Julia Kempe

Silver Professor of Computer Science, Mathematics and Data Science
Center for Data Science and Courant Institute, New York University
Senior Researcher, Meta FAIR Paris
Email: kempe@nyu.edu | Google Scholar | đť•Ź
Publications CV Awards Talks Teaching Group Research Blog Official Bio Blurb
Julia Kempe

I am a Silver Professor of Computer Science, Mathematics and Data Science at NYU's Center for Data Science and the Courant Institute. I am currently also a Senior Researcher at Meta FAIR Paris leading the Foundations of Reasoning Team.

From 2018 to 2023, I served as Director of the Center for Data Science at NYU, where I was a member of the NYU Senior Leadership Team. My current research focuses on the theory and empirics of machine learning and AI.

Research Interests: Machine Learning Theory, Deep Learning & Foundation Models, AI Safety & Robustness, ML for Science (Physics)

For Prospective PhD Students: Applications for the Ph.D. program are reviewed by a department-wide committee and are very competitive. Please apply directly to the PhD Program of the Center for Data Science or the Computer Science Program. Do not forget to mention my name (and of other faculty you'd like to work with) in the application AND in your research statement.

Prospective MS Independent Study Students / Undergraduates (not for 2025/26): I usually work with students that are known to me through the capstone course or other teaching. That said if you feel there is an exceptional match in research interests (mainly theory and empirics of machine learning, LLM reasoning and self-improvement, ML interpretability and robustness, world modeling), please email me and make a connection between your interests and my recent work. Please add [MSU-abc] to your subject line to confirm that you have read this. Please do not apply if you want to work on quantum computing.

Due to increasing volume of emails, I apologize for not being able to reply to all of them.

Publications

Recent publications in Machine Learning (2022- ). For a complete list, see my CV or Google Scholar.

2026
Teaching Models to Teach Themselves: Reasoning at the Edge of Learnability
Shobhita Sundaram, John Quan, Ariel Kwiatkowski, Kartik Ahuja, Yann Ollivier, Julia Kempe
Preprint, 2026Blog
[arXiv] | [External Blog]
OpenApps: Simulating Environment Variations to Measure UI-Agent Reliability
Karen Ullrich, Jingtong Su, Claudia Shi, Arjun Subramonian, Amir Bar, Ivan Evtimov, Nikolaos Tsilivis, Randall Balestriero, Julia Kempe, Mark Ibrahim
ICLR, 2026
[arXiv] | [Code]
How reinforcement learning after next-token prediction facilitates learning
Nikolaos Tsilivis, Eran Malach, Karen Ullrich, Julia Kempe
ICLR, 2026
[arXiv] | [CDS Blog]
Workshop version: "How reinforcement learning after next-token prediction facilitates learning" — NeurIPS Workshop on Principles of Generative Modeling (PriGM), 2025. Oral
Soft Tokens, Hard Truths
Natasha Butt, Ariel Kwiatkowski, Ismail Labiad, Julia Kempe*, Yann Ollivier* (*Equal senior authorship)
ICLR, 2026
[arXiv] | [CDS Blog]
From Concepts to Components: Concept-Agnostic Attention Module Discovery in Transformers
Jingtong Su, Julia Kempe*, Karen Ullrich* (*Equal senior authorship)
ICLR, 2026
[arXiv] | [CDS Blog] | [Code]
2025
Embedding Trust: Semantic Isotropy Predicts Nonfactuality in Long-Form Text Generation
Dhrupad Bhardwaj, Julia Kempe, Tim G. J. Rudner
Preprint, 2025
[arXiv]
Don't Waste Mistakes: Leveraging Negative RL-Groups via Confidence Reweighting
Yunzhen Feng, Parag Jain, Anthony Hartshorn, Yaqi Duan*, Julia Kempe* (*Equal senior authorship)
Preprint, 2025
[arXiv]
Outcome-based Exploration for LLM Reasoning
Yuda Song, Julia Kempe, RĂ©mi Munos
NeurIPS Workshop on Aligning RL Experimentalists and Theorists (2nd Edition), 2025
[arXiv]
What Characterizes Effective Reasoning? Revisiting Length, Review, and Structure of Chain-of-Thought
Yunzhen Feng, Julia Kempe, Cheng Zhang, Parag Jain, Anthony Hartshorn
NeurIPS Workshop on Efficient Reasoning, 2025Spotlight
[arXiv] | [CDS Blog]
Tuning without Peeking: Provable Generalization Bounds and Robust LLM Post-Training
Ismail Labiad, Mathurin Videau, Matthieu Kowalski, Marc Schoenauer, Alessandro Leite, Julia Kempe, Olivier Teytaud
Preprint, 2025
[arXiv]
Asymmetric REINFORCE for Off-Policy Reinforcement Learning: Balancing Positive and Negative Rewards
Charles Arnal, Gaëtan Narozniak, Vivien Cabannes, Yunhao Tang, Julia Kempe, Rémi Munos
NeurIPS, 2025
[arXiv]
PILAF: Optimal Human Preference Sampling for Reward Modeling
Y. Feng, A. Kwiatkowski, K. Zheng, J. Kempe*, Y. Duan* (*Equal senior authorship)
ICML, 2025
[arXiv]
Workshop versions: ICLR Workshop on Bidirectional Human–AI Alignment, 2025; COLT Workshop on Foundations of Post-Training, 2025.
Strong Model Collapse
E. Dohmatob, Y. Feng, A. Subramonian, J. Kempe
ICLR, 2025Spotlight
[arXiv]
DRoP: Distributionally Robust Data Pruning
A. Vysogorets, K. Ahuja, J. Kempe
ICLR, 2025Spotlight
[arXiv] | [CDS Blog]
Workshop version: "Towards Robust Data Pruning" — ICLR Workshop on Datacentric Machine Learning, 2024.
Flavors of Margin: Implicit Bias of Steepest Descent in Homogeneous Neural Networks
Nikolaos Tsilivis, Gal Vardi, Julia Kempe
ICLR, 2025
[arXiv]
Workshop version: NeurIPS Workshop on Mathematics of Modern Machine Learning, 2024.
Beyond Model Collapse: Scaling Up with Synthesized Data Requires Verification
Y. Feng, E. Dohmatob, P. Yang, F. Charton, J. Kempe
ICLR, 2025
[arXiv]
Workshop version: "Beyond Model Collapse: Scaling Up with Synthesized Data Requires Reinforcement" — ICML Workshop on Theoretical Foundations of Foundation Models (TF2M), 2024. Earlier title; main version revised framing and results.
2024
Emergent properties with repeated examples
F. Charton, J. Kempe
NeurIPS Workshop on Scientific Methods for Understanding Deep Learning, 2024🏆 Debunking Challenge WinnerOral
[arXiv]
Workshop version: "Repeated examples help learn arithmetic" — 4th NeurIPS Workshop on Mathematical Reasoning and AI, 2024. Same core contribution, reframed toward learning arithmetic.
The Price of Implicit Bias in Adversarially Robust Generalization
N. Tsilivis, N. Frank, N. Srebro, J. Kempe
NeurIPS, 2024
[arXiv] | [CDS Blog]
Workshop version: "The Best Algorithm for Adversarial Training" — ICLR Workshop on Bridging the Gap Between Practice and Theory in Deep Learning (BGPT), 2024. Earlier title; subset of authors (main version added N. Srebro).
On the Geometry of Regularization in Adversarial Training: High-Dimensional Asymptotics and Generalization Bounds
Matteo Vilucchio, Nikolaos Tsilivis, Bruno Loureiro, Julia Kempe
Preprint, 2024
[arXiv]
Mission Impossible: A Statistical Perspective on Jailbreaking LLMs
J. Su, J. Kempe*, K. Ullrich* (*Equal senior authorship)
NeurIPS, 2024
[arXiv] | [CDS Blog]
Workshop version: ICML Workshop on Theoretical Foundations of Foundation Models (TF2M), 2024.
Iteration Head: A Mechanistic Study of Chain-of-Thought
V. Cabannes, C. Arnal, W. Bouaziz, A. Yang, F. Charton, J. Kempe
NeurIPS, 2024
[arXiv]
Workshop version: ICML Mechanistic Interpretability Workshop, 2024.
Model Collapse Demystified: The Case of Regression
E. Dohmatob, Y. Feng, J. Kempe
NeurIPS, 2024
[arXiv]
Workshop version: "Towards a Theoretical Understanding of Model Collapse" — ICLR Workshop on Bridging the Gap Between Practice and Theory in Deep Learning (BGPT), 2024. Earlier title; main version includes expanded results.
Attacking Bayes: On the Adversarial Robustness of Bayesian Neural Networks
Y. Feng, T. G. J. Rudner, N. Tsilivis, J. Kempe
Transactions on Machine Learning Research, 2024TMLR Certification Award
[arXiv]
Workshop version: "Attacking Bayes: Are Bayesian Neural Networks Inherently Robust?" — 5th Symposium on Advances in Approximate Bayesian Inference (AABI), 2023. Shorter version; revised title and expanded analysis in TMLR.
A Tale of Tails: Model Collapse as a Change of Scaling Laws
E. Dohmatob, Y. Feng, P. Yang, F. Charton, J. Kempe
ICML, 2024
[arXiv] | [CDS Blog] | [New York Times Feature]
Workshop version: ICLR Workshop on Navigating and Addressing Data Problems for Foundation Models (DPFM), 2024.
Deconstructing the Goldilocks Zone of Neural Network Initialization
A. Vysogorets, A. Dawid, J. Kempe
ICML, 2024
[arXiv]
Mind the GAP: Improving Robustness to Subpopulation Shifts with Group-Aware Priors
T. Rudner, Y. Zhang, A. Wilson, J. Kempe
AISTATS, 2024AISTATS Notable Paper AwardOral
[arXiv]
Embarrassingly Simple Dataset Distillation
Y. Feng, R. Vedantam, J. Kempe
ICLR, 2024
[arXiv]
Workshop version: NeurIPS Workshop on Advancing Neural Network Training, 2023.
On the Robustness of Neural Collapse and the Neural Collapse of Robustness
J. Su, Y. Zhang, N. Tsilivis, J. Kempe
Transactions on Machine Learning Research, 2024
[arXiv] | [Code]
Workshop version: NeurIPS Workshop on Unifying Representations in Neural Models, 2023.
Kernels, Data & Physics (Lecture Notes, Les Houches 2022)
F. Cagnetta, D. Oliveira, M. Sabanayagam, N. Tsilivis, J. Kempe
Journal of Statistical Mechanics: Theory and Experiment, 2024
[arXiv]
2023
Galaxy Dataset Distillation by Self-Adaptive Trajectory Matching
H. Guan, X. Zhao, Z. Wang, Z. Li, J. Kempe
NeurIPS Workshop on Machine Learning and the Physical Sciences, 2023
[arXiv]
Connectivity Matters: Neural Network Pruning Through the Lens of Effective Sparsity
A. Vysogorets, J. Kempe
Journal of Machine Learning Research, 2023 (Vol. 24, Issue 99, pp. 1–23)
[arXiv]
2022
What Can The Neural Tangent Kernel Tell Us About Adversarial Robustness?
N. Tsilivis, J. Kempe
NeurIPS, 2022
[arXiv]
Wavelets Beat Monkeys at Adversarial Robustness
J. Su, J. Kempe
NeurIPS Workshop on Machine Learning and the Physical Sciences, 2022
[arXiv]
Adversarial Noise Injection for Learned Turbulence Simulations
J. Su, J. Kempe, D. Fielding, N. Tsilivis, M. Cranmer, S. Ho
NeurIPS Workshop on Machine Learning and the Physical Sciences, 2022
[Paper]
Can We Achieve Robustness from Data Alone?
N. Tsilivis, J. Su, J. Kempe
ICML Workshop on New Frontiers in Adversarial Machine Learning, 2022
[arXiv]

Selected Awards & Honors

2023 Julius Silver, Roslyn S. Silver, and Enid Silver Winslow Professorship, NYU
2022 Fellow of Asia-Pacific Artificial Intelligence Association
2018 Academia Europaea (elected member)
2010 Knight (Chevalier) in the National Order of Merit (Ordre de Mérite), France — Order of State awarded by the President of the French Republic
2010 Femme en Or de la Recherche, France — Awarded yearly to one woman in Science
2009 Krill Prize for Excellence in Scientific Research — Awarded by the Wolf Foundation to 6 Israeli young scientists
2009 Raymond and Beverly Sackler Career Development Chair, Tel Aviv University - awarded for the year 2009, awarded yearly to two faculty in the Exact Sciences
2006 Prix Irène Joliot-Curie de la jeune femme scientifique — Outstanding junior female researcher of the year, French Ministry of Science
2006 Médaille de Bronze du CNRS — Outstanding young researcher of the year

Official Bio Blurb

Julia Kempe is a researcher working on the Foundations of AI and Machine Learning. She is Silver Professor at the Courant Institute of Mathematical Sciences at NYU, currently on leave, and a Senior Researcher at Meta FAIR, where she leads the Foundations of Reasoning Team (FoRT).

Her career spans academia, frontier AI research, and quantitative finance. She previously served as Director of the NYU Center for Data Science, and spent nearly a decade at Renaissance Technologies, one of the world's most influential quantitative hedge funds. Before that, she held a Professor of Computer Science position at Tel Aviv University and was a Senior Scientist at the French CNRS, working on quantum computation and information.

Raised in East Germany, she experienced first-hand the collapse of political and scientific systems during the fall of the Berlin Wall — including qualifying for the International Mathematical Olympiad without ever being able to compete because her country ceased to exist. That early encounter with disrupted institutions shaped her lifelong interest in how knowledge, power, and technology interact.

Today, her work focuses on the foundations of reasoning in large AI models: what they learn, what they can and cannot understand, how they generalize, and how we should think about their growing role in science, the economy, and society - and how we should view AI not only as a technology, but as a new kind of cognitive infrastructure that is reshaping how humans make decisions and build knowledge.