Paul Vicol

Paul Vicol is a Research Scientist at Google DeepMind who works on reasoning in large language models and Gemini Thinking models, and a coauthor of the ICML 2021 Outstanding Paper on Persistent Evolution Strategies. After belief-change research in 2015 and 2016, he earned a machine learning PhD at the University of Toronto under Roger Grosse, with papers that include Flipout and MovieGraphs in 2018 and Self-Tuning Networks in 2019. A 2020 research internship at Google Brain, hosted by Jascha Sohl-Dickstein and Luke Metz, centered on Persistent Evolution Strategies, and ICML gave the paper an Outstanding Paper award in 2021. In October 2022, he joined Google DeepMind, and he coauthored Direct Reward Fine-Tuning for diffusion models in 2023 and a 2025 paper showing that Veo 3 solves tasks zero-shot. He is also the Managing Editor of Transactions on Machine Learning Research.

Now at: Google DeepMind

Timeline

2015–2016

Belief change and Equibel

He publishes belief-change research with James Delgrande and Torsten Schaub. An LPNMR 2015 paper describes Equibel, a software system for belief change operations on arbitrary graph topologies, and an ECAI 2016 paper presents minimization-based approaches to iterated multi-agent belief change.
2018

Flipout and MovieGraphs

In 2018, he coauthors Flipout with Yeming Wen, Jimmy Ba, Dustin Tran and Roger Grosse; the method decorrelates gradients within a mini-batch through pseudo-independent weight perturbations and appears at ICLR 2018. He is first author of MovieGraphs, a dataset of graph-based annotations of social situations in movie clips, which is a spotlight at CVPR 2018.
2019–2020

Gradient-based hyperparameter tuning

Working under Roger Grosse at the University of Toronto, where he obtains his machine learning PhD, he shares first authorship of Self-Tuning Networks at ICLR 2019 with Matthew MacKay; the networks update their own hyperparameters online during training. In 2020, he coauthors Optimizing Millions of Hyperparameters by Implicit Differentiation with Jonathan Lorraine and David Duvenaud at AISTATS.
2020–2021

Research at Google Brain

In 2020, he is a research scientist intern at Google Brain, hosted by Jascha Sohl-Dickstein and Luke Metz, and works on Persistent Evolution Strategies (PES). He continues there as a student researcher from 2020 to 2021, extending PES.
2021

An ICML Outstanding Paper

In 2021, ICML gives an Outstanding Paper award to Unbiased Gradient Estimation in Unrolled Computation Graphs with Persistent Evolution Strategies, which he coauthors with Luke Metz and Jascha Sohl-Dickstein. PES splits a computation graph into truncated unrolls, applies an evolution strategies update after each one, and removes bias by accumulating correction terms across the unrolls.
October 2022

Joining Google DeepMind

After a 2021 research scientist internship at DeepMind on Deep Equilibrium Models, hosted by Jeff Donahue and Karen Simonyan, he joins Google DeepMind as a Research Scientist in October 2022. He works on Gemini Thinking models.
2023–2024

Fine-tuning diffusion models with DRaFT

With Kevin Clark, Kevin Swersky and David J. Fleet, he coauthors Directly Fine-Tuning Diffusion Models on Differentiable Rewards, posted in September 2023 and published at ICLR 2024. The paper introduces Direct Reward Fine-Tuning (DRaFT), which backpropagates the reward gradient through the full sampling procedure and improves the aesthetic quality of images from Stable Diffusion 1.4.
As of October 2026

Gemini reasoning and TMLR

In 2025, he coauthors Video models are zero-shot learners and reasoners, which shows that Veo 3 can solve a broad variety of tasks it was not explicitly trained for, from segmenting objects to detecting edges. He is a Research Scientist at Google DeepMind working on reasoning in LLMs, including Gemini Thinking models such as Gemini 2.5 Pro and the image generation capabilities of Gemini, and he is the Managing Editor of Transactions on Machine Learning Research.