Programme

Hybridization between data-driven and knowledge-based models is increasingly common in the scientific landscape. This presentation introduces ongoing work toward the systematization of such model hybridization, with an emphasis on its reliability. This work is primarily conducted in the context of the Catalyst project (a focused project of the Engineering Digital Twins PEPR). For more information, visit https://edtlab.fr/en/focused-projects/fp1.

Large Language Models (LLMs) are increasingly being applied to scientific domains, and the verification of the output of these models becomes a critical problem. Theorem provers like Rocq, Lean, or Isabelle emerge as natural tools to ground LLMs for rigorous mathematical reasoning. This potential has not been lost on academic and industry players, triggering in the past few months a race toward stronger models for mathematics with impressive achievements. In this talk, I will present our ongoing work on building LLM-based tools to improve model capabilities for formal reasoning with the Rocq prover, and conversely, how LLMs can be leveraged to improve existing formal mathematics projects. I will then discuss the problem of auditing and trusting formalizations at AI scale.

Software Heritage is the largest public archive of code and its development history. It has archived 30 billion unique source code files and 6 billion commits from more than 400 million projects hosted on thousands of software development platforms. Thanks to its breadth and data model, Software Heritage provides unique opportunities to conduct large-scale studies in fields of software science such as cybersecurity, software engineering, artificial intelligence, innovation economics, and more. In this talk, we will briefly present the Software Heritage initiative and review some of the recent research results obtained by leveraging it.

Generative AI is pushing hardware beyond the GEMM engines that defined the first generation of AI accelerators. This talk presents our ongoing work on open-source RISC-V heterogeneous architectures, centered on three key scaling ideas: improving performance beyond GEMM with tightly-coupled heterogeneous clusters; scaling up capability with mesh-of-tiles systems; and exploiting aggressive architectural heterogeneity to build a strong foundation for open-source European infrastructure for low-power and high-performance edge GenAI

This presentation provides an overview of three aspects of software centric integrated digital system design. The first one concerns system integration within an heterogeneous computing infrastructure that includes CPUs and GPUs connected by a 3.5D logical network implemented using interconnected 3D chiplets. The second one examines hardware acceleration of AI using a highly optimized data-flow network generator, though limited to small networks. The third one focuses on the processors at the heart of these systems, seeking to push the performance limits of in-order and out-of-order general purpose processors.

Hardware-software co-design for low-level systems (compilers, runtimes, OS) raises a core question: what should be handled by the programmer, compiler, or hardware? This talk explores the trade-offs (costs vs. gains in efficiency, security, and sustainability) while acknowledging the challenges, from the lens of compiler expertise, offering insights into how responsibilities can be divided effectively.

LINAGORA, a European open source software provider founded in 2000, has become an open source AI model builder. Through the OpenLLM France consortium, we develop the Lucie and Luciole families of small language models (1B, 8B and 23B parameters, plus a speech-text variant), released as fully open artefacts: training data, intermediate checkpoints, code and weights.
This talk first revisits the genesis of that choice: why small, specialized models that can be deployed next to sensitive systems and easily customized matter more to public and regulated organizations than another race toward frontier scale.
It then discusses what training and serving such models on European infrastructures actually involves, drawing on our experience on the Jean Zay supercomputer with the NVIDIA Nemotron stack, including a hybrid Mamba-Transformer 8B model: access to compute, energy, heterogeneous hardware, and the gap between research clusters and production inference.
Finally, it opens on a shift that reshapes the economics of AI infrastructure: today’s token prices do not reflect the true cost of frontier inference, while agentic usage multiplies consumption by orders of magnitude. When that gap closes, matching each task to the right model, through dynamic query routing across models of different sizes combined with agent harnesses, will become a key architectural layer, and a research agenda where open models and European infrastructures can genuinely compete

The convergence of AI and networking is reshaping how intelligent applications are deployed across device, edge, and cloud infrastructures. This talk presents important aspects related to the sustainable orchestration of AI as a Service (AIaaS), organized around two complementary directions: Network for AI and AI for Networks. On the Network for AI side, we address dynamic inference and distributed orchestration of DNNs across the mobile-edge-cloud continuum, introducing graph-based methods (Feasible Inference Graphs) and quantile-constrained reinforcement learning to achieve reliable, energy-efficient deployment—reducing energy consumption while meeting latency and accuracy constraints. On the AI for Networks side, we leverage graph neural networks and quantum-accelerated algorithms for scalable QoS prediction (delay, jitter, loss), and develop DRL- and LLM-RAG-based agentic frameworks for 6G/5G service orchestration, improving profit and acceptance rates. Overall, this talk presents the aspects needed to make AIaaS sustainable at scale, and discusses how intelligent orchestration, spanning inference scheduling to service provisioning, can enable reliable and energy-efficient provisioning of intelligent applications in next-generation networks.

L’inférence des grands modèles mobilise des infrastructures de calcul considérables, nécessitant souvent plusieurs GPU pour stocker et exécuter un seul modèle, tout en exploitant parfois difficilement toute la puissance disponible. Nous verrons comment les différentes phases de l’inférence, la gestion de la mémoire et l’organisation des requêtes influencent cette efficacité et comment le parallélisme et l’ordonnancement des calculs permettent de mieux utiliser les ressources. Nous nous appuierons sur les recherches menées dans le cadre du défi CUPSELI qui vise à étendre ces approches à des infrastructures plus distribuées et hétérogènes et faire face à leur volatilité

Some commercial LLMs are served behind opaque APIs, where providers can silently swap, quantize, or fine-tune the model answering your queries with no notification to users. This talk presents recent contributions on auditing these black boxes from the outside. We first motivate the problem with evidence of model substitution in the wild, then present two complementary change-detection methods: a grey-box statistical test on returned log-probabilities, and a token-efficient black-box approach that probes carefully chosen “border” inputs. We then turn to instance-level fingerprinting, showing how NIST randomness tests over pseudo-random sequences (FLIPS) can identify which specific model instance is running. I will then conclude on the link between these two primitives; these results chart both practical tools for holding LLM APIs accountable.

Limites Numériques is an interdisciplinary research collective at the intersection of Design and Computer Science, and largely influenced by Science and Technology Studies. The presentation will outline our main research directions on sustainable software design, digital practices, and the design of digital alternatives. We will then outline how we think of our research and its impacts in relation to society, how this shapes how we conduct it, and how we communicate about it. Biography: Aurélien Tabard is an Assistant Professor specialized in Human Computer Interaction. With Nolwenn Maudet and Thomas Thibault, they started Limites Numériques in 2021. He co-leads the projects Congrats from PEPR eNSEMBLE and EcoCommons from PEPR NumEco, and the sical team at LIRIS

La curiosité est l’un des moteurs de l’apprentissage chez l’humain, et joue un rôle très important dans l’innovation au niveau collectif. Cette présentation illustrera comment l’équipe-projet Flowers AI & CogSci étudie les mécanismes de la curiosité et leur rôle dans les apprentissages chez les humains et les machines.