Quentin BERTRAND - Home
Since July 2024, I am a tenured researcher (‘chargé de recherche‘) at Inria Lyon in the Malice team, located in Laboratoire Hubert Curien. I am a Mila-affiliated member and a CIFAR Global Scholar, member of the Learning Brain and Machine program.
I currently work on generative models and multi-agent learning. More broadly, I study how learning algorithms can leverage and interact with synthetic data from deep generative models.
From November 2021 to June 2024, I was a post-doctoral researcher at Mila working with Gauthier Gidel and Simon Lacoste-Julien. Prior to this position, I did my Ph. D. at Inria Paris-Saclay (in the Parietal Team) under the supervision of Joseph Salmon and Alexandre Gramfort. I worked on the optimization and statistical aspects of high dimensional brain signal reconstruction.
The material for ICML 2026 Tutorial on the generalization of diffusion models is now online!
In particular,
We are organizing the Sixth ICLR Blog Post Track! Clarifying and consolidating important results matters more than ever, consider submitting!
We just signed the Inria - Mila Agreements, see you soon in Montréal!
We just gave a Tutorial on the generalization of diffusion models at ICML 2026 in Korea!
I was awarded the CIFAR Global Scholar Fellowship, see you soon in Canada!
I was awarded a joint ANR-NSERC 2025 funding to push further our work on the creativity of generative modeling. More details soon!
Our work on the generalization of flow matching deep generative models was accepted at NeurIPS 2025 as an oral (top 0.3%)
I was appointed Mila affiliated member!
I just launched the Inria-Mila associated team!
We just released a friendly blog post on normalizing flows and conditional flow matching techniques
Our recent works on self-consuming generative models and their biases got media coverage from the N.Y. Times, the Globe and Mail, and Sciences et Avenir!
We released a Python package for large scale optimization of sparse problems (4k download/month)
Here is a short resume and my list of publications.
The material for the AISSAI HACE 2026 - HPCAI hybridization Tutorial of Diffusion and Flow Matching can be found here.
The material for the IVADO Bootcamp: Focusing on the Current State of Agents can be found here.
Contact
Email: quentin [dot] bertrand AT inria [dot] fr
News
05-19-2026 I am now a CIFAR Global Scholar, see you soon in Canada!
11-03-2025 I was awarded a joint ANR-NSERC 2025 funding to push further our work on the creativity of generative modeling. More details soon!
I was appointed Mila affiliated member! See you in Montréal!
Our work on the generalization of flow matching deep generative models was accepted at NeurIPS 2025 as an oral! (top 0.3%) see you in San Diego!
08-15-2025 I gave a tutorial on cooperation in multi-agent reinforcement learning at the IVADO Bootcamp: Focusing on the Current State of Agents
We just released a paper that investigates the generalization of flow matching deep generative models!
We organize a minisymphosium on Generative Models and Optimal Transport for the 2025 Biennal SMAI Conference. It will take place from June 2nd to June 6th, see you in Carcans-Maubuissons!
05-02-2025 Our paper Self-Play Q-learners Can Provably Collude in the Iterated Prisoner's Dilemma was just accepted to ICML, see you in Vancouver!
04-08-2025 We just gave a 12h tutorial on deep generative models at the Senegalese Computer Science Society and AI Hub Sénégal Summer School. The material can be found here. Thanks to Inria and the French Embassy for making this possible: we had amazing interactions in Dakar!
03-20-2025 Our Inria-Mila associated team was just created. See you soon in Montréal!
02-01-2025 Our friendly blog post on normalizing flows and conditional flow matching techniques was accepted at the ICLR 2025 Blog Post Track.
01-15-2025 I just gave a talk on how to retrain on synthetic data at the Mathematic Image and Applications conference! Here are the slides
12-09-2024 I was at NeurIPS to present our paper Self-Consuming Generative Models with Curated Data Provably Optimize Human Preferences
11-26-2024 We just released a blog post on normalizing flows and conditional flow matching techniques!
09-29-2024 I was delighted to be a keynote speaker for the ECCV workshop ¨The Dark Side of Generative AIs and Beyond¨, here are the slides
09-26-2024 Our paper showing that Self-Consuming Generative Models with Curated Data Provably Optimize Human Preferences was just accepted to NeurIPS! See you in Vancouver!
08-26-2024 Our recent works on self-consuming generative models and their biases was featured in the N.Y. Times!
07-01-2024 Just started as an Inria researcher in the Malice team
04-22-2024 The recording of the talk On the Stability of Iterative Retraining of Generative Models on their own Data at the Montreal Machine Learning Seminar can be found here
01-16-2024 Our paper On the Stability of Iterative Retraining of Generative Models on their own Data was accepted to ICLR 2024 with a spotlight, see you in Vienna!
12-18-2023 We just released our paper proving that Self-Play Q-learners can provably learn to collude in the iterated prisoner dilemma!
01-07-2023 On July 1st, 2024 I will join Inria as a Research Scientist!
Previous News
05-12-2023 I will present our paper On the Limitations of Elo: Real-World Games are Transitive, not Additive at the Berkeley Multi-Agent Reinforcement Learning Seminar
Our paper Synergies between Disentanglement and Sparsity: Generalization and Identifiability in Multi-Task Learning has been accepted at ICML 2023, see you in Hawaii!
Our paper On the Limitations of Elo: Real-World Games are Transitive, not Additive has been accepted to AISTATS 2023, see you in Spain!
I just presented our paper Synergies between Disentanglement and Sparsity: a Multi-task Learning Perspective at the Canadian Mathematical Society Winter Workshop
I just presented our two papers Beyond L1: Faster and Better Sparse Models with skglm and The Curse of Unrolling: Rate of Differentiating Through Optimization at NeurIPS 2022
I was awarded the top reviewer award at NeurIPS 2022!
Our papers Beyond L1: Faster and Better Sparse Models with skglm and The Curse of Unrolling: Rate of Differentiating Through Optimization have been accepted to NeurIPS 2022!