Andrea Zanette

Timeline

1. By 2017

   Software and fluid dynamics

   Before beginning his Stanford PhD in 2017, he works as a software developer in high-performance computing and at the von Karman Institute for Fluid Dynamics.

2. 2017–2021

   A Stanford doctorate

   He studies at Stanford University's Institute for Computational and Mathematical Engineering from 2017 to 2021, completing his PhD under Emma Brunskill and Mykel J. Kochenderfer. His earlier education includes a bachelor's degree in mechanical engineering. His doctoral research examines exploration, function approximation, adaptivity and learning from offline data.

3. 2021

   A dissertation award

   In 2021, he receives Stanford's Gene Golub Doctoral Dissertation Award for his research on efficient exploration in reinforcement learning.

4. 2021

   Limits of batch reinforcement learning

   His 2021 ICML paper, “Exponential Lower Bounds for Batch Reinforcement Learning,” studies learning from previously collected data. It establishes settings with linear representations of action values in which batch reinforcement learning requires exponentially more data than online learning.

5. 2022

   A Berkeley research fellowship

   In 2022, he participates as a research fellow in the Learning and Games program at UC Berkeley's Simons Institute for the Theory of Computing. As a postdoctoral researcher at the institute, he works primarily with Martin Wainwright on the foundations of reinforcement learning.

6. 2024

   ArCHer trains language model agents

   In 2024, he coauthors ArCHer with Yifei Zhou, Jiayi Pan, Sergey Levine and Aviral Kumar at ICML. The framework trains language model agents across multiple turns, combining a high-level value function with a policy that produces individual tokens.

7. 2025

   Training models to reason efficiently

   With Daman Arora, he coauthors “Training Language Models to Reason Efficiently” in 2025. The method uses reinforcement learning to allocate computation according to task complexity; experiments on two open-weight reasoning models reduce inference costs while preserving most of their accuracy.

8. 2026

   Maximum likelihood reinforcement learning

   In 2026, he coauthors “Maximum Likelihood Reinforcement Learning,” introducing MaxRL. The method uses sampled outputs to connect reinforcement learning with maximum likelihood training, and its objective approaches maximum likelihood as more computation is used for sampling.

9. Since 2024

   Assistant professor at Carnegie Mellon

   He joins Carnegie Mellon University's Department of Electrical and Computer Engineering as an assistant professor in 2024, with a courtesy appointment in the Machine Learning Department. In 2026, he continues in this role and studies foundation models, including reasoning, alignment, efficiency and optimization.