Moritz Hardt

Moritz Hardt studies the scientific foundations of machine learning and algorithmic decision-making, with a focus on social questions. A director at the Max Planck Institute for Intelligent Systems in Tübingen, he has co-developed methods for reliable leaderboards, fairness criteria for learned predictors, and a framework for predictions that influence their own targets. His recent benchmark work asks what AI evaluations can establish—and why benchmarks have worked as well as they have.[4][8][7][6][3]

In his ICLR 2024 invited talk, “The emerging science of benchmarks,” Hardt addressed annotator errors, whether model rankings hold beyond the evaluation that produced them, and the promise of multi-task benchmarks. In a May 2025 essay, he offered a distinctive formulation: “Benchmarks are the mistake that made machine learning. They shouldn't have worked, and yet they did.”[4][10]

Hardt develops these questions in The Emerging Science of Machine Learning Benchmarks, whose hardcover is scheduled for release by Princeton University Press on October 6, 2026. He distinguishes the replication of absolute accuracy numbers from the more substantial replication of model rankings. He also argues that greater diversity in multi-task benchmarks comes at the cost of greater sensitivity to artifacts, using an analogy with voting systems to explain the limits of aggregating evaluations.[3]

This recent work follows an earlier contribution to a specific benchmark problem: repeated submissions can overfit a leaderboard’s test data. Hardt and Avrim Blum addressed that problem in 2015, before his later invited talk and book on benchmarking.[8][4][3]

Hardt earned his doctorate in computer science at Princeton University in 2011 and subsequently worked at IBM Research and Google before joining the University of California, Berkeley. His employment disclosure lists IBM during 2013–2015 and Google during 2015–2017. He also co-founded the conference Fairness, Accountability, and Transparency in Machine Learning.[5][2]

In The Ladder: A Reliable Leaderboard for Machine Learning Competitions, Blum and Hardt introduced leaderboard accuracy, a notion tailored to tracking each team’s best submission under repeated, adaptive evaluation. They developed the Ladder algorithm and reported theoretical guarantees, resistance to practical adversarial attacks, and useful performance on submission files from a real Kaggle competition.[8]

In the 2016 paper Equality of Opportunity in Supervised Learning, Hardt, Eric Price, and Nathan Srebro proposed a discrimination criterion and a method for optimally adjusting a learned predictor to satisfy it, assuming access to predictions, targets, and protected-group membership. Their criterion uses joint statistics rather than interpretations of individual features. They examined the limits of such statistical tests and illustrated their approach with FICO credit scores; the adjustment removes discrimination according to their definition.[7]

Hardt also co-authored Understanding deep learning requires rethinking generalization with Chiyuan Zhang, Samy Bengio, Benjamin Recht, and Oriol Vinyals. First submitted in 2016 and published at ICLR 2017, the paper reported that image-classification convolutional networks trained with stochastic gradient methods could fit random labels—even with explicit regularization, and even when the images were replaced by unstructured random noise. The authors used these findings to challenge conventional explanations of why large neural networks generalize well.[9]

In the 2020 paper Performative Prediction, Juan C. Perdomo, Tijana Zrnic, Celestine Mendler-Dünner, and Hardt developed a framework for predictions that change their targets through the decisions they support. They introduced performative stability, an equilibrium notion calibrated against future outcomes produced by acting on predictions. Their results give conditions under which retraining converges to a stable point of nearly minimal loss.[6]

In 2021, Hardt was appointed a director at the Max Planck Institute for Intelligent Systems and founded its Social Foundations of Computation department. The institute’s October announcement stated that he had served in a secondary appointment since July 1, with a full-time start planned for January 2022. Celestine Mendler-Dünner, who had worked with him at Berkeley, became the department’s first group leader and helped shape its direction.[5]

At the appointment, Hardt advocated developing algorithmic methods that incorporate social foundations from the outset, rather than treating social problems only as consequences of technology. He recalled reaching social questions through his work on differential privacy. His stated goal as director was to make it easier for other researchers to pursue important questions receiving too little attention.[5]

Alongside his research, Hardt co-authored Fairness and Machine Learning: Limitations and Opportunities with Barocas and Narayanan, and Patterns, Predictions, and Actions with Recht. These books extend his writing across algorithmic fairness and the foundations of machine learning.[2][4]