Publications

2025

• Decoding Persuasiveness in Eloquence Competitions: An Investigation into the LLM’s Ability to Assess Public Speaking
  
  • Barkar Alisa
  • Chollet Mathieu
  • Labeau Matthieu
  • Biancardi Beatrice
  • Clavel Chloé
  , 2025, pp.538-546. The increasing importance of public speaking (PS) skills has fueled the development of automated assessment systems, yet the integration of large language models (LLMs) in this domain remains underexplored. This study investigates the application of LLMs for assessing PS by predicting persuasiveness. We propose a novel framework where LLMs evaluate criteria derived from educational literature and feedback from PS coaches, offering new interpretable textual features. We demonstrate that persuasiveness predictions of a regression model with the new features achieve a Root Mean Squared Error (RMSE) of 0.6, underperforming approach with hand-crafted lexical features (RMSE 0.51) and outperforming direct zero-shot LLM persuasiveness predictions (RMSE of 0.8). Furthermore, we find that only LLM-evaluated criteria of language level is predictable from lexical features (F1-score of 0.56), disapproving relations between these features. Based on our findings, we criticise the abilities of LLMs to analyze PS accurately. To ensure reproducibility and adaptability to emerging models, all source code and materials are publicly available on GitHub. (10.5220/0013158400003890)
  DOI : 10.5220/0013158400003890
• Rapid mixing, partition function estimation and universal quantum computation with dissipative quantum Gibbs sampling
  
  • Rouzé Cambyse
  • França Daniel Stilck
  • Alhambra Álvaro M.
  , 2026.
• Efficient Hamiltonian, structure and trace distance learning of Gaussian states
  
  • Fanizza Marco
  • Rouzé Cambyse
  • Stilck Franca Daniel
  , 2024. In this work, we initiate the study of Hamiltonian learning for positive temperature bosonic Gaussian states, the quantum generalization of the widely studied problem of learning Gaussian graphical models. We obtain efficient protocols, both in sample and computational complexity, for the task of inferring the parameters of their underlying quadratic Hamiltonian under the assumption of bounded temperature, squeezing, displacement and maximal degree of the interaction graph. Our protocol only requires heterodyne measurements, which are often experimentally feasible, and has a sample complexity that scales logarithmically with the number of modes. Furthermore, we show that it is possible to learn the underlying interaction graph in a similar setting and sample complexity. Taken together, our results put the status of the quantum Hamiltonian learning problem for continuous variable systems in a much more advanced state when compared to spins, where state-of-the-art results are either unavailable or quantitatively inferior to ours. In addition, we use our techniques to obtain the first results on learning Gaussian states in trace distance with a quadratic scaling in precision and polynomial in the number of modes, albeit imposing certain restrictions on the Gaussian states. Our main technical innovations are several continuity bounds for the covariance and Hamiltonian matrix of a Gaussian state, which are of independent interest, combined with what we call the local inversion technique. In essence, the local inversion technique allows us to reliably infer the Hamiltonian of a Gaussian state by only estimating in parallel submatrices of the covariance matrix whose size scales with the desired precision, but not the number of modes. This way we bypass the need to obtain precise global estimates of the covariance matrix, controlling the sample complexity.
• Lessons for Interactive Theorem Proving Researchers from a Survey of Coq Users
  
  • de Almeida Borges Ana
  • Casanueva Artís Annalí
  • Falleri Jean-Rémy
  • Gallego Arias Emilio Jesús
  • Martin-Dorel Érik
  • Palmskog Karl
  • Serebrenik Alexander
  • Zimmermann Théo
  Journal of Automated Reasoning, Springer Verlag, 2025, 69 (8), pp.1-29. The Coq Community Survey 2022 was an online public survey of users of the Coq proof assistant conducted during February 2022. Broadly, the survey asked about use of Coq features, user interfaces, libraries, plugins, and tools, views on renaming Coq and Coq improvements, and also demographic data such as education and experience with Coq and other proof assistants and programming languages. The survey received 466 submitted responses, making it the largest survey of users of an interactive theorem prover (ITP) so far. We present the design of the survey, a summary of key results, and analysis of answers relevant to ITP technology development and usage. In particular, we analyze user characteristics associated with adoption of tools and libraries and make comparisons to adjacent software communities. Notably, we find that experience has significant impact on Coq user behavior, including on usage of tools, libraries, and integrated development environments (IDEs). (10.1007/s10817-025-09720-1)
  DOI : 10.1007/s10817-025-09720-1
• A Single-Layer Efficient Metasurface Absorber for RF Energy Harvesting Applications
  
  • Sharifi Raziyeh
  • Lepage Anne Claire
  • Niotaki Kyriaki
  • Begaud Xavier
  , 2025. In this contribution, a single-layer metasurface absorber is proposed for collecting the ambient radio frequency energy at 2.45 GHz. A design methodology is proposed to optimize the finite array performance. An absorber of 4×5 cells is designed and simulated to demonstrate that this methodology enables an improvement of the capturing efficiency from 65% to 90%. The proposed finite array is fabricated and its capturing efficiency is measured in order to be compared to a simulation.
• Impact of Sub-Segment Representations in DASH on the Live Streaming Experience
  
  • Ugur Deniz
  • Bouqueau Romain
  • Stattmann Michael
  • Feuvre Jean Le
  , 2025, pp.72-73. The demand for low-latency live streaming continues to grow, necessitating advancements in transport and packaging technologies. Media-over-QUIC (MoQ) has emerged as a promising approach for scalable, low-latency streaming, though its adoption remains in early stages. Meanwhile, the 6th Edition of MPEG-DASH introduces Sub-Segment Representations (SSR) to reduce tune-in times by leveraging dual-encoding strategies with varying Group of Pictures (GOP) lengths. This paper evaluates the feasibility of SSR in ultra-low latency streaming environments, particularly its interaction with Content Delivery Networks (CDNs) and HTTP/3 stream prioritization. Through controlled experiments, we compare the performance of Low-Latency Low-Delay (L3D) playback against standard playback under different CDN configurations. Results indicate that L3D significantly reduces preroll buffering and stabilizes playback, even at sub-second latencies. Additionally, our findings highlight the role of QUIC and HTTP/3 in optimizing live streaming performance. These insights contribute to understanding SSR's potential in improving scalability, latency, and viewer experience in live streaming workflows. (10.1145/3715675.3715807)
  DOI : 10.1145/3715675.3715807
• Distributed Coherent Sensing Over Deployed Fibers for Network as a Sensor Applications
  
  • Guerrier Sterenn
  • Dorize Christian
  • Abdelli Khouloud
  • Mardoyan Haïk
  • Pavani Henrique
  • Antonelli Cristian
  • Mecozzi Antonio
  • Koubaa Amin
  • Darwish Khalid
  • Biyahi Mohammed
  • Awwad Élie
  • Renaudier Jérémie
  Journal of Lightwave Technology, Institute of Electrical and Electronics Engineers (IEEE)/Optical Society of America(OSA), 2025, 43 (4), pp.1736-1745. We discuss the performance of Coherent-MIMO-DFS over deployed optical networks in various configurations and address technological challenges such as adaptation to various fiber types and disturbance identification. Two different field trial results are exploited, demonstrating the adaptability of our distributed fiber sensing interrogator to the diverse environments that can be encountered in the context of sensing over terrestrial telecommunication networks. We discuss the advantages of extracting the full backscattered Jones matrices using a DFS interrogator. Mechanical events including threats to the infrastructure are localized and identified and different identification methods are discussed. We draw a correspondence between lab measurements and field events based on the type of environment, and finally we present a classification method based on transfer learning with 90% accuracy. (10.1109/JLT.2024.3498070)
  DOI : 10.1109/JLT.2024.3498070
• A Fused Gromov-Wasserstein Approach to Subgraph Contrastive Learning
  
  • Sangare Amadou S.
  • Dunou Nicolas
  • Giraldo Jhony H.
  • Malliaros Fragkiskos D.
  Transactions on Machine Learning Research, [Amherst Massachusetts]: OpenReview.net, 2022, 2025. Self-supervised learning has become a key method for training deep learning models when labeled data is scarce or unavailable. While graph machine learning holds great promise across various domains, the design of effective pretext tasks for self-supervised graph representation learning remains challenging. Contrastive learning, a popular approach in graph self-supervised learning, leverages positive and negative pairs to compute a contrastive loss function. However, current graph contrastive learning methods often struggle to fully use structural patterns and node similarities. To address these issues, we present a new method called Fused Gromov-Wasserstein Subgraph Contrastive Learning (FOSSIL). Our method integrates node-level and subgraph-level contrastive learning, seamlessly combining a standard node-level contrastive loss with the Fused Gromov-Wasserstein distance. This combination helps our method capture both node features and graph structure together. Importantly, our approach works well with both homophilic and heterophilic graphs and can dynamically create views for generating positive and negative pairs. Through extensive experiments on benchmark graph datasets, we show that FOSSIL outperforms or achieves competitive performance compared to current state-of-the-art methods.
• Security and Robustness of Autonomous Driving Systems Against Physical Adversarial Attack
  
  • Chi Lijun
  , 2025. With iterative hardware upgrades and advancements in deep neural networks (DNNs), autonomous driving systems (ADS) are increasingly integrated in life. However, before this technology becomes widespread, a security issue that needs to be addressed is physical adversarial attacks. Such attacks can manipulate real-world objects to disrupt the perception of ADSs and cause traffic accidents. In addition, the diversity of physical attacks makes it difficult for passive defenders.This study addresses these challenges by analyzing, evaluating, and developing practical strategies to improve the robustness of ADS.It begins with a review of recent physical adversarial attacks that identifies specific threats to ADSs.It then introduces a novel public attention-based black-box attack that demonstrates how an attacker can exploit ADS awareness without full knowledge of the system, highlighting the need for enhanced defenses.Next, a lightweight detection framework is proposed for real-time laser-based attack detection. Additionally, a defense mechanism called Laser Shield is developed, using polarizers to block harmful laser signals and enhance ADS security. (10.70675/38cb9544z5106z4ec9zaf28z6fedd9c361dc)
  DOI : 10.70675/38cb9544z5106z4ec9zaf28z6fedd9c361dc
• Private Data Analysis over Encrypted Databases : Mixing Functional Encryption with Computational Differential Privacy
  
  • Alborch Escobar Ferran
  , 2025. In our current digitalized society, data is ruling the world. But as it is most of the time related to individuals, its exploitation should respect the privacy of the latter. This issue has raised the differential privacy paradigm, which permits to protect individuals when querying databases containing data about them. But with the emergence of cloud computing, it is becoming increasingly necessary to also consider the confidentiality of "on-cloud'' storage confidentiality of such vast databases, using encryption techniques. This thesis studies how to provide both privacy and confidentiality of such outsourced databases by mixing two primitives: computational differential privacy and functional encryption. First, we study the relationship between computational differential privacy and functional encryption for randomized functions in a generic way. We analyze the privacy of the setting where a malicious analyst may access the encrypted data stored in a server, either by corrupting or breaching it, and prove that a secure randomized functional encryption scheme supporting the appropriate family of functions guarantees the computational differential privacy of the system. Second, we construct efficient randomized functional encryption schemes for certain useful families of functions, and we prove them secure in the standard model under well-known assumptions. The families of functions considered are linear functions, used for example in counting queries, histograms and linear regressions, and quadratic functions, used for example in quadratic regressions and hypothesis testing. The schemes built are then used together with the first result to construct encrypted databases for their corresponding family of queries. Finally, we implement both randomized functional encryption schemes to analyze their efficiency. This shows that our constructions are practical for databases with up to 1 000 000 entries in the case of linear queries and databases with up to 10 000 database entries in the case of quadratic queries. (10.70675/7fd3987az3c65z4befz98f3zc4201f35a247)
  DOI : 10.70675/7fd3987az3c65z4befz98f3zc4201f35a247
• Laser Guard: Efficiently Detecting Laser-Based Physical Adversarial Attacks in Autonomous Driving
  
  • Chi Lijun
  • Msahli Mounira
  IEEE Access, IEEE, 2025, 13, pp.35219-35229. <div><p>The fast development of deep learning (DL) enables even resource-constrained devices to tackle complex artificial intelligence (AI) tasks, especially those related to environment perception in autonomous driving systems (ADS). However, AI models deployed in the real world are exposed to the threats of adversarial examples (AE). One specific type of physical attack utilizes laser beams or spots planted on images rather than crafted pixel-level perturbations to manipulate the victim deep neural networks (DNN) prediction. These attacks easily mislead traffic sign recognition and object detection in ADS. Laser-based adversarial attacks are cognitively stealthy but visually conspicuous, invalidating the previous defenses designed for digital attacks. This study considers two state-of-the-art (SOTA) laser-based attacks and establishes a benchmark comprising thousands of AEs. Such AEs have distinct pattern features, significant occupation, high contrast, and low variance. Based on the observation, a lightweight detection framework, Laser Guard, is proposed. Specifically, preprocessing methods are used to approximate the laserperturbed areas, followed by a statistics-based strategy to determine abnormalities in the given samples. This framework can be applied in a plug-and-play manner with DNNs in intelligent vehicles. Extensive experimental results show that the framework can effectively filter out about 70-75% of laser-based street sign AEs, and extends well to other objects, successfully filtering out 80%. The detection latency of objects AEs is marginal, with the average detection time for laser spots being approximately 24 ms, and for laser beams, it is around 57 ms.</p><p>INDEX TERMS Deep learning, adversarial attacks, detection-based defense, laser-based attacks, preprocessing.</p></div> (10.1109/ACCESS.2025.3540653)
  DOI : 10.1109/ACCESS.2025.3540653
• Ambiguity and Invariance in Machine Listening
  
  • Perera David
  , 2025. Machine listening is a growing field with applications in security (audio surveillance), health (sound-based diagnosis), transportation (autonomous driving), manufacturing (predictive maintenance), and bioacoustics (ecosystem tracking). It addresses tasks like sound event detection, sound source localization, and speech separation. This thesis tackles two key challenges: first, the lack of training data in this field, which hinders deep neural networks, typically most effective using large data sets; second, the ambiguity in many tasks, where input-target relations are non-deterministic, which challenges the use of single-prediction models. To address the data shortage, we apply semi supervised invariance-based learning, which penalizes model variations near training data and enforces invariance, enhancing data efficiency and generalization capabilities. Using sound event detection as a case study, we investigate the impact of different data augmentations, their intensity, and the layer of the neural network used for penalization.To tackle ambiguity, we use Multiple Choice Learning (MCL), a framework that trains a multi-head neural network to produce a small set of plausible and diverse predictions, using a competitive training scheme that promotes the specialization of the predictions in different regions of the prediction space. We investigate the efficiency of MCL for machine listening, addressing two key challenges. First, MCL suffers from hypothesis collapse, where some network heads stop being updated by gradient descent. We mitigate this by introducing annealing, which ensures that collapsed heads receive gradients and guides the optimization toward better solutions. Second, we extend MCL’s discrete predictions to create a non-sparse estimator of the target probability distribution. We find that usual approaches fail to converge to the true target distribution when the number of predictions grows large. We identify the cause of this issue—kernel density leakage between heads—and propose kernel truncation as a solution, proving that it guarantees the convergence of the estimators. These methods are shown to improve performance in speech separation tasks. (10.70675/164092fdz865cz4fe7za2a9z2f88f6f38360)
  DOI : 10.70675/164092fdz865cz4fe7za2a9z2f88f6f38360
• POLSAR2POLSAR: A SEMI-SUPERVISED DESPECKLING ALGORITHM FOR POLARIMETRIC SAR IMAGES
  
  • Mendes Cristiano Ulondu
  • Dalsasso Emanuele
  • Zhang Yi
  • Denis Loïc
  • Tupin Florence
  ISPRS Journal of Photogrammetry and Remote Sensing, Elsevier, 2025, 220 (0924-2716), pp.783-798. <div><p>Polarimetric Synthetic Aperture Radar (PolSAR) imagery is a valuable tool for Earth observation. This imaging technique finds wide application in various fields, including agriculture, forestry, geology, and disaster monitoring. However, due to the inherent presence of speckle noise, filtering is often necessary to improve the interpretability and reliability of PolSAR data. The effectiveness of a speckle filter is measured by its ability to attenuate fluctuations without introducing artifacts or degrading spatial and polarimetric information. Recent advancements in this domain leverage the power of deep learning. These approaches adopt a supervised learning strategy, which requires a large amount of speckle-free images that are costly to produce. In contrast, this paper presents PolSAR2PolSAR, a semi-supervised learning strategy that only requires, from the sensor under consideration, pairs of noisy images of the same location and acquired in the same configuration (same incidence angle and mode as during the revisit of the satellite on its orbit). Our approach applies to a wide range of sensors. Experiments on Radarsat-2 and RCM data demonstrate the capacity of the proposed method to effectively reduce speckle noise and retrieve fine details. The code of the trained models is made freely available at https://gitlab.telecom-paris.fr/ring/polsar2polsar. The repository additionally contains a model fine-tuned on SLC PolSAR images from NASA's UAVSAR sensor.</p></div>
• Graph-based Moving Object Segmentation for underwater videos using semi-supervised learning
  
  • Kapoor Meghna
  • Prummel Wieke
  • Giraldo Jhony
  • Subudhi Badri Narayan
  • Zakharova Anastasia
  • Bouwmans Thierry
  • Bansal Ankur
  Computer Vision and Image Understanding, Elsevier, 2025, 252, pp.104290. Moving object segmentation (MOS) using passive underwater image processing is an important technology for monitoring marine habitats. It aids marine biologists studying biological oceanography and the associated fields of chemical, physical, and geological oceanography to understand marine organisms. Dynamic backgrounds due to marine organisms like algae and seaweed, and improper illumination of the environment pose challenges in detecting moving objects in the scene. Previous graph-learning methods have shown promising results in MOS, but are mostly limited to terrestrial surface videos such as traffic video surveillance. Traditional object modeling fails in underwater scenes, due to fish shape and color degradation in motion and the lack of extensive underwater datasets for deep-learning models. Therefore, we propose a semi-supervised graph-learning approach (GraphMOS-U) to segment moving objects in underwater environments. Additionally, existing datasets were consolidated to form the proposed Teleost Fish Classification Dataset, specifically designed for fish classification tasks in complex environments to avoid unseen scenes, ensuring the replication of the transfer learning process on a ResNet-50 backbone. GraphMOS-U uses a six-step approach with transfer learning using Mask R-CNN and a ResNet-50 backbone for instance segmentation, followed by feature extraction using optical flow, visual saliency, and texture. After concatenating these features, a k-NN Graph is constructed, and graph node classification is applied to label objects as foreground or background. The foreground nodes are used to reconstruct the segmentation map of the moving object from the scene. Quantitative and qualitative experiments demonstrate that GraphMOS-U outperforms state-of-the-art algorithms, accurately detecting moving objects while preserving fine details. The proposed method enables the use of graph-based MOS algorithms in underwater scenes. (10.1016/j.cviu.2025.104290)
  DOI : 10.1016/j.cviu.2025.104290
• LayerFold: A Python library to reduce the depth of neural networks
  
  • Pilo Giommaria
  • Hezbri Nour
  • Pereira E Ferreira André
  • Quétu Victor
  • Tartaglione Enzo
  SoftwareX, Elsevier, 2025, 29, pp.102030. Large-scale models are the backbone of Computer Vision and Natural Language Processing, and their generalizability allows for transfer learning and deployment in different scenarios. However, their large size means that reducing their computational and memory demands remains a challenge. Recent research proposes to achieve “layer collapse”, a condition where multiple layers can be combined due to the collapse of non-linearities to linear operators. While this is an important discovery, most studies remain theoretical, often replacing non-linearities with simple identity functions and not providing a real implementation of the more compact architecture. Our contribution is LayerFold, a library that studies and implements the merging of collapsed layers. We address typical cases, from fully connected to convolutional layers, discussing constraints and prospective challenges. Our tests on edge devices reveal that merely reducing network depth does not always result in faster computation, even when GPU-equipped. This work raises important warnings and opens the door to further advances in efficient model deployment. (10.1016/j.softx.2024.102030)
  DOI : 10.1016/j.softx.2024.102030
• Learning on graphs : from algorithms to socio-technical analyses on AI
  
  • Delarue Simon
  , 2025. This thesis addresses the dual challenge of advancing Artificial Intelligence (AI) methods while critically assessing their societal impact. With AI technologies now embedded in high-stake decision sectors like healthcare and justice, their growing influence demands thorough examination, reflected in emerging international regulations such as the AI Act in Europe. To address these challenges, this work leverages attributed-graph based methods and advocates for a shift from performance-focused AI models to approaches that also prioritise scalability, simplicity, and explainability.The first part of this thesis develops a toolkit of attributed graph-based methods and algorithms aimed at enhancing AI learning techniques. It includes a software contribution that leverages the sparsity of complex networks to reduce computational costs. Additionally, it introduces non-neural graph models for node classification and link predictions tasks, showing how these methods can outperform advanced neural networks while being more computationally efficient. Lastly, it presents a novel pattern mining algorithm that generates concise, human-readable summaries of large networks. Together, these contributions highlight the potential of these approaches to provide efficient and interpretable solutions to AI's technical challenges.The second part adopts an interdisciplinary approach to study AI as a socio-technical system. By framing AI as an ecosystem influenced by various stakeholders and societal concerns, it uses graph-based models to analyse interactions and tensions related to explainability, ethics, and environmental impact. A user study explores the influence of graph-based explanations on user perceptions of AI recommendations, while the building and analysis of a corpus of AI ethics charters and manifestos quantifies the roles of key actors in AI governance. A final study reveals that environmental concerns in AI are primarily framed technically, highlighting the need for a broader approach to the ecological implications of digitalisation. (10.70675/9b959076z9900z480ez90b4z48f18740c846)
  DOI : 10.70675/9b959076z9900z480ez90b4z48f18740c846
• Model-checking Strategic Abilities in Information-sharing Systems
  
  • Belardinelli Francesco
  • Boureanu Ioana
  • Dima Catalin
  • Malvone Vadim
  ACM Transactions on Computational Logic, Association for Computing Machinery, 2025, 26 (1), pp.1-45. We introduce a subclass of concurrent game structures (CGS) with imperfect information in which agents are endowed with private data-sharing capabilities. Importantly, our CGSs are such that it is still decidable to model-check these CGSs against a relevant fragment of ATL. These systems can be thought as a generalization of architectures allowing information forks, that is, cases where strategic abilities lead to certain agents outside a coalition privately sharing information with selected agents inside that coalition. Moreover, in our case, in the initial states of the system, we allow information forks from agents outside a given set \(A\) to agents inside this group \(A\) . For this reason, together with the fact that the communication in our models underpins a specialized form of broadcast, we call our formalism \(A\) -cast systems . To underline, the fragment of ATL for which we show the model-checking problem to be decidable over \(A\) -cast is a large and significant one; it expresses coalitions over agents in any subset of the set \(A\) . Indeed, as we show, our systems and this ATL fragments can encode security problems that are notoriously hard to express faithfully: terrorist-fraud attacks in identity schemes. (10.1145/3704919)
  DOI : 10.1145/3704919
• Strong Converse for Classical-Quantum Degraded Broadcast Channels
  
  • Cheng Hao-Chung
  • Datta Nilanjana
  • Rouzé Cambyse
  , 2019. We consider the transmission of classical information through a degraded broadcast channel, whose outputs are two quantum systems, with the state of one being a degraded version of the other. Yard et al. proved that the capacity region of such a channel is contained in a region characterized by certain entropic quantities. We prove that this region satisfies the strong converse property, that is, the maximal probability of error incurred in transmitting information at rates lying outside this region converges to one exponentially in the number of uses of the channel. In establishing this result, we prove a second-order Fano-type inequality, which might be of independent interest. A powerful analytical tool which we employ in our proofs is the tensorization property of the quantum reverse hypercontractivity for the quantum depolarizing semigroup. (10.48550/arXiv.1905.00874)
  DOI : 10.48550/arXiv.1905.00874
• GraphRAG: Leveraging Graph-Based Efficiency to Minimize Hallucinations in LLM-Driven RAG for Finance Data
  
  • Barry Mariam
  • Caillaut Gaëtan
  • Halftermeyer Pierre
  • Qader Raheel
  • Mouayad Mehdi
  • Cariolaro Dimitri
  • Deit Fabrice Le
  • Gesnouin Joseph
  , 2025. This study explores the integration of graphbased methods into Retrieval-Augmented Generation (RAG) systems to enhance efficiency, reduce hallucinations, and improve explainability, with a particular focus on financial and regulatory document retrieval. We propose two strategies-FactRAG and HybridRAG-which leverage knowledge graphs to improve RAG performance. Experiments conducted using Finance Bench, a benchmark for AI in finance, demonstrate that these approaches achieve a 6% reduction in hallucinations and an 80% decrease in token usage compared to conventional RAG methods. Furthermore, we evaluate HybridRAG by comparing the Digital Operational Resilience Act (DORA) from the European Union with the Federal Financial Institutions Examination Council (FFIEC) guidelines from the United States. The results reveal a significant improvement in computational efficiency, reducing contradiction detection complexity from O(n 2 ) to O(k •n)-where n is the number of chunks-and a remarkable 734-fold decrease in token consumption. Graph-based retrieval methods can improve the efficiency and cost-effectiveness of large language model (LLM) applications, though their performance and token usage depend on the dataset, knowledge graph design, and retrieval task.
• Rapid thermalization of dissipative many-body dynamics of commuting Hamiltonians
  
  • Kochanowski Jan
  • Alhambra Alvaro
  • Capel Angela
  • Rouzé Cambyse
  , 2024. Quantum systems typically reach thermal equilibrium rather quickly when coupled to a thermal environment. The usual way of bounding the speed of this process is by estimating the spectral gap of the dissipative generator. However the gap, by itself, does not always yield a reasonable estimate for the thermalization time in many-body systems: without further structure, a uniform lower bound on it only constrains the thermalization time to grow polynomially with system size. Here, instead, we show that for a large class of geometrically-2-local models of Davies generators with commuting Hamiltonians, the thermalization time is much shorter than one would naïvely estimate from the gap: at most logarithmic in the system size. This yields the so-called rapid mixing of dissipative dynamics. The result is particularly relevant for 1D systems, for which we prove rapid thermalization with a system size independent decay rate only from a positive gap in the generator. We also prove that systems in hypercubic lattices of any dimension, and exponential graphs, such as trees, have rapid mixing at high enough temperatures. We do this by introducing a novel notion of clustering which we call "strong local indistinguishability" based on a max-relative entropy, and then proving that it implies a lower bound on the modified logarithmic Sobolev inequality (MLSI) for nearest neighbour commuting models. This has consequences for the rate of thermalization towards Gibbs states, and also for their relevant Wasserstein distances and transportation cost inequalities. Along the way, we show that several measures of decay of correlations on Gibbs states of commuting Hamiltonians are equivalent, a result of independent interest. At the technical level, we also show a direct relation between properties of Davies and Schmidt dynamics, that allows to transfer results of thermalization between both. (10.48550/arXiv.2404.16780)
  DOI : 10.48550/arXiv.2404.16780
• Quasi-optimal sampling from Gibbs states via non-commutative optimal transport metrics
  
  • Capel Ángela
  • Gondolf Paul
  • Kochanowski Jan
  • Rouzé Cambyse
  , 2024. We study the problem of sampling from and preparing quantum Gibbs states of local commuting Hamiltonians on hypercubic lattices of arbitrary dimension. We prove that any such Gibbs state which satisfies a clustering condition that we coin decay of matrix-valued quantum conditional mutual information (MCMI) can be quasi-optimally prepared on a quantum computer. We do this by controlling the mixing time of the corresponding Davies evolution in a normalized quantum Wasserstein distance of order one. To the best of our knowledge, this is the first time that such a non-commutative transport metric has been used in the study of quantum dynamics, and the first time quasi-rapid mixing is implied by solely an explicit clustering condition. Our result is based on a weak approximate tensorization and a weak modified logarithmic Sobolev inequality for such systems, as well as a new general weak transport cost inequality. If we furthermore assume a constraint on the local gap of the thermalizing dynamics, we obtain rapid mixing in trace distance for interactions beyond the range of two, thereby extending the state-of-the-art results that only cover the nearest neighbor case. We conclude by showing that systems that admit effective local Hamiltonians, like quantum CSS codes at high temperature, satisfy this MCMI decay and can thus be efficiently prepared and sampled from. (10.48550/arXiv.2412.01732)
  DOI : 10.48550/arXiv.2412.01732
• RF Indoor mapping using measurements, sensors network, simulations and artificial intelligence
  
  • Mulugeta Biruk Ashenafi
  , 2025. The rapid advancements in wireless technologies to meet the escalating demand for wireless data traffic, have also raised the public's risk perception due to the increased emission of electromagnetic fields (EMFs). In response to these concerns, international guidelines have been established to avoid overexposure that can potentially induce adverse health effects.Given that people spend more than 70% of their daily time in indoor environments, efforts to assess indoor RF-EMF exposure have intensified. However, assessment of indoor exposure to RF-EMF using a deterministic approach is a challenging and time consuming task as it is affected by uncertainties due to the complex interaction of EMFs with different obstructions whose geometry and dielectric properties are unknown in a dynamic indoor environment. In addition, the available base station information is limited.In this study, we characterize the spatial and temporal indoor RF-EMF exposure in three-level buildings by developing statistical and AI-based models. The buildings are located in more than 200 meters vicinity of outdoor cellular antennas supporting several cellular technologies (2G, 3G, 4G, and 5G). Different measurement campaigns have been conducted to develop the statistical and AI-based models. In addition, RF sensors, that monitors the temporal variation of the total indoor RF-EMF exposure, has been installed in the rear offices on each floor of the buildings to provide a representative measurement.We identify the key parameters that characterize the indoor RF-EMF exposure by performing a one-sample Kolmogorov-Smirnov (K-S) test on exposure measurement data. The result of this statistical test indicates that the indoor RF-EMF exposure on each floor, along the length of a wing, can be modeled using a Gaussian (normal) distribution. The result implies that the exposure levels can be fully characterized by the mean and standard deviation parameters.The evaluation of autocorrelation of long-term sensor measurement data indicates that the time series data exhibit a clear daily periodicity, which is consistent over various weeks, months, seasons, and years. We develop robust polynomial regression and AI-based models using long-term time series sensor measurement data to capture the 24-hour time variability of total indoor RF-EMF exposure. The models provide a good high-level representation of hourly fluctuations in the mean exposure levels of the 24-hour temporal variation.We develop novel multivariate polynomial and linear regression models to predict the mean and standard deviation of the total RF-EMF exposure level, respectively. To more effectively capture the impact of antenna beam gain, we present an innovative approach that segments the main antenna beam into smaller subbeams. This innovative technique allows the models to capture more accurately the differences in beam gain experienced by each wing of the building. The beam factor has been specifically designed to summarize the key information about the base station antennas. This parameter enables the models to adapt whenever there are changes in the infrastructure of the cellular base stations. The measurements of the reference sensors provide real-time input to the models. The statistical models serve as a powerful RF mapping tool by estimating both the mean and standard deviation of the total RF-EMF exposure levels on all floors of the buildings.
• STanH : Parametric Quantization for Variable Rate Learned Image Compression
  
  • Presta Alberto
  • Tartaglione Enzo
  • Fiandrotti Attilio
  • Grangetto Marco
  IEEE Transactions on Image Processing, Institute of Electrical and Electronics Engineers, 2025, 34, pp.639-651. In end-to-end learned image compression, encoder and decoder are jointly trained to minimize a R+λD cost function, where λ controls the trade-off between rate of the quantized latent representation and image quality. Unfortunately, a distinct encoder-decoder pair with millions of parameters must be trained for each λ , hence the need to switch encoders and to store multiple encoders and decoders on the user device for every target rate. This paper proposes to exploit a differentiable quantizer designed around a parametric sum of hyperbolic tangents, called STanH, that relaxes the step-wise quantization function. STanH is implemented as a differentiable activation layer with learnable quantization parameters that can be plugged into a pre-trained fixed rate model and refined to achieve different target bitrates. Experimental results show that our method enables variable rate coding with comparable efficiency to the state-of-the-art, yet with significant savings in terms of ease of deployment, training time, and storage costs. (10.1109/TIP.2025.3527883)
  DOI : 10.1109/TIP.2025.3527883
• Zero-Knowledge Proofs of Quantumness
  
  • Phan Duong Hieu
  • Wen Weiqiang
  • Yan Xingyu
  • Zheng Jinwei
  IACR Communications in Cryptology, International Association for Cryptologic Research (IACR), 2025, 1 (4), pp.1-19. With the rapid development of quantum computers, proofs of quantumness have recently become an interesting and intriguing research direction. However, in all current schemes for proofs of quantumness, quantum provers almost invariably face the risk of being maliciously exploited by classical verifiers. In fact, through malicious strategies in interaction with quantum provers, classical verifiers could solve some instances of hard problems that arise from the specific scheme in use. In other words, malicious verifiers can break some schemes (that quantum provers are not aware of) through interaction with quantum provers. All this is due to the lack of formalization that prevents malicious verifiers from extracting useful information in proofs of quantumness. To address this issue, we formalize zero-knowledge proofs of quantumness. Intuitively, the zero-knowledge property necessitates that the information gained by the classical verifier from interactions with the quantum prover should not surpass what can be simulated using a simulated classical prover interacting with the same verifier. As a result, the new zero-knowledge notion can prevent any malicious verifier from exploiting quantum advantage. Interestingly, we find that the classical zero-knowledge proof is sufficient to compile some existing proofs of quantumness schemes into zero-knowledge proofs of quantumness schemes. Due to some technical reason, it appears to be more general to require zero-knowledge proof on the verifier side instead of the prover side. Intuitively, this helps to regulate the verifier's behavior from malicious to be honest-but-curious. As a result, both parties will play not only one role in the proofs of quantumness but also the dual role in the classical zero-knowledge proof. Specifically, the two principle proofs of quantumness schemes: Shor's factoring-based scheme and learning with errors-based scheme in [Brakerski et al, FOCS, 2018], can be transformed into zero-knowledge proofs of quantumness by requiring an extractable non-interactive zero-knowledge argument on the verifier side. Notably, the zero-knowledge proofs of quantumness can be viewed as an enhanced security notion for proofs of quantumness. To prevent malicious verifiers from exploiting the quantum device's capabilities or knowledge, it is advisable to transition existing proofs of quantumness schemes to this framework whenever feasible. (10.62056/ayiv4fe-3)
  DOI : 10.62056/ayiv4fe-3
• Masked Computation of the Floor Function and Its Application to the FALCON Signature
  
  • Berthet Pierre-Augustin
  • Paillet Justine
  • Tavernier Cédric
  • Colombier Brice
  • Bossuet Lilian
  IACR Communications in Cryptology, International Association for Cryptologic Research (IACR), 2025, 1 (4), pp.1-23. FALCON is a signature selected for standardisation of the new Post-Quantum Cryptography (PQC) primitives by the National Institute of Standards and Technology (NIST). However, it remains a challenge to define efficient countermeasures against side-channel attacks (SCA) for this algorithm. FALCON is a lattice-based signature that relies on rational numbers, which is unusual in the cryptography field. Although recent work proposed a solution to mask the addition and the multiplication, some roadblocks remain, most noticeably, how to protect the floor function. In this work, we propose to complete the first existing tests of hardening FALCON against SCA. We perform the mathematical proofs of our methods as well as formal security proofs in the probing model by ensuring Multiple Input Multiple Output Strong Non-Interference (MIMO-SNI) security. We provide performances on a laptop computer of our gadgets as well as of a complete masked FALCON. We notice significant overhead in doing so and discuss the deployability of our method in a real-world context. (10.62056/ay73zl7s)
  DOI : 10.62056/ay73zl7s