Gal Vardi

Gal Vardi is a theoretical machine-learning researcher whose work studies when deep-learning systems generalize. After completing a BSc in the Talpiot Program and an MSc in 2015 at Hebrew University, he earned a computer-science PhD there in 2019 under Orna Kupferman, then held postdoctoral roles at Weizmann—where he received the 2020 Koshland Prize and a 2022 Feinberg Graduate School Prize—and jointly at the Toyota Technological Institute-Chicago and Hebrew University. His research includes a 2022 study on reconstructing training data from neural networks, a 2025 theory of autoregressive chain-of-thought learning, and a 2026 proof that grokking can arise in ridge regression; ICML gave the latter an Outstanding Paper Honorable Mention in 2026. As of October 2026, he is a Senior Scientist at Weizmann and teaches machine-learning theory.

Now at: Weizmann Institute of Science

Timeline

2015–2019

Talpiot and a PhD

After completing a BSc at the Hebrew University of Jerusalem as a cadet in the Talpiot Program, he earns an MSc there in 2015. He completes a computer-science PhD at Hebrew University in 2019 under Orna Kupferman.
2020–2022

Postdoc with Ohad Shamir

He joins Prof. Ohad Shamir's group at Weizmann's Department of Computer Science and Applied Mathematics as a postdoctoral fellow in 2020 and remains through 2022. He receives the Weizmann Institute's Koshland Prize in 2020 and the Feinberg Graduate School Prize for Outstanding Achievements in Postdoctoral Research in 2022.
2022

Reconstructing training data from neural networks

In 2022, he coauthors Reconstructing Training Data From Trained Neural Networks at NeurIPS. The paper shows that in some cases a significant fraction of the training data can be reconstructed from the parameters of a trained neural-network classifier.
2022–2024

Joint postdoc in Chicago and Jerusalem

From 2022 to 2024, he holds a joint postdoctoral appointment at the Toyota Technological Institute-Chicago and Hebrew University, with support from the National Science Foundation/Simons Collaboration on the Theoretical Foundations of Deep Learning.
June 2024

Principal investigator at Weizmann

In June 2024, he returns to Weizmann as a principal investigator. He studies the mathematical principles behind when deep-learning models generalize and how training shapes their performance.
2025

Theory of chain-of-thought learning

He coauthors A Theory of Learning with Autoregressive Chain of Thought in 2025. The paper formalizes learning prompt-to-answer mappings through iterative token generation and analyzes their sample and computational complexity.
2026

Provable grokking in ridge regression

With Mingyue Xu and Itay Safran, he coauthors To Grok Grokking: Provable Grokking in Ridge Regression in 2026. Their analysis proves that gradient descent with weight decay can overfit early, remain poorly generalized, and eventually achieve arbitrarily small generalization error in over-parameterized linear regression.
As of October 2026

Senior Scientist at Weizmann

ICML lists To Grok Grokking for an Outstanding Paper Honorable Mention in 2026. As of October 2026, he is a Senior Scientist in Weizmann's Department of Computer Science and Applied Mathematics. The institute's 2026 schedule lists him teaching Topics in Machine Learning Theory.