Abstract

Every time you see a product ad, a recommender system has picked a handful of items out of millions — in milliseconds. At Criteo, we build these systems for thousands of retailers worldwide.

Recommendation is increasingly borrowing from language modeling: products become “tokens”, user histories become “sentences”, and LLMs are entering the stack — as content encoders, as generative recommenders over Semantic IDs, and even as models that reason before recommending.

But what really helps in practice? Drawing on research and production experiments at Criteo, this talk walks through when content improve recommendations (and when it does not), what it takes for generative recommenders to compete with classic ID-based models, and whether “thinking before recommending” pays off.

We will close with a question every practitioner faces: how do you know a new model is better before running an expensive A/B test?

No prior recsys background needed.

Short Bio

Artem Betlei is a Senior Researcher at Criteo AI Lab, working on recommender systems and user modeling. He holds a PhD from Université Grenoble Alpes on causal machine learning (uplift modeling) and co-supervises PhD and intern research at Criteo. 

Abstract

Not so long ago, training a robot to manipulate objects required a lab budget, proprietary hardware and a team of specialists. LeRobot, Hugging Face's open-source robotics library, is changing that. With a low-cost arm, a couple of webcams and a PC with a good GPU, anyone can now record demonstrations by teleoperation, share datasets on the Hugging Face Hub, and train state-of-the-art imitation learning policies. 

This talk will be a gentle introduction to this ecosystem: how the hardware is set up, how demonstrations are collected and formatted, and which policies are available.

We'll then see how far the framework can be pushed with a concrete example: SpectRobot, our recent work : J. Rigal, E. Virot, C. Pascal, https://arxiv.org/abs/2609.24621 and https://spectrobot-project.github.io/ . 

By turning vibration signals from a single tactile sensor into spectrograms that a vision encoder can read, we trained a robot arm to shake an opaque box and sort it by its hidden contents. 

This is very much a story of learning by doing, and we hope sharing it will encourage others in the community to give it a try. 

Short Bio

Emmanuel Virot is an R&D engineer at Wormsensing. He graduated from École Centrale Paris (2011) and holds a PhD in mechanics from École Polytechnique, then continued his research at Harvard and later in startups. His work focuses on vibrations, biomechanics, haptics and signal processing, within both scientific and industrial projects. 

Presentation

Emmanuel kindly shared his presentation in case you missed his talk: An introduction to the fabulous world of LeRobot 

Abstract

Parametric body models provide compact and controllable representations of human shape and pose. They play a central role in computer vision, graphics, and robotics; but are often perceived as complex black boxes by ML practitioners.

This talk will demystify parametric body mesh models by introducing their core principles, including rigid transformations, skeletal animation, kinematics, skinning, mesh representations, and shape modeling. 

It will briefly compare model families, explain their intended use cases, and highlight practical applications in animation, motion capture, avatars, and human mesh recovery.

Presentation

Romain kindly shared his presentation in case you missed his talk: Human Parametric Body Mesh Models

Short Bio

Romain Bregier is a Senior Research Scientist at NAVER LABS Europe. His research focuses on 3D computer vision, human modeling, and robotics. He graduated from École Centrale de Lyon in 2013, and received a Ph.D. at Inria and Siléane for work on robotic manipulation in 2019. His recent work includes Anny, an open-source body model which can be used to recover human shape and pose from images.

Abstract

On the surface, little seems to have changed in how we post-train large language models since 2022. We still do some “mid-training” (arguably the worst misnomer in this process), followed by supervised fine-tuning, and finish with some form of reinforcement learning to align the model. Add a bit of context-length extension, and the recipe sounds familiar. Under the surface, however, everything has changed. 

Data curation has become far more complex. Training now depends on carefully designed environments, LLM-as-a-judge systems, and increasingly sophisticated sequencing of reinforcement learning algorithms, objectives, and rewards. What once looked like a straightforward pipeline has evolved into a highly engineered production system.

Two broad approaches have emerged to manage this complexity: the artisan and the model factory.

In this talk, Matthias will draw on his experience leading post-training teams at Cohere and Poolside to explore the state of post-training in 2026—what has changed, what has scaled, and where the field might be heading next.

Presentation

Matthias kindly shared his presentation in case you missed his talk:

Industrialization of LLMs 

Short Bio

Matthias Gallé has been based in Grenoble since 2011. After completing a PhD in theoretical computer science in Rennes, he joined Xerox Research and later Naver Labs, where he led all of  research until 2022 as well as co-leading the BigScience initiative (creating the BLOOM model). Since then, he has worked at LLM startups (Cohere and Poolside), focused on advancing model capabilities in various domains

Abstract

We hear a lot about the ethical and environmental challenges of AI—energy and water consumption, metal extraction, algorithmic bias, reliability issues, or misuses like deepfakes. As data scientists, we’re aware of these impacts. 

We often feel torn: on one hand, we believe we have little choice but to work with AI given its ubiquity; on the other, we strive to adopt responsible practices.

So, what are the concrete levers for prioritizing frugal AI? How can we, as practitioners, drive more responsible AI within our companies and across society? And what role can data science play in mitigating these impacts, rather than just observing them?

Presentation

Elise and Pierre-Loic kindly shared their presentation in case you missed her talk:

Apr 2026 - Frugal and resposible AI - Elise Pupier & Pierre-Loic Bayart 

Short Bio

Elise Pupier 

Alongside ten years of experience as a data analyst, Elise Pupier has long been committed to social and environmental causes. Elise volunteers with several organizations that seek to combine technology and the common good, such as Data for Good, Latitudes, and The Digital Collage (Fresque du numérique).

Whether as a volunteer, freelancer, or teacher, Elise explores how to use technology while considering its impacts and prioritizing projects that serve the public interest. 

Pierre-Loic Bayart

After seven years in nuclear engineering designing cooling systems, Pierre-Loic Bayart became fascinated by the rapid growth of the Python data science ecosystem and transitioned into AI and data science in 2016 as an independent trainer. Since September 2023, Pierre-Loic have been working as an AI educational research engineer at the Grenoble Interdisciplinary Institute for Artificial Intelligence (MIAI). Pierre-Loic 

also organize the Grenoble Python Meetup and volunteer with Data for Good and Latitudes to promote responsible and socially beneficial digital technology. 

Abstract

Recent advances in generative AI have dramatically increased the realism and accessibility of deepfakes. Deepfakes define images or videos that are altered or fully synthesized using AI. They can be misused for disinformation, identity theft, financial fraud, and other malicious activities. 

In this talk, Noemi will present an overview of different manipulation techniques as well as state-of-the-art methods for detecting deepfakes. 

Because this is a rapidly evolving research area, Noemi will also outline several research directions that will be explored during the PhD project of Charbel Yahchouchi taking place between Probayes and the INRIA Centre at Université Côte d'Azur. 

Short Bio

Noemi Roggero is Data Scientist at Probayes since 2023 and works most of her time on Machine Learning and Computer Vision use cases for La Poste. She moved to AI after a PhD on star formation, where she gained an expertise on signal processing and time series. 

She enjoys contributing to concrete, client-facing projects as well as conducting R&D. She is now co-advising a joint PhD thesis between Probayes and INRIA.

Presentation

Noemi kindly shared her presentation in case you missed her talk: 

Detecting Deepfakes in the Age of Generative AI

Short Bio

Julien Ros is CTO and co-founder of Volubile.ai, building voicebots powered by large language models. He holds a PhD in Computer Science and has led R&D projects in computer vision and AI at Orange Labs and Probayes. 

Julien scaled engineering teams at Criteo and QAPA, driving innovation in adtech and HR tech. He has over 20 years of experience in technical leadership and data-driven product development. Passionate about AI and the Web, he now focuses on transforming telephony through conversational AI.