Mathieu Blondel

Career

Blondel obtained a PhD in machine learning from Kobe University in 2013. From 2013 to 2019 Blondel was a research scientist at NTT's Communication Science Laboratories, and since 2019 has been a research scientist at Google DeepMind.¹

Research

Much of Blondel's research is on differentiable programming: making operations such as sorting, ranking, dynamic programming and optimization usable inside gradient-based learning. Examples include Soft-DTW, a differentiable loss for time series (ICML 2017, with Marco Cuturi), "Fast Differentiable Sorting and Ranking" (ICML 2020) and "Efficient and Modular Implicit Differentiation" (NeurIPS 2022).⁸⁹¹⁰³

Blondel's more recent papers concern language models, including decoding-time realignment (ICML 2024), teacher hacking in distillation (ICML 2025) and a 2025 preprint arguing that autoregressive language models can be read as energy-based models; Blondel is also listed with the Gemini 2.5 and Gemma 4 team reports. With Vincent Roulet, Blondel wrote The Elements of Differentiable Programming (2024).⁸¹¹⁵

Open-source work

Blondel was a core developer of scikit-learn and co-author of its 2011 JMLR paper, and led the lightning library for large-scale linear classification, regression and ranking. In JAX, Blondel was lead developer of JAXopt and a contributor to Optax, and now contributes to gemma, a library for using and fine-tuning Gemma models.²⁶⁷⁴¹²

Work and contributions

• JAXopt Open-source code

  Was the lead developer of JAXopt, a library of hardware-accelerated, batchable, differentiable optimizers in JAX, and is first author of "Efficient and Modular Implicit Differentiation" (2021), which the JAXopt README cites for its implicit differentiation framework.²³⁴¹³

• gemma (JAX library for Gemma) Open-source code

  Listed on Blondel's software page as a current contributor to gemma, an Apache-2.0 JAX library for using and fine-tuning Gemma models.²¹²

• The Elements of Differentiable Programming Published research

  Co-author with Vincent Roulet of the 2024 book reviewing differentiable programming from optimization and probability perspectives.⁵

• scikit-learn Open-source code

  Former core developer, co-author of the 2011 JMLR scikit-learn paper, and listed among the emeritus maintainers on the project's About page.²⁶⁷¹⁴

Timeline

1. 2011Co-authored the JMLR paper "Scikit-learn - Machine Learning in Python".⁶
2. 2013Completed a PhD in machine learning and joined NTT Communication Science Laboratories as a research scientist.¹
3. Mar 5, 2017Co-authored Soft-DTW, a differentiable loss function for time series, with Marco Cuturi.⁹
4. 2019Joined Google DeepMind as a research scientist.¹
5. Feb 20, 2020Released "Fast Differentiable Sorting and Ranking" as first author (ICML 2020).¹⁰
6. May 31, 2021Released "Efficient and Modular Implicit Differentiation", the paper behind JAXopt, as first author.³
7. Mar 21, 2024Released The Elements of Differentiable Programming with Vincent Roulet.⁵
8. Dec 17, 2025Released a preprint as first author connecting autoregressive language models and energy-based models.¹¹

Other documented roles

Roles the sources document that are not confirmed as current.

• Research scientist (2013–2019), NTT Communication Science Laboratories¹

Where they work

• Google DeepMind 64 people in the wiki

Related people

• Razvan PascanuAlso at Google DeepMind
• Roy FrostigAlso at Google DeepMind
• Alex GravesAlso at Google DeepMind
• Laurent SartranAlso at Google DeepMind
• Jascha Sohl-DicksteinAlso at Google DeepMind
• Fabian PedregosaML frameworks and libraries · Machine-learning theory
• Martín AbadiML frameworks and libraries · Machine-learning theory
• Alex SmolaMachine-learning theory · ML frameworks and libraries
• Lorenzo StellaML frameworks and libraries · Machine-learning theory
• Mikayla GawareckiML frameworks and libraries