Publications

2026

• Convex combinations of bosonic pure-loss channels
  
  • Catalano Giuseppe
  • Fanizza Marco
  • Mele Francesco Anna
  • de Palma Giacomo
  • Giovannetti Vittorio
  , 2026. The pure-loss channel is a fundamental model for describing noise in bosonic quantum platforms. It is characterised by a single parameter, the transmissivity, which quantifies the fraction of the input energy that reaches the output of the channel. In realistic scenarios, however, such as free-space quantum communication, the transmissivity is not fixed but fluctuates from one channel use to another. In this setting, the overall channel is effectively described as a convex combination of pure-loss channels, known as a fading channel. Despite its practical relevance, the quantum Shannon theory of the fading channel has remained largely unexplored. Here, we address this gap, specifically investigating degradability, anti-degradability, entanglement breakingness, and capacities of the fading channel. Of particular relevance to practical quantum-internet applications, we prove that entanglement distribution and quantum key distribution can always be achieved at a strictly positive rate over any fading channel, no matter how noisy it is or how strongly the transmissivity fluctuates, provided the channel is not completely noisy. Moreover, we prove that thermal states, which are optimal for a broad class of static bosonic Gaussian channels, fail to achieve the entanglement-assisted classical capacity of fading channels: non-Gaussian Fock-diagonal states strictly outperform all Gaussian encodings. Most strikingly, we identify regimes where the coherent information of thermal inputs vanishes, while optimized non-Gaussian states achieve strictly positive values, thereby activating the channel for quantum communication. For a paradigmatic binary fading model we establish this result analytically, deriving the exact capacity-achieving state in closed form. For general fading distributions, we design an iterative variational algorithm to optimize the coherent and mutual information.
• KARLA: Knowledge-base Augmented Retrieval for Language Models
  
  • Crespin Francois
  • Suchanek Fabian M.
  • Holzenberger Nils
  , 2026. We propose a new method that allows an LLM to automatically pull in factual knowledge from a knowledge base during token generation. This means that (1)~factual knowledge in the LLM output can be updated without retraining the LLM, (2)~facts in the LLM output can be traced to the knowledge base for transparency and explainability, and (3)~smaller models can achieve the same factual accuracy as larger models. Our core idea is to train the model to produce special tokens that trigger a query to the knowledge base. Our experiments show that our method improves factual grounding in both short and long-form generation, and allows factual revisions to take effect through KB edits rather than parameter updates.
• Identifying Security Vulnerabilities in Source Code with Safety Verification
  
  • Kissi Salim Yahia
  • Ameur-Boulifa Rabéa
  • Seladji Yassamine
  International Journal of Critical Computer-Based Systems (IJCCBS), 2026, 12 (01), pp.45 - 70. Ensuring the security of modern software systems is critical due to their increasing complexity and interconnectedness. While automated testing and bug-finding tools have made progress, detecting security flaws--especially those linked to runtime behaviour-- remains a significant challenge. Existing research in software security has largely focused on source code analysis, often overlooking the role of the execution environment and underlying platform. This narrow focus can miss vulnerabilities that arise from machine-dependent behaviours, leading to undetected and potentially critical flaws. To bridge this gap, we introduce a novel approach that analyses source code with a deep awareness of its execution environment, recognising that many vulnerabilities arise from subtle mismatches between the two. Our focus is on detecting arithmetic errors in C/C++ programs by accounting for how the execution context shapes program behaviour. The core of our contribution is a knowledge base that captures the influence of the execution environment on program behaviour using precise logical formulas. This knowledge base enables our algorithm to reformulate vulnerability detection as a satisfiability problem, making it amenable to automated reasoning with formal methods. This article presents the construction of our knowledge base, and details how it integrates with our algorithm to support accurate and environment-aware vulnerability detection. (10.1504/IJCCBS.2026.153768)
  DOI : 10.1504/IJCCBS.2026.153768
• Reinforcement learning for quantum processes with memory
  
  • Lumbreras Josep
  • Huang Ruo Cheng
  • Hu Yanglin
  • Fanizza Marco
  • Gu Mile
  , 2026. In reinforcement learning, an agent interacts sequentially with an environment to maximize a reward, receiving only partial, probabilistic feedback. This creates a fundamental exploration-exploitation trade-off: the agent must explore to learn the hidden dynamics while exploiting this knowledge to maximize its target objective. While extensively studied classically, applying this framework to quantum systems requires dealing with hidden quantum states that evolve via unknown dynamics. We formalize this problem via a framework where the environment maintains a hidden quantum memory evolving via unknown quantum channels, and the agent intervenes sequentially using quantum instruments. For this setting, we adapt an optimistic maximum-likelihood estimation algorithm. We extend the analysis to continuous action spaces, allowing us to model general positive operator-valued measures (POVMs). By controlling the propagation of estimation errors through quantum channels and instruments, we prove that the cumulative regret of our strategy scales as $\tilde{O}(\sqrt{K})$ over $K$ episodes. Furthermore, via a reduction to the multi-armed quantum bandit problem, we establish information-theoretic lower bounds demonstrating that this sublinear scaling is strictly optimal up to polylogarithmic factors. As a physical application, we consider state-agnostic work extraction. When extracting free energy from a sequence of non-i.i.d. quantum states correlated by a hidden memory, any lack of knowledge about the source leads to thermodynamic dissipation. In our setting, the mathematical regret exactly quantifies this cumulative dissipation. Using our adaptive algorithm, the agent uses past energy outcomes to improve its extraction protocol on the fly, achieving sublinear cumulative dissipation, and, consequently, an asymptotically zero dissipation rate.
• MMA-RAG: A Survey on Multimodal Agentic Retrieval-Augmented Generation
  
  • Perlić Vladana
  • Lebailly Stéphane
  • Malvone Vadim
  • Nguyen Van-Tam
  • Urard Pascal
  , 2026. Multimodal Agentic Retrieval-Augmented Generation (MMA-RAG) represents a paradigm shift in artificial intelligence, evolving beyond static text-based retrieval to empower systems with autonomous planning and reasoning capabilities over heterogeneous data types, including text, images, audio, and structured data. This survey provides a comprehensive analysis of the MMA-RAG landscape, introducing a unified taxonomy organized along the complementary axes of modality (text-only vs. multimodal) and agency (static vs. agentic). We systematically review foundational literature and state-of-the-art frameworks to identify dominant architectural patterns—specifically Hierarchical Orchestration, Iterative Refinement, and Domain-Specific Workflow Emulation—that characterize current systems. The survey evaluates the application of these patterns across diverse domains, such as healthcare, document understanding, and embodied AI, while surfacing persistent challenges in cross-modal alignment, system scalability, and the fragmentation of evaluation benchmarks. Moving beyond a retrospective review, we propose a consolidated research agenda to transition MMA-RAG from descriptive models to auditable, high-stakes decision-support agents. This roadmap identifies critical architectural opportunities, conceptualizing frameworks such as Reasoning Trace Tables for authentic in-context learning, Calibration-Aware Compositional Embedding Hubs for precise grounding, and Dual-Coded Episodic Memory to overcome text-bottlenecking, accompanied by a call for novel metrics to rigorously validate agentic reasoning and safety.
• Towards sample-optimal learning of bosonic Gaussian quantum states
  
  • Chen Senrui
  • Mele Francesco Anna
  • Fanizza Marco
  • Li Alfred
  • Mann Zachary
  • Huang Hsin-Yuan
  • Chen Yanbei
  • Preskill John
  , 2026. Continuous-variable systems enable key quantum technologies in computation, communication, and sensing. Bosonic Gaussian states emerge naturally in various such applications, including gravitational-wave and dark-matter detection. A fundamental question is how to characterize an unknown bosonic Gaussian state from as few samples as possible. Despite decades-long exploration, the ultimate efficiency limit remains unclear. In this work, we study the necessary and sufficient number of copies to learn an $n$-mode Gaussian state, with energy less than $E$, to $\varepsilon$ trace distance closeness with high probability. We prove a lower bound of $\Omega(n^3/\varepsilon^2)$ for Gaussian measurements, matching the best known upper bound up to doubly-log energy dependence, and ${\Omega}(n^2/\varepsilon^2)$ for arbitrary measurements. We further show an upper bound of $\widetilde{O}(n^2/\varepsilon^2)$ given that the Gaussian state is promised to be either pure or passive. Interestingly, while Gaussian measurements suffice for nearly optimal learning of pure Gaussian states, non-Gaussian measurements are provably required for optimal learning of passive Gaussian states. Finally, focusing on learning single-mode Gaussian states via non-entangling Gaussian measurements, we provide a nearly tight bound of $\widetilde\Theta(E/\varepsilon^2)$ for any non-adaptive schemes, showing adaptivity is indispensable for nearly energy-independent scaling. As a key technical tool of independent interest, we establish stringent bounds on the trace distance between Gaussian states in terms of the total variation distance between their Wigner functions. In particular, this yields a nearly tight sample complexity of $\widetilde{\Theta}(n^{2}/\varepsilon^{2})$ for learning the Wigner distribution of any Gaussian state to $\varepsilon$ total variation distance, achievable with Gaussian measurements. Our results greatly advance quantum learning theory in the bosonic regimes and have practical impact in quantum sensing and benchmarking applications.
• Hardness of M-LWE with General Distributions and Applications to Leaky Variants
  
  • Boudgoust Katharina
  • Jeudy Corentin
  • Tairi Erkan
  • Wen Weiqiang
  , 2026, 16551, pp.3-37. The Module Learning With Errors (M-LWE) problem has become a fundamental hardness assumption for lattice-based cryptography. It offers an attractive trade-o between strong robustness guarantees, sometimes directly based on worst-case lattice problems, and efficiency of the subsequent cryptographic primitives. Different flavors of M-LWE have then been introduced towards improving performance. Such variants look at different secret-error distributions and might allow for additional hints on the secret-error vector. Existing hardness results however only cover restricted classes of said distributions, or are tailored to specific leakage models. This lack of generality hinders the design of efficient and versatile cryptographic schemes, as each new distribution or leakage model requires a separate and nontrivial hardness evaluation In this work, we address this limitation by establishing the hardness of M-LWE under general distributions. As a first step, we show that M-LWE remains hard when the error vector follows an arbitrary bounded distribution with sufficient entropy, with some restriction on the number of samples. Building on this, we then reduce to the Hermite Normal Form (HNF) where the secret-error vector follows said arbitrary distribution. Overall, our result shows the actual shape of the distribution does not matter, as long as it keeps sufficient entropy. To demonstrate the versatility of our framework, we further analyze a range of leakage scenarios. By examining the residual entropy given the leakage, we show that our results of M-LWE with general distributions encompass various types of leakage. More precisely, we cover exact and approximate linear hints which are widely used in recent cryptographic designs, as well as quadratic, and even non-algebraic forms, some of which were not yet covered by any theoretical hardness guarantees. The generality of our results aims at facilitating future cryptographic designs and security analyses. (10.1007/978-3-032-26731-3_1)
  DOI : 10.1007/978-3-032-26731-3_1
• A Noise-Shaping SAR ADC With Second-Order Input Prediction
  
  • Nguyen-Thien Viet
  • Jabbour Chadi
  • Nguyen Minh-Tien
  • Delorme Nicolas
  , 2026, pp.2982-2986. This paper presents a low-power noise-shaping (NS) SAR ADC with a second-order input prediction algorithm. This algorithm uses polynomial interpolation to predict the bits of the next sample. Since the ADC only needs to convert the small difference between the input signal and the predicted code, the energy per conversion and the number of bit cycles per conversion can be reduced. The prototype ADC is designed in a 130-nm CMOS process. It consumes 270 µW from a 1.2-V supply while operating at 16 MS/s. The simulated SNDR is 86.3 dB over a bandwidth of 500 kHz, corresponding to an SNDR-based Schreier figure of merit (FoM) of 179 dB. (10.1109/ISCAS66217.2026.11562036)
  DOI : 10.1109/ISCAS66217.2026.11562036
• Arena: a kubernetes-based testbed for evaluating application deployment across the computing continuum
  
  • Huang Chih-Kai
  • Krouti Konstantinos
  • Markopoulou Stella
  • Tserpes Konstantinos
  • Bouloukakis Georgios
  , 2026. This paper introduces Arena, a Kubernetes-based testbed for evaluating application deployment across computing continuum environments (IoT/Edge/Cloud). Arena enables the emulation of diverse computing nodes using Docker containers and leverages Kubernetes for testbed management. Arena integrates the Chaos Mesh framework to simulate network characteristics and Prometheus with Grafana tools for monitoring and visualization purposes. Experiments on the Grid'5000 platform with a Google microservice application demonstrate that Arena's container-based emulation achieves similar resource usage patterns to virtual machine-based nodes, and its network chaos injection effectively enforces network constraints. Results highlight Arena's capability to provide a practical and reproducible environment for testing containerized applications across diverse computing continuum nodes.
• Co-investment in mobile edge computing with infrastructure update and dynamic participation
  
  • Sakr Amal
  • Araldo Andrea
  • Chahed Tijani
  • Kofman Daniel
  , 2026. Mobile Edge Computing (MEC) requires Network Operators (NOs) to undertake substantial infrastructure investments, while most revenues are captured by Service Providers (SPs) offering end-user applications. This cost-revenue imbalance discourages NOs from investing in MEC deployment, despite increasing demand for low-latency and bandwidth-intensive services. This paper proposes a co-investment scheme in which players, i.e., one NO and multiple SPs, jointly deploy, maintain, and share MEC infrastructure over multiple decision epochs. We devise a new coalitional game model that captures the planning of resources, their allocation among players, and cost and revenue sharing. To address fluctuating user demand and evolving participation incentives, we design a mechanism that updates resources and allows the dynamic entrance and exit of players over time. We sustain cooperation through a compensation scheme. Numerical results show that combining resource updates with dynamic participation increases the total payoff and strengthens the NO's incentive to invest.
• Generic neural network model for estimating exposure levels in industrial environments
  
  • Plets David
  • Apostolidis Christos
  • Wang Shanshan
  • Valič Blaž
  • Martens Luc
  • Samaras Theodoros
  • Gajšek Peter
  Journal of Radiological Protection, IOP Publishing, 2026, 46 (2), pp.021518 (1-14). Abstract This study describes a neural network-based method for estimating exposure levels in industrial environments, without requiring detailed technical inputs, allowing usage of the model by layman people or by workers active in these areas. A pipeline based on Blender environments and MATLAB ray-tracing simulations is created and after defining a set of 11 candidate input parameters for the model, more than 20 000 different wireless configurations are simulated, varying the different environmental and wireless input parameters. A correlation analysis shows that main inputs influencing the exposure levels in the industrial area are the transmit power of the antennas, the density of clutter in the area, the density of transmitters in the area, and the height and location of the transmitters. A multi-layer fully connected neural network regression model is developed to predict median ( E 50 ) and 95th percentile ( E 95 ) exposure levels in industrial areas. Testing the obtained model on an unseen dataset of environments with E 50 values between 0 and 3.25 V m −1 and E 95 values between 0 and 7 V m −1 , demonstrates the good prediction performance of the model: root-mean-square error values below 0.173 V m −1 and R 2 values above 95% are obtained. Subsequently, the model is validated with measurement data collected in three distinct realistic industrial environments. The average absolute deviation of the model predictions with respect to the measurements is limited to 20.4%. This novel and broadly accessible approach demonstrates that it is possible to reliably estimate exposure levels in realistic environments without having to rely on external experts or on dedicated complex software. (10.1088/1361-6498/ae6c32)
  DOI : 10.1088/1361-6498/ae6c32
• Integrating Visual Features in Multiple Hypothesis Tracking Through Self-Supervised Learning
  
  • Reme Raphael
  • Newson Alasdair
  • Angelini Elsa
  • Olivo-Marin Jean-Christophe
  • Lagache Thibault
  , 2026, pp.1-5. Multiple Object Tracking (MOT) is a crucial step in the automatic analysis of biological processes. In this work, we introduce VISUALMHT, a Bayesian Multiple Hypothesis Tracking algorithm integrating self-supervised visual features to leverage robust target identity and consistently track biological objects. This novel tracking algorithm achieves state-of-the-art results on the SINETRA synthetic datasets and demonstrates robust neuron tracking in the freshwater cnidarian Hydra vulgaris, highlighting the benefit of combining probabilistic motion models with learned visual representations. (10.1109/ISBI61048.2026.11515572)
  DOI : 10.1109/ISBI61048.2026.11515572
• Conception d'un redresseur pour les bandes Wi-Fi et millimétrique
  
  • Wang Yibo
  • Niotaki Kyriaki
  • Lepage Anne Claire
  • Begaud Xavier
  , 2026.
• AudioCAN: Enhanced Few-Shot Audio Classification via Energy-Guided Temporal Cross Attention
  
  • Zhuang Xuanyu
  • Peeters Geoffroy
  • Richard Gaël
  IEEE Signal Processing Letters, Institute of Electrical and Electronics Engineers, 2026, 33, pp.2215-2219. <div><p>Despite the growing research interest and practical potential of few-shot audio classification, its efficacy remains limited by the temporal sparsity of target sound events and the presence of non-target acoustic interference within audio samples. In this work, we leverage the cross attention mechanism to address these challenges. While the Cross Attention Network (CAN) has demonstrated superior performance in few-shot image classification by highlighting semantically relevant regions between support and query features, we propose AudioCAN by extending CAN to audio domain through two-fold, audiospecific modifications. In particular, we first reformulate the original 2D spatial cross attention to 1D temporal version to prioritize the time frames containing discriminative acoustic information, thereby mitigating interference from non-target sounds. Secondly, we introduce an energy-guided masking strategy to synthesize a pseudo-query set from the support set via data augmentation, which then serves as a guidance for optimizing the 1D temporal cross attention during training. Experimental results on several few-shot audio classification benchmarks demonstrate that AudioCAN achieves state-of-the-art performance in 5-way-1shot settings while remaining highly competitive in 5-way-5-shot configurations.</p></div> (10.1109/LSP.2026.3694791)
  DOI : 10.1109/LSP.2026.3694791
• [WIM AWARD] Nanogénérateurs Triboélectriques à Glissement Latéral pour la Récupération d’Énergie Hybride
  
  • Lu Yayi
  • Niotaki Kyriaki
  • Mohellebi Reda
  • Ben Kalaia Karim
  • Jabbour Chadi
  , 2026, pp.1. La récupération d'énergie RF est employée dans le but d'alimenter des appareils connectés à faible consommation. Pour augmenter et stabiliser le niveau de puissance récupéré à la sortie, la source RF est souvent associée avec d'autres types d'énergie dans un système complexe, on parle alors de récupération d'énergie hybride multi-sources. La triboélectrité basée sur le frottement a été repérée comme une voie émergente. Nos études paramétriques sur les nanogénérateurs triboélectriques à glissement latéral (Lateral Sliding-Triboelectric Nanogenerators/LS-TENGs) démontrent que : l'impédance de charge, la vitesse de déplacement et la surface de contact sont parmi les principaux facteurs à considérer, afin de maximiser la puissance récupérée.
• Energy efficiency of quantum computers
  
  • Carrasco-Codina Miquel
  • Escofet Pau
  • Hilaire Paul
  • Soret Ariane
  • Nerenberg Sam
  • Champain Victor
  • Milburn Gerard
  • Theophilo Klara
  • Li Sophie
  • Bautista Irais
  • Gómez Andrés
  • Miralles Jose
  • Abadal Sergi
  • Almudéver Carmen
  • Alarcón Eduard
  • Yehia Raja
  , 2026. How much energy does a quantum computer consume? Are they more efficient than their classical counterparts? In this work, we make a step towards answering these questions. We define the energy efficiency of a quantum computer as the ratio of the number of algorithms it can perform during a given time over the energy consumed by the hardware during this time. We analyze the most representative physical platforms currently envisioned to be used as building blocks of quantum computers: superconducting qubits, silicon spin qubits, trapped ions, neutral atoms and photonic qubits. Including insights from experts in all these technologies and taking into account algorithm compilation constraints, we discuss the advantages and inconveniences of each platform from an energy standpoint. Beyond providing concrete values of the energy consumption of current quantum computers, we lay the foundation of a framework to benchmark the energy efficiency of any future quantum computing architecture. (10.48550/arXiv.2605.15090)
  DOI : 10.48550/arXiv.2605.15090
• Complexity of unicity problems
  
  • Hudry Olivier
  , 2026.
• SPOT: An Annotated French Corpus and Benchmark for Detecting Critical Interventions in Online Conversations
  
  • Berriche Manon
  • Nouri Célia
  • Clavel Chloé
  • Cointet Jean-Philippe
  , 2026. We introduce SPOT (Stopping Points in Online Threads), the first annotated corpus translating the sociological concept of stopping point into a reproducible NLP task. Stopping points are ordinary critical interventions that pause or redirect online discussions through a range of forms (irony, subtle doubt or fragmentary arguments) that frameworks like counterspeech or social correction often overlook. We operationalize this concept as a binary classification task and provide reliable annotation guidelines. The corpus contains 43,305 manually annotated French Facebook comments linked to URLs flagged as false information by social media users, enriched with contextual metadata (article, post, parent comment, page or group, and source). We benchmark fine-tuned encoder models (CamemBERT) and instruction-tuned LLMs under various prompting strategies. Results show that fine-tuned encoders outperform prompted LLMs in F1 score by more than 10 percentage points, confirming the importance of supervised learning for emerging non-English social media tasks. Incorporating contextual metadata further improves encoder models F1 scores from 0.75 to 0.78. We release the anonymized dataset, along with the annotation guidelines and code in our code repository, to foster transparency and reproducible research.
• Experimental Evaluation of QoT Uncertainty Reduction Techniques in Network Digital Twins
  
  • Purkayastha Ambashri
  • Delezoide Camille
  • Boitier Fabien
  • Lourdiane Mounia
  • Ware Cédric
  • Layec Patricia
  , 2026. <div><p>We experimentally demonstrate sequential parameter refinement leveraging lightpath diversity in presence of various levels of uncertainty on parameter knowledge to iteratively improve QoT prediction accuracy by up to 1.8 dB.</p></div>
• ENEIDE: A High Quality Silver Standard Dataset for Named Entity Recognition and Linking in Historical Italian
  
  • Santini Cristian
  • Barzaghi Sebastian
  • Sernani Paolo
  • Frontoni Emanuele
  • Melosi Laura
  • Alam Mehwish
  , 2026. This paper introduces ENEIDE (Extracting Named Entities from Italian Digital Editions), a silver standard dataset for Named Entity Recognition and Linking (NERL) in historical Italian texts. The corpus comprises 2,111 documents with over 8,000 entity annotations semi-automatically extracted from two scholarly digital editions: Digital Zibaldone, the philosophical diary of the Italian poet Giacomo Leopardi (1798-1837), and Aldo Moro Digitale, the complete works of the Italian politician Aldo Moro (1916-1978). Annotations cover multiple entity types (person, location, organization, literary work) linked to Wikidata identifiers, including NIL entities that cannot be mapped to the knowledge graph. To the best of our knowledge, ENEIDE represents the first multi-domain, publicly available NERL dataset for historical Italian with training, development, and test splits. We present a methodology for semi-automatic annotations extraction from manually curated scholarly digital editions, including quality control and annotation enhancement procedures. Baseline experiments using state-of-the-art models demonstrate the dataset's challenge for NERL and the gap between zero-shot approaches and fine-tuned models. The dataset's diachronic coverage spanning two centuries makes it particularly suitable for temporal entity disambiguation and cross-domain evaluation.
• Analysing Lightweight Large Language Models for Biomedical Named Entity Recognition on Diverse Ouput Formats
  
  • Epron Pierre
  • Coulet Adrien
  • Alam Mehwish
  , 2026, pp.1-13. Despite their strong linguistic capabilities, Large Language Models (LLMs) are computationally demanding and require substantial resources for fine-tuning, which is unadapted to privacy and budget constraints of many healthcare settings. To address this, we present an experimental analysis focused on Biomedical Named Entity Recognition using lightweight LLMs, we evaluate the impact of different output formats on model performance. The results reveal that lightweight LLMs can achieve competitive performance compared to the larger models, highlighting their potential as lightweight yet effective alternatives for biomedical information extraction. Our analysis shows that instruction tuning over many distinct formats does not improve performance, but identifies several format consistently associated with better performance.
• Human vs LLM in Conversational Repair Annotation: A New Resource and Comparative Study
  
  • Ngo Anh
  • Rollet Nicolas
  • Pelachaud Catherine
  • Clavel Chloé
  , 2026, pp.6880-6892. Addressing the scarcity of annotated data for Other-Initiated Repair (OIR), when recipients interrupt conversation progressivity to signal trouble, prompting speakers to provide repair, this work introduces OIR annotations for the NOXI corpus, achieving considerable reliability. We evaluate whether LLMs can reliably annotate OIR sequences using structured Chain-of-Thought prompting and conduct comparative analysis across two corpora: NOXI (natural dialogue) and CABB-S (Dutch, task-oriented), finding weak alignment between LLMs and human annotations, particularly in recognizing trouble signaling. Analyzing human-LLM disagreement using the LLM-generated explanations revealed limitations: models rely on lexical patterns rather than conversational context, construct reasonable-sounding but misleading narratives, highlighting crucial limitations for both automated annotation of complex interactional phenomena. (10.63317/32stceuq67kb)
  DOI : 10.63317/32stceuq67kb
• POT Python Optimal Transport
  
  • Flamary Rémi
  • Vincent-Cuaz Cédric
  • Courty Nicolas
  • Gramfort Alexandre
  • Kachaiev Oleksii
  • Quang Tran Huy
  • David Laurène
  • Bonet Clément
  • Cassereau Nathan
  • Gnassounou Theo
  • Tanguy Eloi
  • Delon Julie
  • Collas Antoine
  • Mazelet Sonia
  • Chapel Laetitia
  • Kerdoncuff Tanguy
  • Yu Xizheng
  • Feickert Matthew
  • Krzakala Paul
  • Liu Tianlin
  • Fernandes Montesuma Eduardo
  , 2026. (10.5281/ZENODO.17161062)
  DOI : 10.5281/ZENODO.17161062
• Leveraging machine learning for efficient and secure optical networking
  
  • Andrenacci Isaia
  , 2026. Future optical networks face rapidly growing traffic demands while operating close to fundamental capacity limits, making efficient network resource utilization increasingly critical. Massive monitoring and telemetry have emerged as a possible solution for more autonomous and adaptive network operation. In this PhD, I investigate machine learning-based techniques that leverage receiver-side measurements and optical network parameters to enhance the efficiency, reliability, and security of point-to-point optical fiber links. First, I propose machine learning models to accurately predict nonlinear interference spectrum components in wavelength-division multiplexed systems, enabling low-complexity and high-accuracy performance estimation. I then develop closed-loop machine learning controllers to optimize launch power in links affected by polarization-dependent loss using only receiver-side signal-to-noise ratio statistics. Furthermore, I introduce low-cost monitoring techniques to localize anomalous polarization-dependent loss elements and to identify the dominant noise in the link, facilitating future adaptive optimization and self-healing of the network. Finally, I propose a method to quantify fiber macro-bending events from spectral measurements, allowing future discrimination between maintenance activities and physical-layer attacks. Overall, this work demonstrates that practical machine learning methods, combined with pervasive monitoring, can support near-zero-margin operation and pave the way toward intelligent, autonomous, and secure next-generation optical networks.
• Kalman Filtering for Sensing Aided Communication to Mobile Users in Large Cellular Networks
  
  • Balakrishnan Ashutosh
  • Soprano-Loto Nahuel
  • Baccelli François
  , 2026. Upcoming 6G networks are expected to have joint sensing and communication capabilities. In this work, we propose a sensing-aided communication (SAC) framework, wherein a cellular network of randomly located base stations (BSs) sequentially estimates the position and channel gain of a user equipment (UE) moving randomly across the network. At the core of this framework lies a state-dependent Kalman filter that extends the classical formulation in two key respects: (i) measurements are acquired from a dynamically selected BS, the selection based on the a-priori estimate of the UE's position and fading channel; and (ii) the measurement noise covariance is modelled as a function of the sensing distance and fading, coupling the temporal evolution of the state with the measurement quality. This coupling gives rise to pronounced estimation error peaks during handovers, which in turn directly govern the design of initial access and handover protocols in stochastic networks. To quantify performance, we introduce a new metric, the stationary mean of the estimation error covariance, and establish its validity through an ergodic theorem, further supported by numerical convergence studies. The sensing-based UE-BS association is evaluated in one-and two-dimensional Gauss-Markov mobility scenarios and benchmarked against two alternatives: an autonomous noise covariance setting and an ideal omniscient controller with perfect position knowledge. The SAC framework is shown to outperform a position-only sensing framework by around 32%. Overall, the results highlight the advantage of the SAC framework in providing protocol-level insights for handover-aware sensing and demonstrate that the proposed filter achieves stable long-run performance.