Intervenant

Marc Lelarge (Inria), Nathanaël Fijalkow, Guillaume Baudart (Inria), Xavier Hinaut (Inria), Philippe Suignard (EDF), Pierre-Yves Oudeyer (Inria), Yannis Bendi-Ouis (Inria), Théo Stoskopf (ENS Lyon/LIP), Rémi Louf (.txt)

Contexte scientifique

– Architecture / Transformer / Attention
– Reasoning / RLHF
– Agent / RAG / Prompt augmentation / Tooling
– Quantization / Constrained Inference
– Code generation
– Test generation
– Proof generation
– Séminaires / Retour d’expérience

Programme préliminaire

Monday 9-12:
Introduction to Transformers
[Nathanaël Fijalkow](https://games-automata-play.com/) (CNRS)
This session introduces the Transformer architecture and the self-attention mechanism that revolutionized NLP. We will explore how these models process sequential data, providing the necessary foundation for understanding how LLMs handle both natural languages and structured programming languages.

Monday 15-18:
Data cleaning and pre-training
[Wissam Antoun](https://wissamantoun.com/) (INRIA Paris)
We examine the lifecycle of a model before it is “ready,” from massive-scale data collection to self-supervised learning. The course covers the challenges of data quality and diversity, including the specific role that source code plays in enhancing a model’s logical reasoning capabilities.

Tuesday 9-12:
Post-training
[Nathanaël Fijalkow](https://games-automata-play.com/) (CNRS)
This lecture covers the transition from a raw base model to a functional assistant through Supervised Fine-Tuning (SFT) and Reinforcement Learning. We will discuss how these techniques are used to align models with human intent, specifically for following complex technical instructions.

Tuesday 15-18:
LLM and AI Safety
[Wissam Antoun](https://wissamantoun.com/) (INRIA Paris)
This session explores the ethical and technical safeguards required for deploying LLMs. We will discuss alignment, bias mitigation, and “red teaming,” with a focus on ensuring the reliability and security of model outputs in sensitive contexts like software development.

Wednesday 9-12:
State Space Models : Fondements Mathématiques et Efficacité Computationnelle
[Yannis Bendi-Ouis](https://www.naowak.fr/) (INRIA Bordeaux)
Peut-on combiner l’entraînement parallèle des Transformers avec l’inférence rapide des RNNs ? C’est la promesse des State Space Models
(SSM) modernes. Ce cours propose une plongée approfondie dans les State Space Models (SSM), une alternative puissante aux architectures basées sur l’attention. Nous explorerons comment ces modèles font le pont entre les systèmes dynamiques continus et le Deep Learning discret. À travers l’étude de modèles emblématiques comme S4, H3 et Mamba, nous détaillerons les mécanismes mathématiques clés, notamment la diagonalisation et la discrétisation sélective, qui permettent une inférence rapide et une gestion optimale de la mémoire. L’objectif est de fournir les outils théoriques pour comprendre pourquoi et comment ces architectures redéfinissent l’état de l’art en modélisation de séquences.

Thursday 9-12:
Training Agents
[Laetitia Teodorescu](https://dblp.org/pid/261/3333.html) (Adaptive-ML)
Moving beyond simple text generation, this session introduces LLM-based agents that can use tools and interact with environments. Participants will learn how models are trained to plan and execute multi-step tasks, such as navigating a codebase or interacting with a compiler.

Thursday 15-18:
World Models
[Václav Volhejn](https://vvolhejn.com/) (Kyutai)
This lecture delves into the concept of “World Models”, the internal representations LLMs build of the processes they describe. We will discuss how understanding the underlying “rules” of a system (physical or logical) allows models to predict and simulate complex outcomes.

Friday 9-12:
LLMs for test generation
[Xavier Blanc](https://www.labri.fr/perso/xblanc/) (Université de Bordeaux)
This session focuses on the practical application of LLMs to software quality assurance. We will study how models can be leveraged to automatically generate unit tests, identify edge cases, and assist in formal verification, bridging the gap between natural language requirements and executable code.

Friday 15-18:
Mechanistic interpretability
[David Louapre](https://scienceetonnante.com/) (Hugging Face)
The final session asks: how does a model think? By “opening the black box,” we explore methods to reverse-engineer the neurons and circuits of a Transformer. This understanding is crucial for verifying the internal logic of models used in high-stakes programming and mathematical proof tasks.

Informations pratiques

Date

15 au 19 juin 2026

Lieu

centre CNRS La vieille Perrotine du CNRS.

Inscription

Pour pouvoir participer, merci de remplir le formulaire d’inscription et l’envoyer avant le 15 mai 2026 à Régis Vizet et Tifenn Baril-Graffin.

Pré-requis

Contacts

Sécretariat des écoles
Régis Vizet – CEA
Tifenn Baril-Graffin – INRIA
tel: 01 69 26 47 45
Fax: 01 69 26 70 05

Coordinateurs de l’école d’informatique 2026
Nathanaël Fijalkow
Marc Lelarge
Philippe Suignard
Guillaume Baudart
Xavier Hinaut