Publications

2026

• Exploring Representation Autoencoders for Text-to-Audio Generation
  
  • V. Miranda Bernardo
  • Malard Hugo
  • Richard Gaël
  • Fontaine Mathieu
  • Essid Slim
  , 2026. Representation autoencoders (RAE), created by pairing a frozen pretrained self-supervised learning (SSL) encoder with a trained decoder, have recently demonstrated state-of-the-art performance on image generation. Their use in the audio domain, however, remains limited: text-to-audio (TTA) models are typically trained in the latent space of dedicated autoencoders where the encoder and the decoder are trained jointly with reconstruction objectives. This paper presents a study on audio representation autoencoders (AuRA) and their applicability to general audio TTA. We train decoders for two pretrained SSL audio encoders and evaluate their reconstruction quality against autoencoders used by recent TTA systems. After obtaining decoders for the SSL representations, we also train rectified flow models in the SSL and baseline latent spaces to isolate the autoencoder's effect on generation. We find that generation in SSL spaces achieves comparable or superior performance to generation in the baseline spaces, despite differences in reconstruction performance. These results highlight that reconstruction quality alone is not a sufficient proxy for generative performance.
• Radiation Testing Method for Observation of Soft Errors in Smart Sensors
  
  • Minelli de Carvalho Matheus
  • Cheymol Benjamin
  • Naviner Lirida
  • Possamai Bastos Rodrigo
  , 2026, 16869, pp.162-175. This work introduces a testing method for observing radiation-induced soft errors in modern COTS sensors and their interface circuits, herein smart sensors, including an analysis framework targeting the machine learning core of smart sensors, here simulated to characterize its expected radiation-induced soft error behavior. We demonstrate the effectiveness of the method in a case-study involving accelerometers subjected to neutron radiation, with the observation of radiation-induced soft errors in all axes. (10.1007/978-3-032-35506-5_11)
  DOI : 10.1007/978-3-032-35506-5_11
• LMMSE-Based Phase Estimator for Fiber-Optic Coherent Spread-Spectrum FBG-Based Sensing Systems
  
  • Bai Yiyang
  • Awwad Élie
  • Wigger Michèle
  IEEE Photonics Journal, Institute of Electrical and Electronics Engineers (IEEE), 2026, 18 (5), pp.6802407 (1-7). Fiber Bragg grating (FBG)-based sensing enables localized strain and temperature monitoring by tracking the phase changes of the light reflected from FBGs inserted along a fiber. A major source of disturbance in such systems is laser phase noise, and its influence is typically mitigated by means of additional hardware. In this article, we propose to mitigate the influence of laser phase noise by means of digital signal processing. We propose a practical estimator for coherent single-polarization spread-spectrum FBG-based sensing systems, which is motivated by linearizing the complex received signal and approximating the loss function to formulate a minimum mean square error (MMSE) estimation problem. Moreover, a Cramér-Rao lower bound for this FBG-based estimation problem is presented. Our simulation results demonstrate that the performance of our proposed estimator lies close to the Cramér-Rao lower bound, thus indicating that our estimator effectively compensates for phase noise in FBG-based sensing systems. (10.1109/JPHOT.2026.3734836)
  DOI : 10.1109/JPHOT.2026.3734836
• GRA-Net: Performance Inference for Dynamic Service Function Chains on NUMA Servers
  
  • Tao Weichen
  • Wang Qizheng
  • Linguaglossa Leonardo
  , 2026. Modern cloud systems deploy Virtual Network Functions (VNFs) as service function chains (SFCs) on shared multicore servers. However, accurate chain-level performance monitoring remains difficult on multi-core Non-Uniform Memory Access (NUMA) servers: fine-grained monitoring can incur high overhead and perturb packet processing, while predictionbased approaches must cope with fluctuating traffic, multitenant interference, and the combinatorial deployment space of service chains. Although recent work shows that low-overhead hardware telemetry can infer network-function performance without modifying the data plane, existing predictors often abstract away deployment dynamics and struggle when resourcesharing relationships change. This paper presents a non-invasive and placement-aware SFC throughput inference framework that combines Performance Monitoring Unit counters with VNF-tocore placement metadata. We introduce the Graph Relational-Attention Network (GRA-Net), a graph model for NUMA systems that represents logical cores as nodes and hardware relations as typed edges, and a temporal extension for more accurate delayed refinement. To improve scalability, we further introduce a method that groups VNF-to-core mappings with equivalent hardware sharing patterns, reducing redundant measurements. Extensive experiments on a physical Network Function Virtualization testbed built on the Data Plane Development Kit evaluate dynamic traffic, diverse SFC structures, co-located chains, and unseen placements. On seen placements, both models outperform the evaluated multilayer perceptron baseline. Across the broader evaluation, GRA-Net achieves accuracy above 90% in most settings and generalizes to unseen placements.
• Rust Coreutils: Rebuilding Unix Foundations in a Modern Language
  
  • Ledru Sylvestre
  • Tardieu Samuel
  • Zacchiroli Stefano
  IEEE Software, Institute of Electrical and Electronics Engineers, 2026, pp.1-9. GNU core utilities (coreutils) is a crucial package in modern UNIX systems. It comprises around 100 fundamental commands-like ls, cp, and cat-which run every day on millions of computers. However, GNU coreutils is also legacy software, with its C codebase dating back to the early 1990s and arguably feature-complete. If one were to consider reimplementing this essential package, how would they do so effectively, and why? This paper recounts the development of Rust coreutils, a contemporary open source reimplementation of GNU coreutils in the Rust programming language, which has reached the status of a drop-in replacement for GNU coreutils, compatible with most Linux distributions. By comparing Rust coreutils with its ancestor, we offer insights into creating a reliable substitute for critical software and highlight how modern programming features can attract development interest in legacy packages. (10.1109/MS.2026.3733266)
  DOI : 10.1109/MS.2026.3733266
• A High-Speed Digital Delta-Sigma Modulator Based on Temporal Unrolling and Zero-Padding
  
  • Lima Evelyn
  • Mathieu Yves
  • Jabbour Chadi
  IEEE Transactions on Circuits and Systems II: Express Briefs, Institute of Electrical and Electronics Engineers, 2026, 73 (11), pp.1099 - 1103. This brief proposes a novel digital parallel delta-sigma modulator (DSM) architecture that combines temporal unrolling with zero-padding (ZP) to achieve high throughput while preserving continuous internal state evolution across parallel channels. Unlike conventional sliced or time-interleaved (TI) DSM, the proposed approach avoids state discontinuities and mitigates noise degradation associated with parallelization. Synthesis results in a 22-nm process demonstrate a maximum throughput of 108 GS/s for a single-stage implementation with 16 channels. When extended to a multi-stage configuration, the proposed architecture offers a very interesting trade-off with a signal-to-noise and distortion ratio (SNDR) only 0.9 dB lower than a classical second-order error feedback (EFB) DSM, while achieving a throughput of up to 85 GS/s, corresponding to a 7× speed improvement. (10.1109/TCSII.2026.3733779)
  DOI : 10.1109/TCSII.2026.3733779
• Temporal 2G–5G RF-EMF exposure assessment in ten European countries during one year
  
  • van Bladel Han
  • Veludo Adriana Fernandes
  • Loizeau Nicolas
  • Röösli Martin
  • Maule Milena
  • Vecsei Zsuzsanna
  • Molnár Orsolya
  • Vrijkotte Tanja
  • Polanska Kinga
  • Politański Piotr
  • Mamrot Paweł
  • Wang Shanshan
  • Wiart Joe
  • Grellier James
  • Kovalenko Anastasiia
  • Hulls Paige
  • de Vocht Frank
  • Vaupotiĉ Nina
  • Guxens Mònica
  • Joseph Wout
  Science of the Total Environment, Elsevier, 2026, 1047, pp.182037. Due to the rapid evolution and densification of mobile communication networks, there is an increased interest in the long-term assessment of environmental exposure to radiofrequency electromagnetic fields (RF-EMF). While numerous studies have investigated RF-EMF exposure using short-term or spatial measurement campaigns, the number of comprehensive analyses capturing temporal variability across multiple frequency bands and countries remain limited. This pilot study aims to establish a multi-country sensor network to collect data on long-term RF-EMF exposure, including 5G, and to investigate whether temporal trends occur. Twenty frequency-selective sensors were deployed at fixed indoor and outdoor locations in ten European countries, continuously measuring E-field (electric field) strengths in four mobile communication frequency bands (806 MHz, 942 MHz, 1842 MHz, and 3625 MHz) over a period of 17 months with a temporal resolution of 1 s. The raw measurement data were calibrated and analyzed to describe RF-EMF exposure levels, temporal patterns, and variability across environments and spatial characteristics. A strong and recurring diurnal pattern was observed across the four frequency bands. The highest day-night contrasts occurred in the 806 MHz and 1842 MHz bands, where nighttime values decreased by 35.1% and 48.4%, respectively compared to daytime values. Milder contrasts were observed between weekdays and weekends with the most pronounced decrease of 16.6% for the 1842 MHz band. Temporal variability, quantified using the R-factor (i.e. the ratio of the median field strength to the maximum field strength measured during the period of interest), varied substantially across frequency bands and locations, with lower R-factors observed for higher-frequency bands (with a median R-factor of 0.45 for 3625 GHz), indicating a greater variability. Given the scope and duration of this study, it aims to serve as a pilot study for long-term exposure monitoring at fixed sites in multiple countries. (10.1016/j.scitotenv.2026.182037)
  DOI : 10.1016/j.scitotenv.2026.182037
• Efficient Artificial Intelligence Based On Teacher-Student Paradigm : Understanding Low-Rank Learning and Improving Channel Decoding
  
  • Wang Xiaolin
  , 2026. This thesis addresses one question through the teacher--student paradigm: when can a small model do the work of a large one? Its central principle: the sample and parameter efficiency of a learned model is set by the approximation capacity of the task and by the route through which the student absorbs the teacher's structural prior---implicitly via gradient descent on a factored parameterization, or explicitly via a task-aware preprocessing layer.The implicit route is worked out for shallow networks under low-rank teachers. For the quadratic activation, a closed-form optimal student at any prescribed rank exposes the irreducible bias of low-rank learning, and correlated inputs soften the sample-complexity threshold into a two-regime scaling law set by the teacher's effective rank. For ReLU classification, a three-stage transition---random guessing, feature learning, rank-limited saturation---emerges with critical sample size linear in the input dimension and the shared effective rank. Capacity here reduces to the number of singular modes of the teacher.The explicit route is worked out for neural channel decoders of binary linear codes. A syndrome-shifted near-origin reduction lowers the affine-piece complexity of the bit-MAP boundary from exponential in the code dimension to polynomial in the code length, with the degree set by the minimum distance. A code-aware preprocessing layer then closes the capacity gap through closed-form parity-check extrinsic statistics, yielding compact decoders---from a sigmoid perceptron to a pre-norm transformer---that approach the BCJR decoder on short and medium-length BCH codes and a new state of the art on a long BCH code.
• Training-free Uncertainty Guidance for Complex Visual Tasks with MLLMs
  
  • Kim Sanghwan
  • Xiao Rui
  • Alaniz Stephan
  • Xian Yongqin
  • Akata Zeynep
  , 2026. Multimodal Large Language Models (MLLMs) often struggle with fine-grained perception, such as identifying small objects in high-resolution images or detecting key moments in long videos. Existing methods typically rely on complex, task-specific fine-tuning, which reduces generalizability and increases system complexity. In this work, we propose an effective, training-free framework that uses an MLLM's intrinsic uncertainty as proactive guidance. Our core insight is that a model's uncertainty decreases when provided with relevant visual information. We introduce a unified mechanism that scores candidate visual inputs by response uncertainty, enabling the model to autonomously focus on the most informative data. We apply this simple principle to three challenging visual tasks: Visual Search, Long Video Understanding, and Temporal Grounding, allowing off-the-shelf MLLMs to achieve performance competitive with specialized, fine-tuned systems. Our results demonstrate that leveraging intrinsic uncertainty is a powerful strategy for improving fine-grained multimodal performance.
• Statistical wave field theory: Main results
  
  • Badeau Roland
  , 2026. The statistical wave field theory mathematically establishes the statistical laws of the solutions to the wave equation in a bounded domain. It provides the closed-form expressions of the power distribution and the correlations of the wave field jointly over time, frequency and space, which hold at high frequency and after many reflections, in terms of the geometry and the specific admittance of the boundary surface. This paper summarizes the main results of the theory. (10.1121/10.0027914)
  DOI : 10.1121/10.0027914
• Look But Don't Touch with Sparse Autoencoders for Unlearning in Diffusion Models
  
  • Cassano Enrico
  • Renzulli Riccardo
  • Ahmed Rayyan
  • Grangetto Marco
  • Alaniz Stephan
  , 2026. Sparse autoencoders (SAEs) have recently been proposed as interpretable tools for concept-level manipulation, under the assumption that isolated features can serve as controllable intervention points. In this work, we systematically evaluate this assumption in the context of object erasure and steering in diffusion models. We show that while SAEs reliably detect and localize semantic concepts within diffusion model activations, direct intervention in their latent space frequently induces out-of-distribution activations, resulting in severe visual artifacts. To disentangle detection from intervention, we use SAE activations purely as semantic detectors to identify image regions containing the target object, and replace those patch embeddings with the ones that do not contain it. This detection-based replacement preserves the diffusion model's activation statistics and produces significantly cleaner erasure results than latent steering. Our findings reveal a fundamental gap between concept detection and concept intervention in diffusion models: monosemantic or sparse features are not inherently suitable as control knobs for steering. These results position SAEs as powerful interpretability tools for analyzing generative models, but highlight important limitations when used for direct manipulation, such as unlearning.
• Evaluating and Recalibrating Probabilistic Classifiers from One-Bit Threshold Queries
  
  • Plaud Roman
  • Saillenfest Antoine
  • Labeau Matthieu
  • Bonald Thomas
  • Waegeman Willem
  , 2026. Standard calibration evaluation and recalibration methods assume access to predicted probabilities. We study one-bit threshold access, a setting in which probabilities are hidden and the evaluator observes only whether the predicted probability exceeds a single chosen threshold per sample. Under this restriction, we show that classifiers can nevertheless be evaluated and recalibrated without reconstructing their hidden scores. We derive a consistent estimator of True Calibration Error, unbiased estimators of strictly proper scoring rules, and a one-bit objective for Temperature Scaling whose expectation matches the standard log-loss objective. Across these settings, we quantify the statistical cost incurred by replacing standard probability-access procedures by our one-bit estimators. Finally, we instantiate the framework for black-box LLMs, where the logit_bias API parameter allows us to realize the required threshold queries in zero-shot binary question answering.
• Self-Supervised Rotation-Invariant Representations for Unsupervised Motif Discovery in Molecular Dynamics
  
  • Morozov Vsevolod
  • Devijver Emilie
  • Laclau Charlotte
  • Krzakala Paul
  • Jakse Noel
  , 2026, 16948. Identifying local structures in 3D point clouds remains challenging for unsupervised methods, particularly in scientific settings such as molecular dynamics simulations of crystallization, where meaningful motifs correspond to subtle geometric arrangements of atoms. While recent advances in point-cloud representation learning focus primarily on object-level shape understanding, they are less suited to capturing fine-grained local geometric patterns under rotational symmetry. We propose a self-supervised framework for learning rotation-aware representations of atomic neighborhoods and clustering them into local structural motifs. Our method combines a rotation-equivariant encoder with a VICReg objective and a spatially consistent view construction that encourages smooth representations across overlapping neighborhoods. On both synthetic benchmarks and large-scale molecular dynamics simulations, the learned embeddings organize local environments into coherent and interpretable clusters that reveal spatially structured transitions between disordered and ordered atomic configurations. These results demonstrate the potential of self-supervised geometric learning for uncovering physically meaningful structure in large scientific point clouds.
• Towards Trust Calibration in Socially Interactive Agents: Investigating Gendered Multimodal Behaviors Generation with LLMs
  
  • Galland Lucie
  • Clavel Chloé
  • Ochs Magalie
  , 2026, pp.Article No.: 10, Pages 1 - 14. As Socially Interactive Agents (SIAs) become increasingly integrated into daily life, the ability to calibrate user trust to an agent's actual capabilities would help ensure appropriate usage of these agents. In this paper, we explore the capacity of Large Language Models (LLMs) to generate multimodal behaviors (verbal, vocal, gestural, and facial expression modalities) that reflect varying levels of ability and benevolence (A&B), two key dimensions of trustworthiness. We propose a novel method for automatically generating behaviors aligned with specific levels of these traits, a first step towards enabling nuanced and trust-calibrated interactions. By analyzing a large dataset of multimodal transcripts generated by LLMs, we demonstrate that GPT-5.4 is able to produce coherent behavior across different modalities (text, intonation, facial expression, and gesture). Using Random Forest feature importance analysis, we show that the generated behaviors align with theoretical expectations for (A&B). However, we also find that when gender is specified in the prompt, LLMs tend to reproduce societal gender stereotypes, associating male agents' behaviors with high ability and female agents' behaviors with high benevolence. To validate our approach, we conducted a user study on Prolific using a withinsubjects design. Participants perceived different levels of (A&B) in the generated behaviors align with the intended instructions. (10.1145/3806774.3827976)
  DOI : 10.1145/3806774.3827976
• LCoT-GV: Graph Attention Networks for Verifying Long Reasoning Chains in Large Language Models
  
  • Jaulmes Bérénice
  • Alam Mehwish
  , 2026. Large Reasoning Models produce Long Chains-of-Thought (LCoTs) which involve breaking down the problem into smaller reasoning steps before reaching the conclusion. However, these steps often contain contradictions, unsupported inferences, or irrelevant steps, even when the final answer is correct. We propose Long Chain-of-Thought Graph Verifier (LCoT-GV), a graph-based framework that represents LCoTs as reasoning graphs. Each node in the graph represents a reasoning step and the edges encode semantic and logical relations. A Graph Attention Network is then trained to predict chain-of-thought correctness from the reasoning graph. We construct a new graph-oriented verification dataset from multiple reasoning benchmarks for question answering in various domains. The results show that our method is competitive with the most similar approaches. (10.48550/arXiv.2608.30679)
  DOI : 10.48550/arXiv.2608.30679
• Two-dimensional Taxonomy for n-ary Knowledge Representation Learning Methods
  
  • Lu Xiaohua
  • Tupikina Liubov
  • Alam Mehwish
  IEEE Transactions on Knowledge and Data Engineering, Institute of Electrical and Electronics Engineers, 2026, pp.20. <div><p>Real-world knowledge can take various forms, including structured, semi-structured, and unstructured data. Among these, Knowledge Graphs (KGs) are structured representations that integrate heterogeneous data sources into structured representations. However, KGs typically reduce complex n-ary relations to simple triples, thereby losing higher-order relational details. In contrast, hypergraphs naturally represent n-ary relations with hyperedges that directly connect multiple entities. Recent advances have led to many hypergraph representation learning methods; however, they often overlook entity roles in hyperedges, limiting fine-grained semantic modelling. To address these issues, Knowledge Hypergraphs (KHGs) and Hyperrelational Knowledge Graphs (HKGs) combine the advantages of KGs and hypergraphs to better capture complex relational structures and role-specific semantics of real-world knowledge. This survey provides a structured review of representation learning methods for n-ary relational data, with a primary focus on static KHGs and HKGs. We propose a two-dimensional taxonomy: the first dimension categorises models based on their methodology, including translation-based models, tensor factorisation-based models, deep neural network-based models, logic rule-based models, and hyperedge expansion-based models. The second dimension classifies models according to their awareness of entity roles and positions in n-ary relations, dividing them into positionaware, role-aware, and aware-less approaches. Finally, we summarise benchmark datasets, training settings, negative sampling strategies, and benchmark-style performance comparisons, and outline open challenges to inspire future research.</p></div> (10.1109/TKDE.2026.3731554)
  DOI : 10.1109/TKDE.2026.3731554
• Scale-Aware Feature-Level Encoding for Historical Document Stamp Detection
  
  • Chelali Mohamed
  • Cloppet Florence
  • Kurtz Camille
  • Bloch Isabelle
  • Foliard Daniel
  , 2026, 16851, pp.195-212. Stamps are generally used to authenticate, validate, and officially identify a document. Their analysis is particularly complex, especially for historical documents, which can be damaged. This article focuses on the particular task of stamp detection from old photographs. These stamps allow historians to trace the provenance and distribution of photographs within historical press agencies and, therefore, convey important information. Their varied, sometimes complex shapes, their non-standardized spatial location on the document, and the presence of textual information at different scales make their detection very challenging. In addition, the print quality of the stamp on the paper also has an impact, as some stamps may be faded or even completely invisible. To cope with the multi-scale nature of the features to be detected and with the potential degradation of scanned documents, we introduce a Scale-Aware Positional Encoding (SAPE) module designed for vision Transformer-based architectures that rely on multi-scale feature maps. Our module encodes the normalized logarithmic stride as a continuous variable using two complementary components: (1) an interpolation within a learnable table of scale anchors, and (2) a sinusoidal projection capturing fine-grained variations. The resulting embedding provides a smooth and adaptive modulation of multi-resolution features, leading to better spatial coherence across feature-pyramid levels. Experimental results demonstrate that SAPE improves the DeepSolo baseline and achieves competitive performance against representative detection approaches. These results confirm the effectiveness of SAPE in handling high scale variability and degradation in archival scans. The code is available at: https://github.com/mchelali/SA-DETR. (10.1007/978-3-032-36207-0_12)
  DOI : 10.1007/978-3-032-36207-0_12
• Determinants of radiofrequency electromagnetic fields emitted by smartphones in French cellular telephony networks
  
  • Fontaine Maëlle
  • Guida Florence
  • Moissonnier Monika
  • Beranger Remi
  • Lagroye Isabelle
  • Orlacchio Rosa
  • Laplanche Alexia
  • Dejardin Olivier
  • Conil Emmanuelle
  • Huss Anke
  • Mazloum Taghrid
  • Wiart Joe
  • Danjou Aurélie
  • Schüz Joachim
  • Bories Serge
  • Deltour Isabelle
  Environmental Research, Elsevier, 2026, 305, Part 2, pp.125023-1:125023-11. Smartphones are a major source of radiofrequency electromagnetic field exposure, yet real-life determinants of uplink emissions remain poorly characterised. We conducted the first epidemiological study of smartphone cellular network uplink emissions in a general population sample, recruiting 167 volunteers across three French cities in 2022-2023. Emissions were measured quasi-continuously for a week during everyday activities using the novel DEVIN exposimeter, while the XMobiSensePlus application simultaneously recorded sensor values and smartphone activities in the background. Determinants of UL emission occurrence and level were analysed using logistic and linear regressions. Over 8001 h analysed, voice calls were recorded in 2.4% of time and data uploads in 81.3%. Voice calls were associated to emission occurrence (OR = 7.09, 95% CI: 5.96-8.43). In fully adjusted models, emission levels were associated to calls (+9.16 dBm, CI: 7.15; 11.16), and to Wi-Fi connection with substantially lower cellular emissions outside call periods (-15.37 dBm, CI: -17.28, -13.46). Cellular upload rates differentially increased emissions depending on Wi-Fi connection, while received signal quality differentially reduced emissions depending on call status. Legacy technologies (2G, 3G) were associated with higher emissions than 4G. Results varied across centers. Smartphone brand, operator, and Android version showed no independent association with emitted power after System-on-Chip adjustment. These findings demonstrate that real-life smartphone radiofrequency electromagnetic cellular emissions are shaped by a complex interplay of usage patterns, network conditions, and device characteristics. (10.1016/j.envres.2026.125023)
  DOI : 10.1016/j.envres.2026.125023
• Digital Predistortion Based on Advanced Voltage-Aware Memory-Polynomial Modeling for Massive MIMO Systems
  
  • Ghoniem Ahmed
  • Mohellebi Reda
  • Pham Dang-Kien Germain
  • Jabbour Chadi
  IEEE Transactions on Microwave Theory and Techniques, Institute of Electrical and Electronics Engineers, 2026, 74 (9), pp.7840-7853. (10.1109/TMTT.2026.3715377)
  DOI : 10.1109/TMTT.2026.3715377
• HELIOS: Lightweight real-time scheduling with energy readiness guarantees
  
  • Qiu Ruizhe
  • Robert Thomas
  • Tardieu Samuel
  • Pautet Laurent
  • Singhoff Frank
  Journal of Systems Architecture, Elsevier, 2026, 178, pp.1038856-1:1038856-10. Energy-harvesting embedded systems, like sensors and medical implants, must satisfy real-time constraints under strict energy limitations. Traditional schedulers, which prioritize timing over energy, often fail to produce valid schedules, particularly for non-preemptive task models. We introduce HELIOS, a lightweight, table-driven scheduler that treats timing and energy as joint first-class constraints while supporting limited-preemptive execution. HELIOS introduces energy readiness, ensuring jobs execute only when both temporal and energy conditions are met. The scheduling problem is formulated as an integer linear programming (ILP) problem, jointly optimizing task ordering, preemption points, and execution under energy constraints. By leveraging non-preemption and energy readiness, HELIOS maintains complexity proportional to job count. A key innovation is the use of worst-case energy footprints to guarantee safe execution via an energy readiness threshold. The table-driven runtime ensures compatibility with resource-constrained kernels. Experiments demonstrate that HELIOS achieves higher schedulability than priority-based baselines. (10.1016/j.sysarc.2026.103885)
  DOI : 10.1016/j.sysarc.2026.103885
• A posteriori closure of turbulence models: Are symmetries preserved?
  
  • Freitas André
  • Um Kiwon
  • Desbrun Mathieu
  • Buzzicotti Michele
  • Biferale Luca
  European Journal of Mechanics - B/Fluids, Elsevier, 2026, 119, pp.204496:1-204496:8. Turbulence modeling remains a longstanding challenge in fluid dynamics. Recent advances in data- driven methods have led to a surge of novel approaches aimed at addressing this problem. This work builds upon our recent work [Phys. Rev. Fluids 10, 044602 (2025)], where we introduced a new closure for a shell model of turbulence using an a posteriori (or solver-in-the-loop) approach. Unlike most deep learning-based models, our method explicitly incorporates physical equations into the neural network framework, ensuring that the closure remains constrained by the underlying physics benefiting from enhanced stability and generalizability. In this paper, we further analyze the learned closure, probing its capabilities and limitations. In particular, we look at joint probability density functions between resolved and unresolved variables, as well as the scale invariance of multipliers (ratios between adjacent shells) within the inertial range. Although our model excels in reproducing high-order statistical moments, it breaks this known symmetry near the cutoff, indicating a fundamental limitation. We discuss the implications of these findings for subgrid-scale modeling in 3D turbulence and outline directions for future research. (10.1016/j.euromechflu.2026.204496)
  DOI : 10.1016/j.euromechflu.2026.204496
• The role of Mrs. Gerber’s Lemma for evaluating the information leakage of secret sharing schemes
  
  • Rioul Olivier
  • Béguinot Julien
  , 2026. (10.1007/978-3-032-13992-4_5)
  DOI : 10.1007/978-3-032-13992-4_5
• Structural Bias Beyond Homophily: A Study of Fairness in Link Prediction
  
  • Marey Lilian
  • Perez Mathilde
  • Viard Tiphaine
  • Laclau Charlotte
  Transactions on Machine Learning Research, [Amherst Massachusetts]: OpenReview.net, 2022, 2026, pp.2835-8856. Graph link prediction (LP) plays a critical role in socially impactful applications such as job recommendation and friendship formation, making fairness an essential concern in this task. While many fairness-aware methods manipulate graph structures to mitigate prediction disparities, the topological biases inherent to social graphs remain poorly understood and are consistently conflated with homophily alone. In this work, we study the relationship between structural biases and fairness outcomes in LP. To this end, we formalize a taxonomy of topological bias measures and introduce a graph generation method producing a diverse corpus of synthetic graphs with controlled structural properties. Using this corpus, we show empirically that fairness outcomes are strongly correlated with graph topology, and that current fairness-aware methods remain sensitive to structural biases beyond homophily. These findings highlight the need for structurally grounded evaluations in fair graph learning.
• Towards Understanding Low-Rank Learning for Classification
  
  • Wang Xiaolin
  • Rioul Olivier
  • Mokraoui Anissa
  • Duhamel Pierre
  , 2026. <div><p>Understanding the sampling complexity of a neural network-the number of training examples it needs to generalize-is a fundamentally unresolved question, especially for modern architectures that use low-rank constraints for compression and fine-tuning. Although low-rank methods such as LoRA are widely adopted in practice, their effect on sampling complexity for classification remains poorly understood. In this work, we investigate shallow ReLU networks with low-rank constrained weights in a teacher-student context. It is shown that the inherent structure of classification tasks is a prerequisite for successful low-rank learning. A distinctive three-stage sample complexity curve is observed : (i) initial phase of random guessing; (ii) abrupt transition to rapid improvement once a critical sample size is reached; and (iii) final plateau whose height depends on the network's rank. It is shown that the critical sample size is determined by the effective task structure rather than by the raw dimension of the input data, and that the rank constraint limits the maximum achievable accuracy. Simulations on synthetic data validate our predictions, revealing a rigorous trade-off between compression and accuracy. This can guide rank selection in low-rank learning and constitutes a crucial step toward understanding low-rank training in more complex architectures and real-world contexts.</p></div>
• SCALMU: Synthetically-trained Unrolled Coupled Nonnegative Matrix Factorization for Hyperspectral-Multispectral Data Fusion
  
  • Xu Xinxin
  • Gousseau Yann
  • Kervazo Christophe
  • Saïd Ladjal
  , 2026. HyperSpectral-MultiSpectral Image (HSI-MSI) fusion enables high-resolution hyperspectral imaging by combining the rich spectral information of low-spatial-resolution hyperspectral images with the detailed spatial structure of multispectral images. Classical methods such as Coupled Nonnegative Matrix Factorization (CNMF) benefit from a strong physical interpretability but suffer from inferior results compared to their deep-learning counterparts. To address this limitation, we propose SCALMU (Synthetically-trained Coupling of Adaptive Learned Multiplicative Updates), a novel unrolled neural network architecture that integrates adaptive learnable matrices within the classical framework of CNMF multiplicative updates improving its results. Due to its architectural proximity with CNMF, the resulting algorithm preserves physical interpretability and nonnegativity constraints. To overcome data scarcity, we additionally generate a synthetic HSI-MSI dataset via the dead leaves model, enabling ground-truth supervision under realistic spatialspectral degradations. SCALMU is then trained end-to-end on this dataset. Experiments demonstrate SCALMU's superiority over state-of-the-art methods on the Urban dataset. The code will be made available at https://github.com/xinxinxu99/SCALMU.git