Léonard Blier

Timeline

1. 2018

   Description length of deep networks

   With Yann Ollivier, he co-authors "The Description Length of Deep Learning Models", presented at NeurIPS 2018. The paper shows that deep neural networks can compress their training data losslessly even when the cost of encoding their parameters is counted, and that variational methods give poor compression bounds.

2. October 2018

   The Alrao optimization method

   In October 2018, with Pierre Wolinski and Yann Ollivier, he introduces Alrao (All Learning Rates At Once), in which each unit or feature of a neural network gets its own learning rate sampled from a random distribution. Stochastic gradient descent with Alrao performs close to stochastic gradient descent with an optimally tuned learning rate.

3. 2019

   Q-learning and time discretization

   In 2019, he co-authors a paper with Corentin Tallec and Yann Ollivier at the 36th International Conference on Machine Learning on making deep Q-learning robust to time discretization. It proves that Q-learning does not exist in continuous time and builds an off-policy algorithm that performs similarly over a wide range of time discretizations.

4. 2021

   Successor states and sparse goals

   In January 2021, with Tallec and Ollivier, he derives temporal difference algorithms for learning successor states and goal-dependent value functions. In June 2021, he and Ollivier prove that Hindsight Experience Replay is unbiased in deterministic environments and introduce unbiased deep Q-learning and actor-critic algorithms for infinitely sparse rewards.

5. 2022

   A doctorate at Paris-Saclay

   He earns a PhD from Université Paris-Saclay in 2022 with the dissertation "Some Principled Methods for Deep Reinforcement Learning". His advisors are Yann Ollivier and Marc Schoenauer.

6. June 2025

   Magistral, Mistral's first reasoning model

   In June 2025, he co-authors the report introducing Magistral, Mistral's first reasoning model, along with Mistral's own scalable reinforcement learning pipeline. Magistral Medium is trained for reasoning on top of Mistral Medium 3 with reinforcement learning alone, and Magistral Small is open-sourced under the Apache 2.0 license.

7. July 2025–January 2026

   Voxtral, Devstral and Ministral 3

   In July 2025, he co-authors Voxtral, a pair of multimodal audio chat models that understand spoken audio and text documents. In August 2025, he co-authors the report on Devstral-Small, a lightweight open-source model for code agents, and in January 2026, the report on the Ministral 3 series of language models in 3B, 8B and 14B sizes.

8. As of October 2026

   AI scientist at Mistral AI

   He is an AI scientist at Mistral AI. In July 2026, he co-authors Shieldstral, a 3B-parameter multimodal safety classifier, and Robostral Navigate, an 8B vision-language model that guides robots from a stream of monocular RGB images.