large language models, reasoning, fine-tuning, test-time computation, reinforcement learning with human feedback, world models
News RSS feed
News: new Gave an adidas AI talk, “When AI Learns to Tie Its Shoes: A Future Beyond LLMs”, for adidas AG in Herzogenaurach, Germany. 👟
News: new Invited to STVR’s Do kríža for a television debate on current AI developments, benefits, risks, and broader implications.
News: new Interview with Libération journalist Olivier Monod on rewards in AI, conflicting objectives, and limitations of reward-based approaches.
News: new Speaking at the Centre Jean Gol conference “China: Partner, Competitor, or Enemy of the EU?” in Brussels on October 16. event
News: new Joining NVIDIA Startup Day Berlin at GTC Berlin on October 20 for the panel “Building and Scaling European Models”.
News: new Written SME interview for Respect Forum / Slovak Telekom.
News: new France Culture: joining La Fabrique de l’information for a live public discussion on AI risk narratives, superintelligence, and separating concrete risks from alarmist myths.
Bio
Michal is the Founding Researcher at Isara Labs, tenured researcher at Inria, and a lecturer at MVA at ENS Paris-Saclay. Michal is primarily interested in designing algorithms that would require as little human supervision as possible. He works on methods and settings that are able to deal with minimal feedback, such as deep reinforcement learning, bandit algorithms, self-supervised learning, or self play. Michal has recently worked on representation learning, world models and deep (reinforcement) learning algorithms that have some theoretical underpinning. In the past he has also worked on sequential algorithms with structured decisions where exploiting the structure leads to provably faster learning. Michal is now working on a new generation of large language models (LLMs), in addition to providing algorithmic solutions for their scalable test-time inference, fine-tuning and alignment. He received his PhD in 2011 from the University of Pittsburgh, before getting a tenure at Inria in 2012 and co-creating Google DeepMind Paris with R. Munos. In 2024, he became a Principal Llama Scientist at Meta, building online reinforcement learning stack and research for Llama 3. In 2025, he joined Isara Labs as a founding researcher.
Selected work
Research threads spanning frontier models, representation learning, bandits, and sparsification.
Nash-MD / NLHF
Preference learning as a two-player game, connecting LLM alignment, human feedback, and provable Nash-equilibrium algorithms.
NLHF · Nash-MD · preference learning
IX / implicit exploration
Implicit exploration replaces explicit exploration mixing with a biased loss estimator, giving sharp guarantees under partial and side-observation feedback.
side observations · data-dependent bounds · contextual bandits
Bandits & structured exploration
A long line on learning faster from structure: graphs, side observations, combinatorial actions, and Bayesian exploration.
Spectral · graph bandits · best-arm ID · semi-bandits · Thompson sampling
SQUEAK & sparsification
Single-pass sparsification of kernel matrices and graphs using ridge leverage scores, with scalable distributed variants.
Kernel approximation · ridge leverage scores · spectral sparsification
Coming up
Contact
Paris, France
40 avenue Halley
59650 Villeneuve d'Ascq, France
+33 3 59 57 78 01
4, avenue des Sciences
91190 Gif-sur-Yvette, France