Publications

2026

• Optimizing Coexistence of Quantum Continuous Variable and Classical Signals in the C-band
  
  • Sharma Tanya
  • Purkayastha Ambashri
  • Schiavon Matteo
  • Alvarez Velasquez Juan Rafael
  • Lourdiane Mounia
  • Ware Cédric
  • Fabre Nicolas
  , 2026, PC14078-65. Integrating Quantum Key Distribution (QKD) with existing wavelength division multiplexing (WDM) networks is hindered by noise from high-power classical channels, particularly spontaneous Raman scattering, which cannot be filtered effectively. Continuous-variable QKD (CV-QKD) employs coherent detection with narrow spectral filtering, offering higher tolerance to in-band noise compared to discrete-variable QKD, which cannot distinguish signal from noise photons. Our numerical simulations quantifies the impact of launch power, channel spacing, and bandwidth on CV-QKD performance in C-band WDM systems, establishing a framework for predicting secure key rates and demonstrating the practical coexistence of quantum and classical channels. (10.1117/12.3106320)
  DOI : 10.1117/12.3106320
• Direction-aware Ambisonic Speech Separation Using Edge-Conditioned Graph Convolution
  
  • Kelley Liam
  • Fontaine Mathieu
  • Peeters Geoffroy
  , 2026. The conventional Direction-Aware Ambisonic Speech Separation (DAASS) paradigm involves spatial filtering followed by post-filtering refinement. A number of approaches have been proposed to jointly perform spatial filtering and refinement using end-to-end neural methods, using as input full ambisonic signals and spatially filtered signals without explicitly modeling their spatial relationships. In this paper, we propose to unify these different input types directly in the model architecture with a spatial relationship graph. In particular, we design a novel edge-conditioned graph convolution block which aims to replace standard 1D convolutional blocks and be better suited to ambisonic signals. For our experiments, we develop a new synthetic DAASS dataset called Spatial LibriMix which will be available online upon acceptance. Our method matches or outperforms baseline performance with a significantly reduced parameter count (66M to 7.5M).
• Synthetic Abundance Maps for Unsupervised Super-Resolution of Hyperspectral Remote Sensing Images
  
  • Xu Xinxin
  • Gousseau Yann
  • Kervazo Christophe
  • Ladjal Saïd
  IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, IEEE, 2026. Hyperspectral single image super-resolution (HS-SISR) aims to enhance the spatial resolution of hyperspectral images to fully exploit their spectral information. While considerable progress has been made in this field, most existing methods are supervised and require ground truth data for training—data that is often unavailable in practice. To overcome this limitation, we propose a novel unsupervised training framework for HS-SISR, based on synthetic abundance data, where no high-resolution ground-truth reference is required for training. The approach begins by unmixing the hyperspectral image into endmembers and abundances. A neural network is then trained to perform abundance super-resolution using synthetic abundances only. These synthetic abundance maps are generated from a dead leaves model whose characteristics are inherited from the low-resolution image to be super-resolved and from the known point spread function (PSF) of the hyperspectral sensor. This trained network is subsequently used to enhance the spatial resolution of the original image's abundances, and the final super-resolution hyperspectral image is reconstructed by combining them with the endmembers. Experimental results demonstrate both the training value of the synthetic data and the effectiveness of the proposed method across 3 datasets, 3 scaling factors, and several evaluation metrics. The code is available at https://github.com/xinxinxu99/SISR-DL.git
• Structure preserving adversarial diffusion for unpaired medical image synthesis
  
  • Paniagua Kévin
  • Conze Pierre-Henri
  • Jaouen Vincent
  • Angelini Elsa
  , 2026, pp.1-5. Recent works have established diffusion models as viable candidates for unpaired medical image translation, by incorporating supervision through non-diffusive models in an end-to-end fashion.. Existing models typically rely on cycleconsistent, two-sided, non-diffusive modules of relatively high computational cost where structure preservation is not explicit. In this work, we propose SPAD, a novel one-sided approach that explicitly preserves the source domain firstorder (gradient) orientations, while enabling flexible contrast adaptation between source and target domains. Our approach provides direct edge guidance for the diffusive module, leading to better structural preservation. Experiments performed on image synthesis tasks using BraTS, CERMEP, and Gold Atlas datasets demonstrate that our model improved translation fidelity and reduced training time by over 50% compared to SynDiff while keeping the same diffusion architecture. The reduction in computational cost allowed us to train 3D variants, which achieved consistently higher quantitative scores. Code will be publicly shared on GitHub upon acceptance. (10.1109/ISBI61048.2026.11515588)
  DOI : 10.1109/ISBI61048.2026.11515588
• Leveraging Whole Slide Difficulty in Multiple Instance Learning to Improve Prostate Cancer Grading
  
  • Arrivat Marie
  • Peyret Rémy
  • Angelini Elsa
  • Gori Pietro
  , 2026, pp.1-4. Multiple Instance Learning (MIL) has been widely applied in histopathology to classify Whole Slide Images (WSIs) with slide-level diagnoses. While the ground truth is established by expert pathologists, the slides can be difficult to diagnose for non-experts and lead to disagreements between the annotators. In this paper, we introduce the notion of Whole Slide Difficulty (WSD), based on the disagreement between an expert and a non-expert pathologist. We propose two different methods to leverage WSD, a multi-task approach and a weighted classification loss approach, and we apply them to Gleason grading of prostate cancer slides. Results show that integrating WSD during training consistently improves the classification performance across different feature encoders and MIL methods, particularly for higher Gleason grades (i.e., worse diagnosis). (10.1109/ISBI61048.2026.11515564)
  DOI : 10.1109/ISBI61048.2026.11515564
• Adaptive gradient domain normalization for one-sided unsupervised medical image synthesis
  
  • Paniagua Kévin
  • Conze Pierre-Henri
  • Jaouen Vincent
  • Angelini Elsa
  , 2026.
• Measuring and Aligning Abstraction in Vision-Language Models with Medical Taxonomies
  
  • Schaper Ben
  • Di Folco Maxime
  • Kainz Bernhard
  • Schnabel Julia A
  • Bercea Cosmin I
  , 2026, pp.1-4. <div><p>Vision-Language Models (VLMs) show strong zero-shot performance for chest X-ray classification, but standard flat metrics fail to distinguish between clinically minor and severe errors. This work investigates how to quantify and mitigate abstraction errors by leveraging medical taxonomies. We benchmark several state-of-the-art VLMs using hierarchical metrics and introduce Catastrophic Abstraction Errors to capture cross-branch mistakes. Our results reveal substantial misalignment of VLMs with clinical taxonomies despite high flat performance. To address this, we propose risk-constrained thresholding and taxonomy-aware fine-tuning with radial embeddings, which reduce severe abstraction errors to below 2% while maintaining competitive performance. These findings highlight the importance of hierarchical evaluation and representation-level alignment for safer and more clinically meaningful deployment of VLMs.</p></div> (10.1109/ISBI61048.2026.11515430)
  DOI : 10.1109/ISBI61048.2026.11515430
• COVARIANCE DESCRIPTORS MEET GENERAL VISION ENCODERS: RIEMANNIAN DEEP LEARNING FOR MEDICAL IMAGE CLASSIFICATION
  
  • Mayr Josef
  • Reithmeir Anna
  • Di Folco Maxime
  • Schnabel Julia
  , 2025, pp.1-5. <div><p>Covariance descriptors capture second-order statistics of image features. They have shown strong performance in general computer vision tasks, but remain underexplored in medical imaging. We investigate their effectiveness for both conventional and learning-based medical image classification, with a particular focus on SPDNet, a classification network specifically designed for symmetric positive definite (SPD) matrices. We propose constructing covariance descriptors from features extracted by pre-trained general vision encoders (GVEs) and comparing them with handcrafted descriptors. Two GVEs -DINOv2 and MedSAM -are evaluated across eleven binary and multi-class datasets from the MedMNSIT benchmark. Our results show that covariance descriptors derived from GVE features consistently outperform those derived from handcrafted features. Moreover, SPDNet yields superior performance to state-of-the-art methods when combined with DINOv2 features. Our findings highlight the potential of combining covariance descriptors with powerful pretrained vision encoders for medical image analysis.</p></div> (10.1109/ISBI61048.2026.11515496)
  DOI : 10.1109/ISBI61048.2026.11515496
• Fast tomographic SAR inversion by unrolling a greedy algorithm for sparse reconstruction
  
  • Mendes Cristiano Ulondu
  • Denis Loïc
  • Kervazo Christophe
  • Tupin Florence
  , 2026. <div><p>As a range-based imaging technique, synthetic aperture radar (SAR) is agnostic to targets' elevation. In urban areas, tomographic synthetic aperture radar (TomoSAR) retrieves the position of sparsely distributed targets along the elevation direction thanks to the phase information from multiple acquisitions of the same area. Compressive sensing (CS)-based methods have received a lot of attention during the last decade, producing satisfactory reconstructions from a limited number of acquisitions. To improve their performance and scalability for pixel-wise TomoSAR inversion, data-driven methods based on the deep unrolling of basis pursuit algorithms have recently been proposed. However, they depend on a discretization of the elevation direction, often leading to artifacts which have to be corrected in a post-processing step. In this paper, we present an approach inspired by the iterations of a greedy algorithm for sparse reconstruction that estimates continuous elevation positions of targets together with well-calibrated uncertainty predictions. Our approach achieves performances comparable to reference CS-based methods while offering significantly faster processing. The trained models and the code used in the different experiments are available at place holder.</p></div>
• Diff-CA: Separating Common and Salient Factors with Diffusion Models
  
  • Soumm Michaël
  • Fournier-Montgieux Alexandre
  • He Yunlong
  • Gori Pietro
  • Newson Alasdair
  , 2026. Contrastive Analysis aims to separate factors that are common between two data distributions from those that are salient to only one of them. Existing contrastive methods are based on generative models (e.g., VAEs or GANs) that often suffer from limited reconstruction and image quality, which hampers effective latent factor separation and limits their applicability to high-fidelity image generation and edition. We propose a novel conditioning framework for diffusion models that enables contrastive decomposition without compromising generation quality. We first train a prompt-free, image-conditioned diffusion model, and then learn to decompose the conditioning into a common and a salient factor, using weak supervision. We prove that the additive contrastive factorization, commonly assumed in prior work, is identifiable under mild conditions. This factorization enables targeted operations by swapping or interpolating only the salient factor.
• Evaluating DV/CV-QKD Architectures for SAFE Long-Term Secure Storage: A Risk Model and ILP-Based Cost Optimization Approach
  
  • Tasdighi Alireza
  • Alléaume Romain
  , 2026, pp.1-9. This paper presents a unified cost model and optimization framework for evaluating hybrid discrete variable / continuous variable quantum key distribution (DV/CV-QKD) architectures within the SAFE (Secure and Efficient) longterm secure storage (LTSS) protocol. The proposed methodology jointly models the information-theoretic and computational aspects of storage security over extended time horizons and integrates a stochastic integer linear programming (ILP) formulation with a quantitative risk model. By combining Sample-Average Approximation (SAA) with an analytically tractable compromise probability, the framework identifies feasible and cost-optimal SAFE topologies under heterogeneous transmission budgets, keyrate constraints, and security tolerances. Simulation results over randomly generated metropolitan-scale deployments reveal how DV, CV, and multiplexed (MUX) QKD modalities are adaptively selected to minimize infrastructure cost while maintaining a target global compromise probability ϵ. The findings show that mixed DV/CV configurations—particularly when combined with multiplexing—achieve significant cost reductions compared to single-modality networks, paving the way towards the deployment of cost-efficient network infrastructures able to leverage existing QKD technology to provide practical cryptographic advantage for long-term secure storage. (10.1109/QCNC69040.2026.00029)
  DOI : 10.1109/QCNC69040.2026.00029
• Automated Derivation of Formal Properties from Requirements
  
  • Sultan Bastien
  • Apvrille Ludovic
  • Saqui-Sannes Pierre De
  , 2026, pp.1-8. Early detection of design errors in the life cycle of systems is one of the many benefits one may expect from using an model-based systems engineering approach. The formal verification of models is thus an issue. Among verification techniques, the paper focuses on model checking. Taking as input a set of properties and a model, model checking outputs a “yes/no” answer for each property saying whether the property is verified or not. The relevance of this proof relies on the relevance of the defined properties. The paper proposes solutions for automated generation of properties from a SysML model and, more specifically, from the requirements expressed in the SysML model. We introduce an AI- and syntactic rule-based process that assists systems engineers in the generation of syntactically correct CTL properties, along with complementary contributions for computing and propagating requirement coverage from formal verification results, backtracing them directly in the SysML requirement diagram. The approach is mathematically defined, with explanatory comments accompanying the formalism to support comprehension. The proposed approach is further implemented by the free software TTool-AI. The latter is an open source SysML toolkit extended with an LLM-based assistant. A satellite system and an embedded maritime system serve as case studies. (10.1109/SysCon66367.2026.11503633)
  DOI : 10.1109/SysCon66367.2026.11503633
• The FORBIN Dataset: A Collection of Historical Photographs With Archival Metadata
  
  • Chelali Mohamed
  • Gosselet Sylvain-Karl
  • Cloppet Florence
  • Kurtz Camille
  • Bloch Isabelle
  • Foliard Daniel
  Journal of Open Humanities Data, Ubiquity Press, 2026, 12, pp.56. Early photography and the first massification of the medium at the turn of the 20th century is foundational to present day visual cultures. This is specifically true of the visual archives of the first generation of news picture agencies from the 1900s to the 1920s. The latest generation of artificial intelligence models provides the ability to analyse such photographs beyond their purely pictorial dimension, either by grouping them according to their semantic content or by generating textual descriptions. Cross-seeding approaches between historical epistemologies and computer vision expertise can unlock new perspectives on the computational analysis of large and poorly curated photographic archives to provide new insights on early photographic cultures. The automated detection of visual and textual contents, as well as the exploration of the transformation of continuous-tone photographs into halftone images, can help us understand the complexity and historical variability of systems and structures that created, stored and circulated news photographs. In this paper, we introduce the FORBIN dataset. A Paris-based journalist, Victor Forbin, created his own news picture agency. He bought original prints from around the world and then sold them to major French newspapers from the late 1900s to the 1930s. The FORBIN dataset is composed of 62,135 photographs where both the front and back sides are digitized. This paper includes information of the history of this collection, its digitization and its ensuing transformation into a dataset. The Forbin dataset can be used for various computer vision tasks, such as text recognition, image similarity analysis, caption generation, and metadata extraction. (10.5334/johd.487)
  DOI : 10.5334/johd.487
• Breaking exponential complexity in games of ordered preference: A tractable reformulation
  
  • Lee Dong Ho
  • Li Jingqi
  • Peters Lasse
  • Bakirtzis Georgios
  • Fridovich-Keil David
  , 2026. Games of ordered preference (GOOPs) model multi-player equilibrium problems in which each player maintains a distinct hierarchy of strictly prioritized objectives. Existing approaches solve GOOPs by deriving and enforcing the necessary optimality conditions that characterize lexicographically constrained Nash equilibria through a single-level reformulation. However, the number of primal and dual variables in the resulting KKT system grows exponentially with the number of preference levels, leading to severe scalability challenges. We derive a compact reformulation of these necessary conditions that preserves the essential primal stationarity structure across hierarchy levels, yielding a "reduced" KKT system whose size grows polynomially with both the number of players and the number of preference levels. The reduced system constitutes a relaxation of the complete KKT system, yet it remains a valid necessary condition for local GOOP equilibria. For GOOPs with quadratic objectives and linear constraints, we prove that the primal solution sets of the reduced and complete KKT systems coincide. More generally, for GOOPs with arbitrary (but smooth) nonlinear objectives and constraints, the reduced KKT conditions recover all local GOOP equilibria but may admit spurious non-equilibrium solutions. We introduce a second-order sufficient condition to certify when a candidate point corresponds to a local GOOP equilibrium. We also develop a primal-dual interior-point method for computing a local GOOP equilibrium with local quadratic convergence. The resulting framework enables scalable and efficient computation of GOOP equilibria beyond the tractable range of existing exponentially complex formulations.
• Importance Sampling Optimization with Laplace Principle
  
  • Dragomir Radu-Alexandru
  • Portier François
  • Priser Victor
  , 2026. Grid search and random search are widely used techniques for hyperparameter tuning in machine learning, especially when gradient information is unavailable. In these methods, a finite set of candidate configurations is evaluated, and the best-performing one is selected. We propose a simple and computationally inexpensive refinement of this paradigm: instead of selecting a single best point, we form a weighted average of the evaluated configurations, where the weights are chosen using an importance sampling scheme inspired by the Laplace principle. This scheme can be implemented as a post-processing step on top of a random search, with no additional function evaluations. We also propose an iterative variant, where the sampling distributions are chosen adaptively to generate new candidate points around the previous estimate, in the spirit of Evolution Strategy (ES) methods. In a general non-convex setting, we show that, after n evaluations, the error of the proposed methods is of smaller order than n -2/(d+2) . This compares favorably to random search or grid search rates of n -1/d as soon as d &gt; 2. We illustrate the practical benefits of this averaging strategy on several examples.
• Self-Supervised Quantification of Despeckling Uncertainties Applied to Unsupervised Radiometric Change Detection in SAR imaging
  
  • Bultingaire Thomas
  • Kervazo Christophe
  • Denis Loïc
  • Tupin Florence
  , 2026. The Synthetic Aperture Radar (SAR) imaging modality is well-suited for change detection tasks. In particular, its robustness to day-night cycles and cloud coverage is critical for certain applications, such as disaster assessment and monitoring the deforestation in old-growth forests that are frequently cloudcovered. However, its use is often impeded by the speckle phenomenon, causing strong fluctuations in the radar images, which are difficult to distinguish from actual temporal changes. To suppress these fluctuations, we propose to resort to a stateof-art despeckling method. In addition, to better take into account the residual fluctuations in the areas that are the most challenging to restore, we quantify despeckling uncertainties, which enables reaching a constant rate of false positive detections. We also show that incorporating spatial correlations into the despeckling uncertainties model can further improve change detection, especially in highly textured regions such as building areas. The proposed SAR Despeckling Uncertainty Change detection frameworK (SAR-DUCK) is applied to TerraSAR-X and Sentinel-1 images, demonstrating improved detection rates compared to methods from the state-of-the-art, as well as a constant and controlled false alarm rate. Supplementary results and an implementation of the proposed method are available on GitLab at https://gitlab.telecom-paris.fr/ring/sar-duck.
• The Role of Averages in CV-QKD over Fast Fading Channels
  
  • Castillo-Celeita Miguel
  • Schiavon Matteo
  Entropy, MDPI, 2026, 28 (4), pp.388. This work presents a study of continuous-variable quantum key distribution (CV-QKD) protocols over fast-fading channels, typically found in free-space communication links. Two eavesdropping models are considered to evaluate their security under collective attacks: Holevo bound average (HBA) and covariance matrix average (CMA). In the HBA approach, the Holevo bound is averaged over the channel transmittance. In contrast, the CMA method calculates the Holevo bound from the average covariance matrix. Analytical expressions are developed for both strategies. The two methods also differ in how they calculate the mutual information between the legitimate parties. The results demonstrate that the SKR is significantly influenced by how you treat channel fluctuations, highlighting the importance of choosing the model that best describes the actual implementation of the protocol. (10.3390/e28040388)
  DOI : 10.3390/e28040388
• Decentralized Ranking Aggregation: Gossip Algorithms for Borda and Copeland Consensus
  
  • van Elst Anna
  • Le Caillec Kerrian
  • Colin Igor
  • Clémençon Stéphan
  , 2026. <div><p>The concept of ranking aggregation plays a central role in preference analysis, and numerous algorithms for calculating median rankings, often originating in social choice theory, have been documented in the literature, offering theoretical guarantees in a centralized setting, i.e., when all the ranking data to be aggregated can be brought together in a single computing unit. For many technologies (e.g. peer-to-peer networks, IoT, multi-agent systems), extending the ability to calculate consensus rankings with guarantees in a decentralized setting, i.e., when preference data is initially distributed across a communicating network, remains a major methodological challenge. Indeed, in recent years, the literature on decentralized computation has mainly focused on computing or optimizing statistics such as arithmetic means using gossip algorithms. The purpose of this article is precisely to study how to achieve reliable consensus on collective rankings using classical rules (e.g. Borda, Copeland) in a decentralized setting, thereby raising new questions, robustness to corrupted nodes, and scalability through reduced communication costs in particular. The approach proposed and analyzed here relies on random gossip communication, allowing autonomous agents to compute global ranking consensus using only local interactions, without coordination or central authority. We provide rigorous convergence guarantees, including explicit rate bounds, for the Borda and Copeland consensus methods. Beyond these rules, we also provide a decentralized implementation of consensus according to the median rank rule and local Kemenization. Extensive empirical evaluations on various network topologies and real and synthetic ranking datasets demonstrate that our algorithms converge quickly and reliably to the correct ranking aggregation. This work paves the way for principled collective decision-making in fully decentralized systems.</p></div>
• Consensus Ranking Distribution as Ranking Quantization: A K-medians Algorithm
  
  • Clémençon Stéphan
  • Irurozki Ekhine
  • Benedetti Gabriel
  , 2026.
• Taylor-SWFT: fast discrete Statistical Wave Field Theory using Taylor expansion for late reverberation
  
  • Rodrigues Marius
  • Lalay Louis
  • Badeau Roland
  • Richard Gaël
  • Fontaine Mathieu
  , 2026. Dynamic room acoustic simulation aims to render the acoustic effects of an environment in real time while accounting for potentially moving sources and receivers. In this context, the efficient synthesis of the long-term room response, also known as late reverberation, remains challenging because of the intricate relationship between room geometry and acoustic behavior. This paper introduces Taylor-SWFT, an efficient implementation of key results from Statistical Wave Field Theory (SWFT) for the geometry-aware dynamic synthesis of late reverberation. The method is evaluated on the Benchmark for Room Acoustical Simulation (BRAS) and achieves competitive performance compared with classical approaches, while substantially reducing computational cost.
• Cross-Core Covert Channel for RISC-V: Implementation, Countermeasures and Cross-Platform Analysis
  
  • Khan Mahreen
  • Mushtaq Maria
  • Pacalet Renaud
  • Apvrille Ludovic
  , 2026. <div><p>Cache-based covert channels exploit microarchitectural timing differences to enable unauthorized communication between processes. While extensively studied on x86 architectures, such channels remain underexplored in the emerging RISC-V ecosystem. This paper presents the design and implementation of a novel prefetcher and cache timing covert channel for RISC-V platforms that exploits the timing difference between cached and uncached memory accesses. Our implementation supports both standardized RISC-V cache management extensions (Zicbom and Zicbop) and vendor-specific instructions (T-Head C910 custom instructions), demonstrating cross-platform portability across heterogeneous RISC-V implementations. The sender encodes bits by selectively prefetching or flushing a shared cache line, while the receiver decodes information by measuring memory access latency. Through careful synchronization using POSIX shared memory and atomic operations, we achieve reliable bit transmission on both the RISC-V gem5 full-system simulator (Sifive U54 core) and physical RISC-V Beagle-V Ahead (T-Head C910 core). Our paper contributes to understanding the security implications of cache and prefetcher management instructions in RISC-V systems and provides a foundation for developing detection and mitigation strategies for this emerging architecture.</p></div>
• Virtual PUF: Built-in Model for Highly Reliable and Secure PUF
  
  • Nasir Neelam
  • Cheng Wei
  • Kühne Ulrich
  • Graba Tarik
  • Danger Jean-Luc
  , 2026. Strong Physical Unclonable Functions (PUFs) with challengeresponse protocols provide a cost-effective authentication solution for resource-limited devices. However, they are susceptible to modeling attacks. An effective countermeasure for multi-bin PUFs -such as the Ring Oscillator PUF (RO-PUF) and Loop PUF -is the use of Non-Monotonic Quantization (NMQ) of the response. However, NMQ requires quite high quantization levels in order to effectively improve the security. This makes the PUF unreliable, rendering it impractical for authentication purposes. In this paper, we present a solution called the Virtual PUF : It generates a lightweight PUF model at device start-up, which is then used instead of the physical PUF during authentication. This allows the elimination of the noise impact and renders the PUF fully reliable and secure against ML attacks at the same time. As a proof of concept, we present an FPGA implementation based on a Loop PUF -a PUF relying on a single ring oscillator -and we show that beside its perfect reliability and high security, the Virtual PUF can be designed in a lightweight manner. However, the model built by the Virtual PUF may differ from the enrolled model and cause mismatch errors between the two models, which could deteriorate the authentication protocol. We show that the mismatch error measured in different configurations and environments can be optimized towards an acceptable range and/or managed by the authentication protocol.
• LLM-as-a-qualitative-judge: automating error analysis in natural language generation
  
  • Chirkova Nadezhda
  • Ajayi Tunde Oluwaseyi
  • Aycock Seth
  • Mujahid Zain Muhammad
  • Perlić Vladana
  • Borisova Ekaterina
  • Vartampetian Markarit
  , 2026, pp.99-132. Prompting large language models (LLMs) to evaluate generated text, known as LLM-as-ajudge, has become a standard evaluation approach in natural language generation (NLG), but is primarily used as a quantitative tool, i.e. with numerical scores as main outputs. In this work, we propose LLM-as-a-qualitative-judge, an LLM-based evaluation approach with the main output being a structured report of common issue types in the NLG system outputs. Our approach is targeted at providing developers with meaningful insights on what improvements can be done to a given NLG system and consists of two main steps, namely open-ended per-instance issue analysis and clustering of the discovered issues using an intuitive cumulative algorithm. We also introduce a strategy for evaluating the proposed approach, coupled with ∼300 annotations of issues in instances from 12 NLG datasets. Our results show that instancespecific issues output by LLM-as-a-qualitativejudge match those annotated by humans in 2/3 cases, and that LLM-as-a-qualitative-judge is capable of producing error type reports resembling the reports composed by human annotators. We also demonstrate in a case study how the use of LLM-as-a-qualitative-judge can substantially improve NLG systems performance. Our code and data are publicly available. (10.18653/v1/2026.mme-main.7)
  DOI : 10.18653/v1/2026.mme-main.7
• New approaches to CLT for stable random variables
  
  • Coutin Laure
  • Decreusefond Laurent
  • Huang Lorick
  , 2026. In this paper, we present novel approaches to the Central Limit Theorem (CLT) for stable random variables, particularly focusing on cases where the distribution has heavy tails. The study develops three independent methods, each addressing different aspects of convergence in the context of α-stable distributions. We explore both non-integrable and integrable cases, offering new derivations and highlighting the distinct normalization required when dealing with heavy-tailed distributions. Special attention is given to the Stein method, adapted to handle stable laws, leveraging Fourier techniques and Poisson process representations. We provide convergence rates and quantify the Wasserstein distance between sums of independent and identically distributed random variables and their limiting stable distributions. Our findings extend existing results by offering intrinsic methods for stable CLTs, with applications to distributions such as Pareto, where classical CLT normalization fail.
• Data Metrics with Statistical Depths and Kernel Methods
  
  • Castellanos Salinas Arturo
  , 2026. Data is ubiquitous nowadays, with more and more virtualization of the real world. That raises the question: how well can we understand and picture data? “Mathematics and the Picturing of Data” was the title of the seminal work by John Tukey in 1975 that started the field of statistical data depths, and it is suggesting, of course, to rely on mathematical techniques, while keeping the idea of simplicity of a picture. The challenge is to keep the balance between fidelity to the complexity of the data and simplicity to be understandable by the human mind. In computational statistics, this challenge is often translated as using computational power to analyze a finite number of samples, that should approximate well an underlying distribution, yet to compute relevant characteristics of such distribution. Because the computational complexity is bound to increase with the number of samples, a trade-off appears where the statistical approximation error is hoped to decrease fast enough with the number of samples.In this work, we design tools to understand better data under the constraints that their sample complexity should not grow too high. First, we focus on designing new data depth functions with two kind of extensions of the Tukey depth. One extension is done by changing its geometrical nature, inspired by kernel methods to change from the classical Euclidean inner product. The other is by seeing the Tukey depth as classification risk and changing the space of classifiers as well as the loss function. In both cases, we manage to guarantee parametric rate for this non-parametric methods.In a another part, we also consider understanding distribution by looking at distances to discriminate them. The Wasserstein distance from optimal transport is a well-known distance to quantify the difference of two distributions, however its sample complexity suffers from the curse of dimensionality. On the other hand, we show that the Maximum Mean Discrepancy (MMD), a kernel-based distance enjoying good sample complexity, is sometimes too lenient for discriminating distributions, and we suggest to use a distance in-between those two that has higher discriminating power than MMD at the cost of a slight increase of the sample complexity.All of these methods can be used to prevent contamination of data to ruin our statistical analyses, whether it is by detecting outliers using data depth scores or by enhancing the discrimination of distributions via distances.