As Postdoctoral Researcher, Adrien Deliège aims to develop « generative explicit rule induction and composition » to enable generative AI models to infer, compose, and apply human-interpretable rules for visual reasoning.
Adrien Deliège tells us about his career path and the research topic that led him to obtain a position as a Postdoctoral Researcher at the FRS-FNRS.
His background
I studied mathematics at the University of Liège (2008–2013), where I developed a strong foundation in mathematical analysis before pursuing a PhD in wavelet-based methods for geoscience signal processing (2013–2017, F.R.S.– FNRS fellow), under the supervision of Professor Samuel Nicolay. My doctoral work combined mathematical concepts with real-world data science, designing algorithms based on continuous and discrete wavelet transforms to extract multifractal and oscillatory structures from environmental and planetary datasets. This research, which earned multiple publications (Springer Nature, Physical Review Journals) and the first prize at the international “Ma Thèse en 180 secondes” contest, already showed my interest in making complex models interpretable and tied to tangible data, which is a perspective that continues to shape my work today.
From 2017 onward, my work evolved toward applied artificial intelligence and computer vision, first through the DeepSport project (ULiège, 2017–2021, coordinated by Professor Louis Wehenkel, with Professor Marc Van Droogenbroeck as partner, Call WALInnov), where I co-led efforts in deep learning for sports video understanding. This project yielded five Best Paper Awards at CVPR (Conference on Computer Vision and Pattern Recognition) workshops, led to the creation of the SoccerNet benchmark, and mobilized large-scale annotation campaigns and international collaborations with Aalborg University - Visual Analysis and Perception Laboratory (Denmark) and KAUST - Image and Video Understanding Lab (Saudi Arabia). I later directed research as Lead Machine Learning Engineer at Maddox AI (a German startup of Tübingen, Germany, 2022–2023), guiding a team of engineers in developing industrial computer vision systems for quality control. Since mid-2023, I have conducted research with the Centre de Sémiotique & Rhétorique (Faculty of Philosophy and Letters, ULiège), exploring how computer vision and data-driven methods can contribute to art history and visual semiotics. I studied how text-to-image models encode visual forms and dynamism of historical artworks, and how to operationalize art historians’ evaluation criteria to characterize stereotypes in AI-generated productions. In parallel, under the supervision of Professor Pierre Geurts, I have served as ULiège’s coordinating post-doc of the ARIAC project (Digital Wallonia 4 Artifical Intelligence) since 2025, which is the largest Belgian AI project to date, encompassing multiple universities and research centers located in Wallon Region.
Across these projects, I have published around 30 peer-reviewed papers, gathered 1000+ citations, co-authored nine funding proposals, obtained four industrial sponsorships for scientific challenges, and collaborated with several dozens of researchers worldwide, all while engaging in scientific communication, teaching, and thesis juries.
His research
My future research aims to study a new paradigm in generative AI, that I call Generative Explicit Rule Induction and Composition (GenERIC), which seeks to endow models with the ability to infer, represent, and systematically compose human-interpretable rules. While current models excel at pattern replication, they often lack compositional reasoning and internal structure. I want to bridge this gap by enabling models to discover “reasoning primitives” (e.g., geometric transformations, color manipulations, or temporal shifts) from data, and to organize them into higher-level generative rules. I will begin by investigating this framework in controlled reasoning settings such as the ARC-AGI benchmark, studying how to internalize rule execution within models rather than relying on exhaustive search. Then, I plan to extend GenERIC to real-world generative domains, including text-to-image synthesis and temporal sequence modeling, where rule-based reasoning can provide interpretability and compositional control. Beyond technical development, I intend to contribute to new benchmarks and evaluation methodologies that assess reasoning capacity rather than surface correctness, thus shaping how the AI community conceptualizes and measures progress toward better reasoning-capable generative systems.
