Publications

2026

• Understanding Build Reproducibility in the F-Droid Ecosystem
  
  • Nanni Denise
  • Malka Julien
  • Zacchiroli Stefano
  • Zimmermann Théo
  • d'Angelo Gabriele
  , 2026. The security of open source applications benefits considerably from the possibility of rebuilding their source and verifying the output. F-Droid, a prominent distribution for open source Android applications, systematically rebuilds them from source and tests their bitwise reproducibility at app publishing time. However, F-Droid offers no guarantee that app reproducibility will continue to hold in the future. As software ecosystems evolve, reproducibility may degrade, with potential negative consequences for software preservation and security. We present the first empirical study of build reproducibility in the F-Droid app ecosystem. Analyzing historical reproducibility logs, we find that the overall bitwise reproducibility rate has been steadily increasing over time (as new versions of apps are published). We then evaluate how reproducibility holds in time for fixed app versions, by attempting to rebuild 18 904 app versions that F-Droid had previously confirmed bitwise reproducible, published between September 2018 and February 2026, achieving an 83% rebuild success rate, and identify missing dependencies as the dominant cause of failure, accounting for 76% of non-rebuildable cases. Among suc- cessfully rebuilt apps, 94% are also bitwise reproducible—i.e., they still yield bitwise identical artifacts upon rebuild. Together, these results show that while bitwise reproducibility largely holds for apps that can be rebuilt, rebuildability itself is highly sensitive to temporal decay. (10.1145/3820002.3828593)
  DOI : 10.1145/3820002.3828593
• Mutating the "Immutable": A Large-Scale Study of Git Tag Alterations
  
  • Rapaport Solal
  • Pautet Laurent
  • Tardieu Samuel
  • Zacchiroli Stefano
  • Zimmermann Théo
  , 2026. Git tags are commonly viewed as immutable references in software development, marking releases and specific repository states that underpin build reproducibility and software supply-chain integrity. Despite their intended immutability, Git allows tags to be altered through deletion or modification via force-pushed updates. The prevalence of such alterations threatens reproducible builds and dependency integrity. We conduct the first large-scale empirical study of tag alterations in public code repositories, analyzing 328.4 M software repositories from Software Heritage and identifying 10.2 M tag alterations affecting 189 k unique repositories. A cross-analysis with Nixpkgs reveals that 32 packages reference tags altered in our dataset, with 7 exhibiting confirmed build errors, providing concrete evidence that tag alterations break reproducible package builds. Our findings challenge the widespread assumption that tags are immutable anchors for released software. We therefore recommend that build systems and package managers pin dependencies to cryptographic commit hashes, that development forges expose tagmutation audit logs, and that the community adopt systematic monitoring of tag alterations as a standard supply-chain security practice. (10.1145/3820002.3828585)
  DOI : 10.1145/3820002.3828585
• NeuralSketch2Surf: Fast Neural Surfacing of Unoriented 3D Sketches
  
  • Ye Hongsheng
  • Sureshkumar Anandhu
  • Zhonghan Wang
  • Cani Marie-Paule
  • Hahmann Stefanie
  • Bonneau Georges-Pierre
  • Parakkat Amal Dev
  , 2026, pp.123:1-12. With recent advances in VR, 3D sketches have emerged as a powerful medium for 3D model creation. However, while they already provide the user with a good perception of the intended shape, they must be surfaced before any reuse in a downstream application. This remains a challenge when sketches are unoriented, i.e., when they are simply sets of unsorted 3D strokes, without any additional normal information. We introduce NeuralSketch2Surf, the first fast and robust neural surfacing solution that processes arbitrarily unoriented sketches at interactive rates. Our approach uses S2V-Net, a new transformer network designed to mesh 3D sketches. Instead of directly inferring complex functions to represent shapes, we focus on predicting an occupancy grid, then refined using a custom smoothing function to create the desired surface. Thanks to a lightweight architecture that enables fast predictions, our method produces results in less than 2 seconds, in contrast to SOTA techniques that can take minutes or even hours. Extensive evaluations demonstrate that our method is not only fast but also generates closed surfaces with high geometric, topological, and perceptual accuracy. (10.1145/3799902.3811227)
  DOI : 10.1145/3799902.3811227
• Structure-Guided Learning for Audio Analysis under Low Supervision
  
  • Torres Diniz Bernardo
  , 2026. For many audio analysis tasks, supervised learning is either inadequate or inefficient. Navigating a low-supervision landscape within modern data-driven frameworks requires carefully integrating expert domain knowledge to guide the learning process and constrain the set of possible solutions. In this dissertation, three contributions that provide structure to learning in the form of signal (or physical) priors or geometric constraints are presented. The first contribution builds on a geometric understanding of the Fourier frequency space through Optimal Transport. Building on this, a frequency-aware training objective for audio tasks, named Spectral Optimal Transport (SOT), is proposed. SOT is evaluated on synthetic data for harmonic parameter estimation using unsupervised analysis and synthesis, and compared to traditional spectral losses based on Lp norms. The second contribution presents an extensive evaluation of a pitch estimator trained using equivariant self-supervised learning across cross-dataset scenarios, along with practical improvements informed by domain knowledge. The improvements are shown to noticeably improve performance, enabling a lightweight model to match or exceed the performance of supervised baselines, while enabling real-time inference.The third and final part addresses the underexplored task of separating individual drum instruments from a drums-only mixture. The event-driven nature of percussive recordings, in which each sound arises from a distinct, discrete strike, is modeled as a prior within a synthesis model. This enables training a modular separation model that matches the performance of larger, fully supervised models while bypassing the need for separated tracks by relying instead on transcription annotations. Together, these three contributions demonstrate that providing signal and geometric structure to data-driven learning models is beneficial, particularly for musical audio analysis tasks under low supervision.
• ExposNet: A Deep Learning Framework for EMF Exposure Prediction in Complex Urban Environments
  
  • Zhang Yarui
  • Wang Shanshan
  • Wiart Joe
  IEEE Transactions on Machine Learning in Communications and Networking, Institute of Electrical and Electronics Engineers, 2026, pp.1-1. The prediction of the electric field (E-field) plays a crucial role in monitoring radiofrequency electromagnetic field (RF-EMF) exposure induced by cellular networks. In this paper, a deep learning framework is proposed to predict E-field levels in complex urban environments. First, the drive test measurement in Paris and Lyon and publicly accessible databases used to construct the training dataset are introduced, with a detailed explanation provided on how these datasets are formulated and integrated to enhance their suitability for Convolutional Neural Networks (CNNs)-based models. Then, the proposed model, ExposNet, which is a lightweight CNN-based framework, is presented. Two variants of the network structure are proposed, enabling per-frequency prediction and total E-field prediction. Extensive experimental analyses are conducted, including the comparison with a standard U-Net baseline and a classical Kriging baseline. Moreover, a detailed ablation study is conducted to quantify the contribution of each input component, and additional generalization experiments are performed on different test subsets. The overall results indicate that, despite being trained and tested on real-world measurements, the model performs well and achieves better accuracy compared to previous studies. (10.1109/TMLCN.2026.3714431)
  DOI : 10.1109/TMLCN.2026.3714431
• Indirect Core Halt via Trap Handling: Halting a RISC-V Core from User Mode Through Kernel-Mediated Fault Handling
  
  • Awais Muhammad
  • Mushtaq Maria
  • Naviner Lirida
  • Yahya Jawad Haj
  • Bruguier Florent
  , 2026. We show that a RISC-V CPU core can be persistently halted from user mode by deliberately triggering illegal CSR access instructions, requiring no kernel privilege or physical access. The fault propagates through the Linux S-mode trap handler to OpenSBI's unhandled-fault path, terminating in sbi_hart_hang(), an infinite wfi() loop independent of any debugger configuration. We validate this behaviour on a SiFive HiFive Premier P550 running Linux with OpenSBI v1.6, and demonstrate that user-triggered exceptions can cause system-wide denial of service, stressing the need for careful validation of exception-handling paths in RISC-V platforms.
• TRAPHALT: Indirect Core Halt via Trap Handling to Halt a RISC-V Core from User Mode Through Kernel-Mediated Fault Handling
  
  • Awais Muhammad
  • Mushtaq Maria
  • Naviner Lirida
  • Haj Jawad
  • Bruguier Florent
  , 2026. Modern processors rely on strict privilege separation to prevent user applications from manipulating privileged architectural state directly. In RISC-V systems, user-mode programs cannot access privileged control and status registers (CSRs) or execute privileged instructions. Still, exceptions raised by user programs are handled by the operating system kernel through the trap mechanism, which runs at a higher privilege level. This interaction between user-level faults and kernel trap handling can introduce subtle system behaviours that affect processor control state. In this work, we show that a RISC-V CPU core can be indirectly halted from user mode by deliberately triggering kernel trap handling with carefully crafted illegal instruction sequences. A user-space program forces the kernel to process a fault condition that propagates through OpenSBI's unhandled-fault path to sbi hart hang(), suspending the hart in an infinite wfi() loop. We demonstrate this behaviour on a SiFive HiFive Premier P550 platform running Linux with OpenSBI v1.6. Our results highlight how usertriggered exceptions can cause unexpected system-level effects and stress the need for careful validation of exception-handling paths in RISC-V platforms.
• Automatic Discovery of Disease Subgroups by Contrasting with Healthy Controls
  
  • Louiset Robin
  • Duchesnay Edouard
  • Dufumier Benoit
  • Grigis Antoine
  • Gori Pietro
  Data Mining and Knowledge Discovery, Springer, 2026, 40, pp.74 (1-30). n biomedical Subgroup Discovery, practitioners are interested in discovering interpretable and homogeneous subgroups within a group of patients. In this paper, assuming that healthy subjects (i.e., controls) share common but irrelevant factors of variation with the patients, we motivate and develop a Contrastive Subgroup Discovery method, entitled Deep UCSL. By contrasting patients with controls, Deep UCSL identifies subgroups driven solely by pathological factors, ignoring common variability shared with healthy subjects. Our framework employs a deep feature extractor to learn a discriminative representation space. Mathematically, we derive a novel loss based on the conditional joint likelihood of latent clusters and patient/control labels, optimized via an Expectation-Maximization strategy alternating between subgroup inference and feature encoder updates. A regularization term further encourages representations to capture disease-specific variability while ignoring variability shared with controls. Compared to previous related works, our approach quantitatively improves the quality of the estimated subgroups, as demonstrated on a MNIST example and four distinct real medical imaging datasets. (10.1007/s10618-026-01233-8)
  DOI : 10.1007/s10618-026-01233-8
• Green Energy Driven Integrated Smart Grid and Wireless Networks
  
  • Wang Li-Chun
  • De Swades
  • Balakrishnan Ashutosh
  , 2026, pp.XXIII, 140. Green energy and next-generation wireless systems are no longer independent domains, rather they are rapidly converging to redefine the future of sustainable connectivity. This book presents a forward looking study outlining the design of green energy driven, integrated smart grid and wireless networks. It reimagines the integration of renewable energy with the traditional power grid, by enabling each grid user to not only consume grid energy but to also be a potential energy source, thereby revising the conventional idea of power grids. A grid networked system of such distributed ambient powered nodes is hence envisioned to potentially act as a carbon-free energy producer system to the power grid, in addition to an energy prosumer system. Through its chapters, the book outlines and highlights the importance of sustainability in 6G networks. The authors present insights on the design constraints and challenges in system analysis. The study broadly pertains to analytical modeling of spatio-temporal randomness in cellular networks, presenting novel strategies to mitigate the effects of the dual randomness on network performance. The key ideas presented include joint load and energy balancing, optimum resource provisioning, and aerial offloading. The authors also demonstrate a wider perspective by extending the concept of energy balancing to energy aware residential networks. The book concludes by discussing the scope of integration of AI in wireless networks and motivate the need for green-AI aided future networks. Written for graduate level students in computer science and electrical engineering, as well as industry professionals deploying large-scale green service solutions, this book serves as both a foundational reference and a roadmap toward sustainable, scalable, and intelligent future communication networks. (10.1007/978-3-032-20509-4)
  DOI : 10.1007/978-3-032-20509-4
• Trust-Based Admission and Departure Control for Cooperative Platooning
  
  • Braiteh Emma
  • Bassi Francesca
  • Khatoun Rida
  , 2026, pp.2502-2507. <div><p>Reliable data exchange is essential for safety and coordination in Cooperative Intelligent Transport Systems (C-ITS), especially within vehicular platoons. However, this requires interaction with previously unknown vehicles, introducing security risks and requiring trust evaluation prior to sharing sensitive information. This paper presents a cooperative platooning protocol with a trust management framework based on Subjective Logic to assess the trustworthiness of vehicles based on the information they share. The framework processes heterogeneous evidence types to estimate individual trust levels at different stages. A fuzzification module classifies these trust values, enabling adaptive and interpretable trust-based decisionmaking. The resulting trust assessments are used to determine whether to accept vehicle requests to join or leave a platoon and to detect misbehavior among existing platoon members. The complete model was thoroughly evaluated using PLEXE and SUMO simulators in dynamic and realistic platooning scenarios. Empirical results demonstrate its effectiveness in identifying faulty or malicious vehicles, thereby significantly enhancing the reliability of cooperative driving decisions.</p></div> (10.1109/CoDIT70676.2026.11631269)
  DOI : 10.1109/CoDIT70676.2026.11631269
• Antennes et métamatériaux
  
  • Begaud Xavier
  , 2026. Les antennes sont des éléments clés des systèmes de télécommunications, des Radar ou pour l’internet des objets mais leurs performances en termes de bande de fréquences de fonctionnement, de gain et d’encombrement sont contraintes notamment par les limites physiques des matériaux conventionnels. Les métamatériaux sont apparus au début des années 2000 et ont suscité depuis de nombreux travaux. Il s’agit de structures artificielles ayant des propriétés électromagnétiques extraordinaires, telles qu’un indice de réfraction négatif ou une permittivité et une perméabilité contrôlées. Ces propriétés qui n’existent pas à l’état naturel offrent de nouveaux degrés de liberté pour la conception d’antennes et permettent de concevoir de nouveau systèmes antennaires capables d’améliorer le rayonnement, de le contrôler et de réduire son encombrement. L’étude des interactions entre l’antenne et le métamatériau représente un domaine de recherche dynamique associant la physique, l’électromagnétisme et l’ingénierie des systèmes de télécommunications. Les travaux sur les métamatériaux sont très nombreux et il est important de définir précisément les termes utilisés pour s’y retrouver et mieux se rendre compte de l’étendue des applications potentielles. Une fois la terminologie connue, les premiers travaux portant sur la permittivité et la perméabilité négatives rappelés vont démontrer comment il est possible de synthétiser ces matériaux sous les hypothèses importantes de modélisations et d’homogénéisation. D’autres métamatériaux comme les ligne composites main droite main gauche, les surfaces hautes impédances, les conducteurs magnétiques artificiels vont aussi être présentés et leurs caractéristiques essentielles permettant de les dimensionner seront détaillées. Cette première phase est essentielle pour utiliser correctement les métamatériaux avec les antennes. Dans de nombreuses applications, les antennes sont en très grande proximité des métamatériaux et elles interagissent nécessairement avec ces derniers. Les problématiques sont variées et adressent ainsi la réduction des dimensions en volume ou en épaisseur, l’augmentation de la directivité voire du gain, la formation ou le balayage de faisceau. L’idée est donc de donner au lecteur un certain nombre de méthodes pour dimensionner l’antenne avec le métamatériaux quand cela est possible, de préciser certaines limites ou plus simplement de définir l’intérêt de la solution à métamatériaux proposée. Le lecteur trouvera en fin d’article un glossaire des termes et des expressions importantes de l’article, ainsi qu’un tableau des notations et des symboles utilisés. (10.51257/a-v1-e3315)
  DOI : 10.51257/a-v1-e3315
• Statistical wave field theory: Random wave model of the complex Robin Laplacian
  
  • Badeau Roland
  , 2026. The statistical wave field theory mathematically establishes the statistical laws of the solutions to the wave equation in a bounded domain, which hold at high frequency. It provides the closed-form expressions of the power distribution and the correlations of the wave field jointly over time, frequency and space, in terms of the geometry and the specific admittance of the boundary surface. In this paper, the mathematical foundations of the theory are reworked at the fundamental level of the eigenfunctions of the Laplace operator. We thus introduce a random wave model of the complex Robin Laplacian, which encompasses and generalizes previous works dedicated to the real Robin Laplacian, and which accounts for the absorption of energy at the boundary surface. This model will allow us to establish the first and second order statistics of the room impulse response, and to finally achieve the unification of geometrical acoustics, which is generally used to predict the first echoes that appear in early reverberation, and of probabilistic models, which are generally used to predict the statistical properties of late reverberation.
• Geometrical acoustics meets the statistical wave field theory
  
  • Badeau Roland
  , 2026. The statistical wave field theory mathematically establishes the statistical laws of the solutions to the wave equation in a bounded domain, which hold at high frequency. It provides the closed-form expressions of the power distribution and the correlations of the wave field jointly over time, frequency and space, in terms of the geometry and the specific admittance of the boundary surface. In our previous article, the mathematical foundations of the theory have been reworked at the fundamental level of the eigenfunctions of the Laplace operator. We thus introduced a random wave model of the complex Robin Laplacian, which accounts for the absortion of energy at the boundary surface. In this article, based on this model, we establish the first and second order statistics of the room impulse response. The previous predictions of the theory are thus retrieved, and their accuracy is improved. In addition, this approach allows us to achieve the unification of geometrical acoustics, which is generally used to predict the first echoes that appear in early reverberation, and of probabilistic models, which are generally used to predict the statistical properties of late reverberation.
• Slice Collision-Induced Unfolding and Molecular Dynamics Reveal How Post-Translational Succination Reshapes SUMO1 Unfolding Pathways
  
  • Groignet Louis
  • Robert Thomas
  • Duez Quentin
  • Claessens Jehan
  • Brocorens Patrick
  • Gerbaux Pascal
  • de Winter Julien
  Journal of The American Society for Mass Spectrometry, American Chemical Society, 2026, 37 (8), pp.1872-1882. Abstract SUMO1 (small ubiquitin-like modifier 1) is a central feature of post-translational SUMOylation, modifying a broad range of substrate proteins. SUMO1 itself is prone to succination, i.e., a post-translational Michael addition of cysteine onto fumarate, which results in the formation of succinated SUMO1 with modified properties. The present study assesses the structural and gas-phase stability modifications in SUMO1 induced by diethyl fumarate succination using advanced ion mobility spectrometry-mass spectrometry (IMS-MS) techniques. Among them, collision-induced unfolding (CIU) and slice-CIU highlight a modified unfolding process resulting from the creation of specific charge-dipole interactions involving the appended dicarbonyl moiety. The experimental results are further supported by molecular dynamics simulations to understand, at the atomistic level, the mechanisms underlying the CIU of gaseous (derivatized) SUMO1 ions. (10.1021/jasms.6c00118)
  DOI : 10.1021/jasms.6c00118
• Learning High-Dimensional Parity Functions with Product Networks using Gradient Descent
  
  • Larue Guillaume
  • Dufrène Louis-Adrien
  • Lampin Quentin
  • Ghauch Hadi
  • Rekaya Ben Othman Ghaya
  , 2026. Parity functions are fundamental Boolean operations with critical applications across machine learning, cryptography, and error correction. Yet, learning high-dimensional parity functions poses significant challenges: in a general setting, standard neural network architectures typically require exponential sample complexity, making gradient-based optimization intractable for large number of inputs N. We demonstrate that compact product-based neural architectures combined with stochastic data sparsity (Bernoulli inputs with pe≤1/N) and appropriate hyperparameter choice enable efficient parity learning, with theoretical guarantees of convergence. Experiments validate our theory across dimensions up to N=100,000, with empirical evidence showing optimal hyperparameter choices for pe and learning rate α, as well as polynomial complexity scaling laws. This work establishes fundamental connections between architectural inductive bias and data sparsity, opening new possibilities for neural arithmetic, structured reasoning, binary neural networks, and machine learning applied to automated protocol discovery.
• Feature-aware (Hyper)graph Generation via Next-Scale Prediction
  
  • Gailhard Dorian
  • Tartaglione Enzo
  • Naviner Lirida
  • Giraldo Jhony H.
  , 2026, pp.306:32640-32674. Graph generative models perform well on small structured data but struggle to scale to large, complex structures. Hierarchical approaches improve scalability but often ignore node and edge features, which are critical in real-world applications, particularly for hypergraphs that model higher-order relationships. In this paper, we propose FAHNES (feature-aware (hyper)graph generation via next-scale prediction), a hierarchical framework that jointly generates topology and features for graphs and hypergraphs. FAHNES builds multi-scale representations through node coarsening and localized expansion, guided by a novel hierarchical scale encoding that controls granularity and ensures cross-scale consistency. Experiments on synthetic, 3D mesh, and graph point cloud datasets demonstrate competitive or state-of-the-art performance while uniquely scaling to featured large-scale graphs and hypergraphs. Our code is open source.
• TimeSAE: Causal Sparse Decoding for Faithful Explanations of Black-Box Time Series Models
  
  • Oublal Khalid
  • Bouniot Quentin
  • Gan Qi
  • Clémençon Stéphan
  • Akata Zeynep
  , 2026. As black box models and pretrained models gain traction in time series applications, understanding and explaining their predictions becomes increasingly vital, especially in high-stakes domains where interpretability and trust are essential. However, most of the existing methods involve only in-distribution explanation, and do not generalize outside the training support, which requires the learning capability of generalization. In this work, we aim to provide a framework to explain blackbox models for time series data through the dual lenses of Sparse Autoencoders (SAEs) and causality. We show that many current explanation methods are sensitive to distributional shifts, limiting their effectiveness in real-world scenarios. Building on the concept of Sparse Autoencoder, we introduce TimeSAE , a framework for black-box model explanation. We conduct extensive evaluations of TimeSAE on both synthetic and realworld time series datasets, comparing it to leading baselines. The results, supported by both quantitative metrics and qualitative insights, show that TimeSAE delivers more faithful and robust explanations.
• Spatiotemporal imputation with graph-informed flow matching
  
  • Zhang Zepeng
  • Einizade Aref
  • Giraldo Jhony H
  • Fink Olga
  , 2026, pp.306:155195-155211. Missing data is a common challenge in spatiotemporal systems, arising in applications such as air quality monitoring and urban traffic management. Traditional machine learning approaches, like recurrent and graph neural networks, rely on iterative propagation, which tends to accumulate errors over time and space. Recent diffusionbased methods mitigate error propagation but require iterative sampling and often depend on problem-agnostic Gaussian priors, limiting both efficiency and effectiveness. To address these limitations, we propose GiFlow, a Graph-Informed Flow Matching framework for spatiotemporal imputation. GiFlow replaces the typical Gaussian prior with a graph-informed prior constructed via spatiotemporal filtering of observable signals, which better aligns the source distribution to the target and thereby simplifies the generation trajectory. The flow field is parameterized by a hybrid vector field model that integrates spatial attention, temporal attention, and spatiotemporal propagation, enabling joint modeling of spatial and temporal dependencies. Extensive experiments on both synthetic and real-world datasets demonstrate that the proposed GiFlow outperforms the state-of-the-art approaches in spatiotemporal imputation. (10.48550/arXiv.2606.06682)
  DOI : 10.48550/arXiv.2606.06682
• SOTAlign: Semi-Supervised Alignment of Unimodal Vision and Language Models via Optimal Transport
  
  • Roschmann Simon
  • Krzakala Paul
  • Mazelet Sonia
  • Bouniot Quentin
  • Akata Zeynep
  , 2026. The Platonic Representation Hypothesis posits that neural networks trained on different modalities converge toward a shared statistical model of the world. Recent work exploits this convergence by aligning frozen pretrained vision and language models with lightweight alignment layers, but typically relies on contrastive losses and millions of paired samples. In this work, we ask whether meaningful alignment can be achieved with substantially less supervision. We introduce a semi-supervised setting in which pretrained unimodal encoders are aligned using a small number of image-text pairs together with large amounts of unpaired data. To address this challenge, we propose SOTAlign, a two-stage framework that first recovers a coarse shared geometry from limited paired data using a linear teacher, and then refines the alignment on unpaired samples via an optimal-transport-based divergence that transfers relational structure without overconstraining the target space. SOTAlign effectively leverages unpaired images and text, learning robust joint embeddings across datasets and encoder pairs, and significantly outperforming supervised and semi-supervised baselines.
• On the consistency of state machines, use cases and block diagrams using dependency graphs and Large Language Models
  
  • Sultan Bastien
  • Apvrille Ludovic
  • Coudert Sophie
  Software and Systems Modeling, Springer Verlag, 2026, pp.1-40. Abstract Model-Driven Engineering aims to support rigorous reasoning about system analysis and design by relying on multiple interconnected modeling views to represent complex systems. However, maintaining consistency across these heterogeneous views remains a major challenge, especially when capturing the subtle semantic and logical dependencies inherent to such systems. Traditional approaches to consistency rely on formal rule-based methods, sometimes complemented by ontologies. Yet, these techniques often fall short when dealing with deeper semantic issues that cannot be explicitly expressed as rules. This paper introduces a combined approach for the automated detection and correction of inconsistencies in multi-view SysML models. The proposed framework articulates three complementary techniques: (1) a Large Language Model (LLM)-based technique for inconsistency detection, (2) a dependency graph-based technique for detecting inconsistencies in a class of logical relationships between model elements and (3) a correction method relying on LLMs constrained by formal rules to enforce consistency constraints. The graph-based technique relies on transforming SysML design models into dependency graphs: The paper first proves that these graphs are in bijective correspondence with the original design models. It then formalizes consistency rules for three diagram types, presents the integrated framework and evaluates its implementation on representative SysML analysis and design diagrams. Overall, the practical illustration shows that the approach is effective for identifying and correcting inconsistencies, although the LLM-based component produces both false positives and false negatives. It also suggests that the dependency graph-based technique is a relevant complement to the LLM-based one for identifying inconsistencies in logical dependencies between model elements. (10.1007/s10270-026-01388-4)
  DOI : 10.1007/s10270-026-01388-4
• Tailoring Strictly Proper Scoring Rules for Downstream Tasks: An Application to Causal Inference
  
  • Plaud Roman
  • Perez-Lebel Alexandre
  • Saillenfest Antoine
  • Bonald Thomas
  • Le Morvan Marine
  • Varoquaux Gaël
  • Labeau Matthieu
  , 2026. Probabilistic models are typically trained using task-agnostic objectives like log-loss, which can lead to significant errors in downstream estimation. This disconnect is especially critical in Inverse Probability Weighting (IPW) for causal inference, where propensity score errors near 0 and 1 often lead to high bias and variance. We propose a principled framework for deriving task-specific strictly proper scoring rules by matching the local curvature of the downstream error metric. We apply this to the Average Treatment Effect (ATE) estimation, deriving a closed-form loss and its corresponding canonical probability mapping that can be readily integrated with any model like a neural network or a gradient boosting algorithm. Extensive evaluations on causal inference benchmarks demonstrate that our tailored objective consistently outperforms standard likelihood-based and covariate-balancing approaches.
• NLP-based fraud detection on a French public benefits platform : the technical, organisational and regulatory context
  
  • Breidenstein Alicia
  , 2026. An increasing number of public service platforms are being developed on behalf of the State to fulfill missions of general interest. However, as these platforms automate the distribution of public benefits, they become particularly vulnerable to fraud. This thesis focuses on Mon Compte Formation (MCF), France's public vocational training platform managed by the Caisse des Dépôts et Consignations (CDC). Fraud detection is a pressing concern for MCF due to the sheer volume of training offers and the impracticality of manual verification at scale.We explore Natural Language Processing (NLP) methods to identify textual anomalies within the MCF catalog. Our work assesses the influence of domain-specific pre-training, fine-tuning, and diverse textual representations on tasks such as similarity assessment, classification, and anomaly detection. Our results show that simple models with carefully selected embeddings can outperform more complex models on certain tasks, facilitating their industrial deployment. In more structurally complex scenarios, lexical representations prove most effective without fine-tuning, while BERT-based models excel once adapted to domain-specific data. Conversely, large language models (LLMs) underperform, constrained by the dataset's structural complexity and specific vocabulary.Beyond technical solutions, we examine the legal and organizational frameworks governing AI deployment at the CDC, particularly in relation to the EU AI Act and French regulatory requirements. Through semi-structured interviews, we analyze existing verification processes and the integration of AI in fraud prevention, highlighting persistent challenges in transparency, multi-stakeholder collaboration, and regulatory compliance. This interdisciplinary research, conducted in collaboration with the CDC, bridges technical NLP advancements and real-world public service challenges, providing actionable insights for algorithmic fraud detection in public administration
• Medium Identification Based on Correlation Analysis: A DAS Field Trial with Telecom Fibers
  
  • Luo Jiawei
  • Irurozki Ekhine
  • Pruvost Pierre
  • Jaouën Yves
  • Awwad Élie
  , 2026. Distributed acoustic sensing (DAS) enables infrastructure monitoring. However, DAS response strongly depends on cable deployment conditions. Based on field trial data, we propose medium-identification methods using cross-correlation, paving the way toward enhanced DAS data processing.
• Microarchitectural Analysis of Speculative Execution Patterns in RISC-V using Machine Learning and ISA-Level Masking Wrappers
  
  • Awais Muhammad
  • Mushtaq Maria
  • Naviner Lirida
  • Haj Jawad
  • Bruguier Florent
  , 2026, pp.1-7. Speculative execution attacks, such as Spectre Variant 1, leak sensitive data through transient microarchitectural effects. This paper presents a machine learning-based framework for analyzing and mitigating speculative execution vulnerability patterns in RISC-V systems. We extract branch, cache, and timing features from gem5 simulations and validate leakage behavior on a SiFive HiFive Premier P550 platform. Supervised models achieve up to 97.1% accuracy in distinguishing benign and speculative-leak execution windows, while unsupervised methods (K-means, HDBSCAN, and Isolation Forest) independently reveal structural and anomaly-based separation. Feature analysis shows that speculative loads and memory latency are the strongest leakage indicators. We further propose lightweight RISC-V assembly wrappers based on index masking and fence instructions. Results demonstrate that these wrappers shift vulnerable execution toward benign microarchitectural behavior. The proposed approach combines simulation, hardware validation, and learning-based analysis for practical speculative vulnerability analysis and mitigation in RISC-V processors. (10.1109/IOLTS69666.2026.11633845)
  DOI : 10.1109/IOLTS69666.2026.11633845
• Dual representations of classical and quantum entropies: Theory and perspectives
  
  • Rioul Olivier
  , 2026. The MAXENT (maximum entropy) principle gives a natural correspondance between (classical and quantum) statistical mechanics and information theory. Its general formulation leverages on log-concavity (or Klein's inequality) and the celebrated information inequality, which is equivalent to Gibbs' inequality. The resulting convex optimization allows one to relate entropy to the log-canonical partition function by Legendre-Fenchel duality. This can be further generalized to dual variational characterization of relative entropy or Kullback-Leibler divergence, relating Gibbs variationnal principle to Donsker-Vardhan variational representations. From an information theoretical point of view, the Gibbs principle and its associated evidence lower bound serves as the basis of modern machine learning algorithms, such as the expectation-maximization (EM) algorithm and its restriction to parameters output by a neural network. Interestingly, a large class of alternating optimization algorithms including EM and Blahut-Arimoto for computing channel capacity and source rate-distortion functions can be cast in a generic information geometric alternating projections on convex sets of measures. On the other hand, Donsker-Vardhan representations, when restricted to the output of a neural network, allows one to efficiently estimate information measures, and when applied to parametric estimators with quadratic risk, provide HCR-like and CRB-like statistical lower bounds. An alternative derivation on such lower bounds in the Bayesian setting makes use of the Weyl-Heisenberg uncertainty principle, which opens up new perspectives. In this keynote, I will select and review several of the aforementioned topics, and discuss the interactions between classical and quantum interpretations.