Résumé : After a brief introduction on how current LLMs are trained, we will discuss some limitations, in terms of compositional generalization, reasoning capacities and explainability, of modern LLMs in different in domains, such as causal reasoning and driving LLMs. Whenever possible, we’ll adopt in this discussion a larger perspective than the one solely associated with LLMs.

After a PhD on probabilistic models for bilingual lexicon extraction from parallel corpora, I joined the Xerox Research Centre Europe (XRCE) in 1996 to work on information retrieval. In 1999, I spent 6 months at PARC (Palo Alto Research Center) to develop hierarchical versions of Probabilistic Latent Semantic Indexing. I then led a research team on textual information access, and later became area manager of the group Learning and Content Analysis at XRCE, prior to joining the Université Grenoble Alpes and the Laboratoire d’Informatique de Grenoble (LIG) as a professor in September 2006. I was Director of LIG from 2016 to 2020 and Director of the Grenoble Interdisciplinary Institute in Artificial Intelligence from 2019 till 2025. Since 2023, I am a member of the Institute Universitaire de France.

Résumé : In the last 50 years, the amount and complexity of information stored in databases has grown and the needs of database users have evolved. These changes have led to the creation of new models such as the graph database.

In a graph database, data is stored as a “graph”, or a collection of nodes, representing the entities, connected together by edges, representing relationships between the entities. Additional information about the entities and their relationships can be attached to the nodes and edges. This powerful model is now a staple in various industries such as biology, social networks and banking. By putting the links between the data points front and center, graph databases allow users to reason not only about individual elements but also about the structure of the graph. Accordingly, the goal of a typical graph query is to find a “path” connecting specific nodes.
Because graph traversals inherently rely on transitivity, traditional query languages are not suitable in the graph context, and thus new languages have been created. In 2019, a new International Organisation for Standardization (ISO) group was created to oversee the standardization of practical graph languages. This led to the creation of two new standards: SQL/PGQ and GQL. While this standardization work is a step in the right direction, one crucial ingredient was missing: a corresponding theoretical model that would allow us to study these languages.

In this talk, I will present the models of GQL and SQL/PGQ that we have developed and show how they have allowed us to uncover flaws in the design of GQL and SQL/PGQ.

Résumé : Knowledge Graphs (KGs) in the Semantic Web have become central to numerous applications in artificial intelligence, semantic search, and data integration. Their collaborative, crowdsourced nature, represented by platforms like Wikidata, DBpedia, and YAGO, offers immense scale and coverage but also raises important questions about their structural dynamics and analytic utility. While much research has focused on ontological (or schema) modeling, less attention has been devoted to understanding how facts evolve and self-organize in the factual layer of these graphs. This work addresses this gap by modeling the dynamic behavior of relationships in KGs, discovering domain-specific ranking indicators, and assessing the robustness of resulting rankings under perturbations. Our investigation is guided by four research questions. First, we explore whether the distributed editing processes in large crowdsourced KGs result in stable global structures (RQ1). Second, we seek to understand how to model the knowledge accumulation process (RQ2). Building on these insights, we address the problem of automatically discovering meaningful and interpretable ranking indicators tailored to specific domains and use cases (RQ3). Finally, we assess the resilience of such rankings in the face of structural errors and vandalism (RQ4). To tackle these questions, we first introduce KRELM, a generative Knowledge Relationship Model that treats each KG relationship as a bipartite network governed by an asymmetric attachment process: subjects receive new facts uniformly, while popular objects attract new facts preferentially. This model explains the persistent emergence of exponential and power-law degree distributions in real-world KGs. We provide theoretical results for these convergence behaviors and empirically validate KRELM across four crowdsourced KGs and eight historical Wikidata snapshots. Our results demonstrate that the structural regularities observed in KGs are not coincidental but are rooted in reproducible generative mechanisms. Building on this foundation, we propose RIPM (Ranking-Indicator Pattern Miner), a scalable algorithm for the automatic extraction of domain-specific ranking indicators. RIPM identifies, filters, and favors relationship class pairs based on their statistical inequality (measured by the Gini coefficient) and coverage (the proportion of covered entities). We derive an efficient approximation for the Gini coefficient, enabling RIPM to operate within the constraints of public SPARQL endpoints. An experimental evaluation, including a user study with 19 participants, confirms the utility, diversity, and interpretability of the indicators discovered by RIPM. Finally, we examine the robustness of rankings under structural perturbations by proposing a probabilistic model of ranking stability. We formalize and quantify the likelihood that an entity’s rank changes under perturbations. Building upon KRELM’s growth dynamics, we establish theoretical thresholds for tolerable perturbations and empirically validate these results across multiple KGs. Overall, this work presents a framework that links the dynamic structure of KGs with knowledge graph analysis, in particular, ranking indicator mining and robustness analysis. By combining insights from complex networks and statistical modeling, it advances the theoretical understanding and practical exploitation of large-scale knowledge graphs.

Résumé : j’introduis tout d’abord une famille de problèmes et de paradoxes relevant du rééchantillonnage de données dans de nombreuses disciplines, et dont l’effet est particulièrement visible en géographie : l’erreur écologique (ecological fallacy), le problème de changement du support (COSP), le problème des unités surfaciques modifiables (MAUP), les paradoxes de Simpson et de Robinson. Après avoir montré leurs effets néfastes sur la cartographie multiscalaire et, en généralisant, sur toute restitution statistique agrégée, désagrégée ou recomposée, je présente et discute les solutions proposées dans la littérature pour réduire ou éviter les problèmes et paradoxes évoqués.

Biographie : J’ai soutenu une thèse en géographie quantitative en 1995 à Grenoble sur les méthodes d’induction statistique (arbres de décision et réseaux de neurones) appliquées à la géographie (prédiction de l’évolution de l’occupation du sol agricole). J’ai ensuite intégré le CNRS en 1996, passé mon HDR en 2010 et été détâché 3 ans au Laboratoire d’Informatique d’Avignon (LIA). J’ai (co)dirigé le GdR MAGIS en 2017, pour ensuite diriger le laboratoire ESPACE jusqu’en 2023, UMR multisite et multi-tutelles de l’Institut des Sciences Humaines et Sociales du sud de la France. Dès ma thèse, j’ai travaillé en interdisciplinarité profonde (géomatique, informatique, mathématique, écologie, ingénierie des transports) avec des outils de modélisation , m’intéressant à des thématiques impactant la société (mobilité, dérèglement climatique) et à des questions plus théoriques et des problèmes méthodologiques, comme ceux abordés dans cette présentation.