Alireza Makhzani

Alireza Makhzani is a machine learning researcher whose work spans sparse representations, generative models and probabilistic inference in language models. He earned a bachelor's degree at Amirkabir University of Technology in 2010 and a master's degree at the University of Toronto in 2012, proposed k-sparse autoencoders in 2014, and coauthored adversarial autoencoders in 2016. From 2018 to 2024, he was a faculty member at the Vector Institute and a Canada CIFAR AI Chair. His coauthored twisted sequential Monte Carlo paper received an ICML 2024 Best Paper Award. He is a research scientist at Google DeepMind, where he works on Gemini post-training for reasoning and coding.

Now at: Google DeepMind
Badges: Bridge
Links: Website, Google Scholar and X

Timeline

2010–2012

Degrees in Iran and Toronto

He earns a bachelor's degree from Amirkabir University of Technology in 2010 and a master's degree from the University of Toronto in 2012.
2014

K-sparse autoencoders

In 2014, he coauthors “k-Sparse Autoencoders” with Brendan Frey at ICLR. The method retains the k highest activities in an autoencoder's hidden layers and reports better MNIST and NORB classification than denoising autoencoders, dropout-trained networks and restricted Boltzmann machines.
2015–2017

Google Brain and StarCraft II

He interns with Google Brain in 2015 on generative models of images and with Google DeepMind in 2016 on deep reinforcement learning for StarCraft II. He coauthors the 2017 paper “StarCraft II: A New Challenge for Reinforcement Learning,” which introduces the StarCraft II Learning Environment for reinforcement-learning research.
2016

Adversarial autoencoders

In 2016, he coauthors “Adversarial Autoencoders,” which uses a generative adversarial network to match an autoencoder's aggregated latent-code distribution to an arbitrary prior. The method can disentangle image style and content and achieves competitive generative performance on MNIST, Street View House Numbers and Toronto Face datasets.
2017–2018

A Toronto PhD

At the University of Toronto, he presents “Unsupervised Representation Learning with Autoencoders” at a thesis examination in October 2017 with Brendan Frey as supervisor. He completes a PhD in 2018 as part of the Machine Learning Group.
2018–2024

Faculty at the Vector Institute

From 2018 to 2024, he is a faculty member at the Vector Institute and a Canada CIFAR AI Chair. His research in that period covers language-model reasoning, alignment and inference; diffusion models; variational inference and Monte Carlo methods; and computational optimal transport.
2024

An ICML Best Paper Award

He coauthors “Probabilistic Inference in Language Models via Twisted Sequential Monte Carlo,” which receives an ICML 2024 Best Paper Award. The work uses learned twist functions to direct sampling toward target sequences and introduces bidirectional sequential Monte Carlo bounds for evaluating inference accuracy.
As of October 2026

Research scientist at Google DeepMind

He is a research scientist at Google DeepMind in San Francisco, where he works on Gemini post-training for reasoning and coding. His current research covers reinforcement learning for large language models, test-time compute, probabilistic and RL-based search and planning, and knowledge distillation.