped her as a researcher and person. VDS has seen success in reducing cost by 55% in SOTA prostate cancer detection in collaboration with the Chinnaiyan Group at the University of Michigan Medical School. She is leading a project called “Green Shielding” in collaboration with the Kornblith Group at UCSF and others. GreenShielding is a new user-centric approach towards trustworthy AI and a conceptual and empirical framework for deployment-aware LLM or genAI evaluation illustrated in medical diagnosis.

She recently became a co-PI at TRUST - the new Norwegian Centre for Trustworthy AI. She co-leads at TRUST a theme on veridical AI with TRUST Director Arnoldo Frigessi of University of Oslo. She won the 2025 Peter Sather grant for her work on Veridical AI together with Frigessi. This grant supports international research and educational collaboration between UC Berkeley and Norway. She has also been awarded a UK AI Security Institute (AISI) “Alignment” grant and received a gift from Coefficient Giving that supports her group for a retreat to discuss and plan for research related to AI Safety.

Co-authored with Statistics alum and her former Ph.D. student Rebecca Barter, Ph.D.‘19, Professor Yu’s VDS book was published by MIT Press in 2024 in their machine learning series, with an online version available at vdsbook.com and a very positive inaugural book review of Harvard Data Science Review (HDRS) by Yuval Benjamini (Statistics alum) and Yoav Benjamini. The book is authored to be highly accessible and to train critical thinking through narratives and case studies. It is designed for upper-division undergraduates and beginning graduate students, and domain experts alike.

As described by CambioML founder and Forbes 30 under 30 winner Rachel Hu, M.A. ‘19, “Most bad ‘AI decisions’ aren’t model problems. They’re human-decision problems around what was filtered out, how the data was cleaned, and which columns you silently dropped at 1 am!” said Hu. “That’s why I’m excited about VDS and its simple but sharp framework.”

VDS has been found essential in the Bridgewater AIA Lab and AI Start-up Traversa,l and has been recommended to the PIs of the Early Detection Research Network (EDRN) of the National Cancer Institute.

From Jas Sekhon, Chief Scientist, Head of AI, Bridgewater Associates (now Chief Strategy Officer, Google DeepMind), which is one of the largest hedge funds in the US: 

In AIA Labs at Bridgewater, we use artificial intelligence and machine learning to understand the global economy and manage investments in a noisy, adversarial, high‑stakes environment. Standard statistical inference techniques make strong assumptions for the sake of analysis; when applied blindly, without verifying these assumptions or accounting for researcher judgement calls, they are dangerous and convey a false sense of confidence. The PCS framework influences how we develop, apply, and govern our methods by forcing us to interrogate both collected data and the theoretical data-generating process as we develop algorithms. We lean heavily on predictive checks with collaborators’ domain knowledge, stability analysis of downstream results, and systematic documentation and probing of researcher judgement calls. Instead of treating scientific inference as purely algorithmic, we require both substantive and statistical justification. This disciplined use of PCS is essential to trusting our scientific process and deploying our models with confidence in real‑world investment decisions.

From Raaz Dwivedi, co-founder of Traversal,  CTO, and Assistant Prof. Cornell Tech. Traversal is an AI start-up that has raised $48M and featured in FORTUNE in 2025:

Traversal builds an AI Site Reliability Engineer (AI-SRE) that helps companies like American Express, Pepsi, and DigitalOcean diagnose outages by searching petabytes of telemetry and code to deliver fast, trustworthy root causes and resolutions. In these high-pressure settings—where signals are very weak and the search space is enormous—the PCS framework is essential. First get accuracy (P). Then speed (C). Then stability (S). That’s how we develop at Traversal AI.

Moreover, the Steering Committee meeting of EDRN (Early Detection Research Network) was held in March 2026. EDRN is a research network funded by NCI (National Cancer Institute) and this March meeting gathered ~100 guarantee researchers and NCI project directors. EDRN Coordinating Center Principal Investigator, Dr. Ziding Feng of Hutchinson Cancer Center, recommended that EDRN data analysis projects follow the PCS framework for Veridical Data Science. This recommendation was well received by the attendees.

Yu is the CDSS Chancellor's Distinguished Professor, Departments of Statistics and Electrical Engineering & Computer Sciences, and Center for Computational Biology at the University of California, Berkeley. She obtained her BS Degree in Mathematics from Peking University and her MS and Ph.D. Degrees in Statistics from UC Berkeley. She was an Assistant Professor at UW-Madison, a Visiting Assistant Professor at Yale University, a Member of the Technical Staff at Lucent Bell-Labs, and a Miller Research Professor at Berkeley. She was a Visiting Faculty member at MIT, ETH, the Poincaré Institute, Peking University, INRIA-Paris, the Fields Institute at the University of Toronto, the Newton Institute at Cambridge University, and the Flatiron Institute in NYC. She has also served as Chair of the Department of Statistics at UC Berkeley. She has played a crucial role in shaping the intellectual and organizational vision for the Division of Computing, Data Science, and Society (CDSS) at UC Berkeley as a faculty advisory committee member.

(photo via David Apilado, Jr.)