Publications

2026

• MEDIATE: multi-faceted implementation of a mixed software/hardware-based zero trust framework for the computing continuum
  
  • Fournaris Apostolos P
  • Haleplidis Evangelos
  • Abdoul-Soukour Shahin
  • Huang Chih-Kai
  • Khoder Niemat
  • Bouloukakis Georgios
  • Brokalakis Andreas
  • Georgopoulos Konstantinos
  • Ioannidis Sotiris
  , 2026. This paper introduces MEDIATE, a framework that brings zero trust principles to the Cloud-Edge-IoT Computing Continuum. MEDIATE integrates several software-based security and policy services operating at the cloud and edge layers with hardware-and software-based monitoring components as IoT devices deployed close to the assets that require security protection. The framework is designed to a three-layer architecture with Orchestrator (Cloud), Overwatch (Edge), and Sentinel (IoT) that collectively enables cross-layer policy enforcement and coordinated threat mitigation. The Orchestrator provides global view and policy consistency, the Overwatch performs local analysis and threat correlation, and the Sentinel acts as lightweight agent for real-time monitoring near critical assets. These components can provide an adaptive security foundation suitable for distributed and large-scale operational domains such as Fourth-Party Logistics (4PL).
• Efficient 3D Deep Learning for Joint DBT Reconstruction and Faithful Uncertainty Estimation: Preliminary Clinical Evaluation
  
  • Quillent Arnaud
  • Bismuth Vincent
  • Bloch Isabelle
  • Kervazo Christophe
  • Ladjal Saïd
  , 2026. Purpose: Digital breast tomosynthesis (DBT) reconstruction is an ill-posed inverse problem due to the limited-angle acquisition and sparse projection sampling, leading to streak artefacts and poor depth resolution in the reconstructed volumes. Although deep learning provides data-driven priors, its application remains hindered by the lack of ground-truth clinical data and concerns regarding model reliability. This study introduces a 3D deep learning framework trained on virtual phantoms to jointly reconstruct DBT volumes and estimate voxel-wise uncertainty. Methods: We generate a synthetic database of phantom-reconstruction pairs derived from breast computed tomography (CT) images on which we train a post-processing 3D deep ensemble to estimate the reconstructed volume. To ensure reliability, we implement faithful regression and decouple the estimation of uncertainty from the main reconstruction task, mitigating training instability. The approach is then evaluated on the synthetic test set along with textured phantoms and clinical images. Results: On synthetic data, our method improves peak signal-to-noise ratio (PSNR) by 30%, normalised root mean squared error (NRMSE) by 50%, and structural similarity index measure (SSIM) by 43% compared to baseline iterative reconstruction. The qualitative analysis demonstrates significantly reduced tissue superposition and improved depth resolution in coronal and sagittal planes. Additionally, the uncertainty maps correlate well with the reconstruction error, but they slightly underestimate its true magnitude. A preliminary clinical evaluation yields plausible glandular localisation, although domain shift introduces some artefacts and texture mismatch. Lastly, we discuss the potential clinical value of uncertainty estimation as a tool to guide radiologists' confidence in reconstructed volumes. Conclusion: We successfully demonstrate a robust 3D deep learning approach for DBT that improves reconstruction quality while providing faithful uncertainty estimates. The model generalises well to synthetic phantoms, with promising translation to clinical images.
• Towards Reliable and Secure RISC-V Systems: Survey of Testability and Security Mechanisms
  
  • Khan Mahreen
  • Mushtaq Maria
  • Apvrille Ludovic
  , 2026. <div><p>RISC-V has emerged as a versatile open-source instruction set architecture, enabling extensible microarchitectures, custom accelerators, and domain-specific processors. Its openness facilitates innovation in testability, safety, and security for safety-critical and security-sensitive applications. This survey provides a comprehensive review of recent research in RISC-V verification and protection mechanisms. We analyze AI-assisted test generation, statistical fault injection frameworks, systemlevel testing, design-for-test architectures, and hardware-software co-verification methods. In the safety domain, we discuss temporal isolation, performance monitoring, debug support, and fault containment strategies. Security mechanisms, including trusted execution environments, memory protection, cryptographic ISA extensions, post-quantum acceleration, and secure debug practices, are evaluated. Open challenges in scalable test coverage, AI-enabled certification, side-channel resilience, and lifecycle management are highlighted. Finally, we outline future research directions that leverage RISC-V's modularity and openness to enable trustworthy computing systems in automotive, aerospace, telecommunications, and edge computing domains.</p></div>
• Joint Visibility Analysis of RIS in Non-Terrestrial Networks through Stochastic Geometry
  
  • Balakrishnan Ashutosh
  • Lee Junse
  • Baccelli François
  , 2026. Non-Terrestrial networks (NTNs) are a key theme in upcoming 6G communications, especially for ubiquitous coverage. Urban environments, comprising of high rise buildings often result in blocking the line of sight (LoS) path between the user equipment (UE) and the NTN base station (NTN-BS). In this paper we investigate the situation where reconfigurable intelligent surfaces (RIS) are deployed on the building roof-tops to ensure multi-hop connectivity between the UE and the NTN-BS. In such a scenario, it becomes crucial to statistically study the LoS visibility of the RIS from the UE as well as from the NTN-BS, hence termed as joint visibility. In this work, accounting for the dual stochasticity arising from the locations of the RIS deployed buildings and the respective random building heights, we statistically study the probability of joint RIS visibility in a two-dimensional (2D) scenario considering a deterministic location of the NTN-BS. Further, we study the joint RIS visibility statistics conditional on the UE-NTN link being LoS or non-LoS. For the RISs deployed as a point point process (PPP) having exponentially distributed heights, the expected RISs jointly visible under the unconditional and conditional geometric settings are derived in closed form. Interestingly, in the 2D setting, the maximum expected RISs jointly visible, unconditionally, is twice the Basel number (π^2 /6). The simulated results are analyzed over building density, average building height, the altitude and position of the NTN-BS. We also illustrate probability heatmaps, demonstrating the strongest chance to have a RIS used conditioned on the system geometry. This study is expected to be useful in planning the deployment of RIS in urban areas, improving the signal and for assessing economic aspects.
• Transitions as the Native Objects of Dispersive Light-Matter Dynamics
  
  • Mohamed Meguebel
  • Federico Maxime
  • Garbe Louis
  • Belabas Nadia
  • Fabre Nicolas
  , 2026. We introduce a framework where light-matter transitions, rather than states, are the primary dynamical objects. Successive compositions of elementary transitions yield multiphoton processes with compact diagrammatic bookkeeping of resonant and off-resonant pathways. This approach enables transparent derivations of effective high-order Hamiltonians in the dispersive regime, foundational to quantum-information applications. Applied to the paradigmatic Jaynes-Cummings model, our framework reveals a photon-number-independent intrinsic Rabi frequency and persistent polaritonic hybridization in the dispersive regime, unifying resonant and dispersive limits.
• Time-frequency Talbot effect as Clifford operations on entangled time-frequency GKP states
  
  • Pousset Thomas
  • Dalidet Romain
  • Labonté Laurent
  • Fabre Nicolas
  , 2026. The Talbot effect-a near-field diffraction phenomenon in which a periodic wavefront self-images at regular distances-can be transposed to the time-frequency domain via the space-time duality between diffraction and dispersive broadening. We exploit this analogy to define the time-frequency (TF) Talbot effect and show that it implements different Clifford operations on TF Gottesman-Kitaev-Preskill (TF-GKP) qubits (Phys. Rev. A 102, 012607), a class of qubit states encoded in the discretised frequency and time-of-arrival degrees of freedom of entangled photon pairs, whose logical basis corresponds to even and odd components of an entangled frequency combs. These states are intrinsically robust against small frequency and temporal displacements, which can be further corrected by linear or nonlinear quantum error-correction schemes. We analyse the role of the comb envelope and peak width relative to the free spectral range, and show that a compromise must be made between the gate fidelity of the Clifford gates induced by TF-Talbot operation and the error-correction capacity of the code. We then demonstrate that the signature of the TF-Talbot effect is directly accessible via the generalised Hong-Ou-Mandel interferometer: all six logical GKP states can be unambiguously distinguished by introducing a frequency shift of half the comb periodicity in one interferometer arm. We conclude with a feasibility analysis based on current experimental technology, identifying the comb finesse as the key figure of merit for both gate performance and correctability. This conclusion extends naturally to quadrature GKP states, where a shear in quadrature phase space is precisely a Talbot effect. (10.1103/s8jp-l4hf)
  DOI : 10.1103/s8jp-l4hf
• Effective Hamiltonians in Cavity and Waveguide QED from Transition-Operator Diagrammatic Perturbation Theory
  
  • Meguebel Mohamed
  • Federico Maxime
  • Garbe Louis
  • Belabas Nadia
  • Fabre Nicolas
  , 2026. We propose an adiabatic-elimination formalism in the dispersive regime based on a transitioncentric perturbation theory. The perturbative expansion is recast into a diagrammatic framework, while adiabatic elimination is implemented through controlled projections onto transition subspaces. Our approach applies systematically at arbitrary perturbation order, and is suited to multilevel systems and multiple qubits in both cavity and waveguide quantum electrodynamics. It ultimately enables the explicit construction of effective higher-order Hamiltonians while bypassing important limitations of existing techniques, thereby providing a practical toolbox for multiphoton processes in the dispersive regime.
• Methods for Sizing and Deploying Virtualised Mobile Networks : Application to 5G/6G, Optimisation for Smart Electrical Grids
  
  • Tawil Rosy Al-
  , 2026. Modernising industrial smart grids operated by EDF requires highly reliable communications for Protection, Automation, and Control (PAC) systems. Although 5G network slicing supports mission-critical messages via URLLC, industrial clients lack visibility into the operator's core network, making it hard to guarantee the required 10−5 packet loss probability. Furthermore, traditional allocation models based on Poisson's law fail to capture real-world traffic burstiness, causing resource deficits 30% of the time. Tobridge this visibility gap, a Machine Learning classification model was developed. A LightGBM model was implemented to classify trafficinto URLLC and eMBB slices using experimental data restricted to the lower PHY and MAC layers. Achieving a 97.6% accuracy demonstrates that industrial clients can autonomously monitor traffic without operator-level KPIs. Building on this, a dynamic resource allocation model using Deep Reinforcement Learning (DRL) was introduced. A Deep Q-Network (DQN) agent continuously interacts with the network environment across a continuous state space (URLLC/eMBB loads, cell capacity) to manage the explorationexploitation trade-off. The agent dynamically allocates the optimal guard channels to minimize URLLC packet loss while maintaining fairness toward the eMBB buffer. Combining queuing limits, ML classification, and DQN, this thesis empowers industrial clients to autonomously predict, dimension, and adjust their 5G network slice resources instantaneously.
• CALICE: Continuous bitrate control with Adapted LIC modEl
  
  • Spadaro Gabriele
  • Presta Alberto
  • Giraldo Jhony
  • Fiandrotti Attilio
  • Grangetto Marco
  • Tartaglione Enzo
  ACM Transactions on Multimedia Computing, Communications and Applications, Association for Computing Machinery, 2026, pp.1-21. Learned image compression (LIC) has drawn much attention recently as it outperforms standardized codecs in rate-distortion (RD) efficiency. However, a LIC model is typically trained for a specific RD trade-off, and achieving a different target rate requires retraining the model and storing the weights as a whole, limiting the practical applicability of LIC. In this paper, we introduce CALICE, a framework for achieving continuous bitrate control by plugging into a pre-trained LIC model a set of modular adapters. Unlike similar methods that require a distinct set of adapters for each target rate, our method achieves continuous bitrate control by modulating a single set of adapters via a scalar parameter \(\boldsymbol{\alpha}\) , with a total overhead of less than \(\mathbf{0.35}\boldsymbol{\%}\) of the parameters of the LIC model. This design enables efficient support for multiple distortion objectives by learning lightweight, distortion-aware adapters. We also extend our strategy beyond rate control, demonstrating its ability to provide fine-grained adaptation of perceptual quality along the distortion-perception trade-off. To our knowledge, this is the first method that jointly addresses rate and perceptual control using a unified, low-cost strategy. We publicly released the code at https://github.com/EIDOSLAB/CALICE . (10.1145/3820658)
  DOI : 10.1145/3820658
• Characterizing Perceived Readability in Data Visualization: Design and Reader Factors
  
  • Cabouat Anne-Flore
  • Kurzhals Kuno
  • Matzen Laura
  • Isenberg Tobias
  • Huron Samuel
  • Isenberg Petra
  , 2026. We report results from an online study designed to characterize perceived readability across a controlled set of static visual data representations.
• Group Conversational Agents
  
  • Yeo Shunyi
  • Zhang Tianyi
  • Bateman Scott
  • Hsieh Gary
  • Kim Young-Ho
  • Perrault Simon Tangi
  • Li Jiannan
  • Tang Anthony
  , 2026, pp.2765-2779. Conversational agents that participate in or mediate group interaction introduce challenges that extend beyond supporting individual users, raising new questions about how agents participate in and influence groups. To characterise this emerging design space, we present a systematic review of 53 peer-reviewed studies on group conversational agents (GCAs). We analyse how GCAs intervene in group-level processes, including participation regulation, conflict mediation, task alignment, and execution support. Using concepts from group research as an analytic lens, we organise prior GCA work around recurring group interactional challenges (orientation, conflict, alignment, and execution), and examine the roles agents are designed to play in addressing these challenges. We find that GCAs are predominantly designed as short-term, role-bounded interventions targeting isolated challenges in bounded interactional contexts. We further identify recurring structural tensions in GCA design, including tradeoffs between visibility and discretion, proactivity and group autonomy, and agent authority and group ownership. Together, these findings clarify how current GCAs are positioned within group interaction, surface the implicit assumptions embedded in their designs, and outline open questions for future research on conversational agents as group-level interventions. (10.1145/3800645.3812934)
  DOI : 10.1145/3800645.3812934
• Parser Instrumentation for Semantic-Aware Applicative Intrusion Detection
  
  • Quetel Grégor
  • Gimenez Pierre-François
  • Robert Thomas
  • Pautet Laurent
  , 2026, 787, pp.359–373. Intrusion Detection Systems (IDS) are common security tools for protecting modern information systems, yet their effectiveness at detecting application-layer attacks is often limited by the semantic gap between low-level host or network observations and the actual behavior of applications. Existing work overlooks the data collection phase and typically focuses on designing complex decision engines and preprocessing functions such as embedding-based representations. Unfortunately, these approaches incur significant computational overhead at inference time and remain brittle against adversarial inputs. In this paper, we present a parser-based instrumentation approach for application-level intrusion detection that provides lexical, syntactic and explicit semantic observation with minimal overhead. We introduce gaur, an implementation for instrumenting parsers, it produces observations during parsing by associating semantic tags to grammar rules, eliminating the need for runtime natural language processing. Our evaluation demonstrates the low overhead and collection time of our data collector. Furthermore, empirical results show that incorporating explicit semantic information into decision engines not only improves detection performance over traditional mechanisms but also enables faster inference and greater robustness than approaches relying on implicit semantic representations. (10.1007/978-3-032-27993-4_25)
  DOI : 10.1007/978-3-032-27993-4_25
• On the Informativeness of Security Commit Messages: A Large-scale Replication Study
  
  • Islam Syful
  • Zacchiroli Stefano
  , 2026. The informativeness of security-related commit messages is crucial for patch triage: when high, it enables the rapid distribution and deployment of security fixes. Prior research (Reis et al., 2023) reported, however, that commit messages are often too uninformative to support these activities. To assess the robustness of this negative result, we independently replicate the original study using only the information provided in the paper, without reusing any of the original artifacts (data, analysis pipeline, etc.). Unlike the original study, we source commit data not only from GitHub, but from the entire Software Heritage archive, which includes projects hosted on many other platforms and using multiple version control systems. We retrieve 50673 security-related commits and analyze their informativeness using an independent re-implementation of the techniques introduced by Reis et al. For the same source (i.e., GitHub) and time period (from June 1999 to August 2022) as the original study, our replication confirms the original findings in a statistically significant way: security-related commit messages are, in general, not informative enough for security-focused purposes. We then extend the original study in several ways. Over a longer time period (from June 1999 to October 2025), we find that commit-message informativeness is worsening. Breaking results down by software ecosystem (Linux kernel, Ubuntu, Go, PyPI, etc.), we observe significant differences in informativeness. Finally, we examine emerging best practices for writing commit messages, such as the Conventional Commits Specification (CCS), and again find significant differences in an unexpected direction: CCS-compliant commits are less informative than non-compliant ones. Our findings highlight the need for cross-ecosystem analyses to understand platform- and community-specific commit-message practices, and to inform the development and adoption of universally applicable guidelines for writing informative security-related commit messages. (10.1145/nnnnnnn.nnnnnnn)
  DOI : 10.1145/nnnnnnn.nnnnnnn
• On the Use of Commit Messages for Corrective Software Maintenance: A Systematic Mapping Study
  
  • Islam Syful
  • Zacchiroli Stefano
  , 2026. Corrective maintenance is crucial to ensure the quality of software, thereby improving reliability and user experience. In a version control system (VCS), developers write commit messages to document their changes and support later maintenance. Therefore, the utilization of commit messages to accomplish corrective maintenance has become a common practice among software engineering practitioners and researchers. Still, to this day, no secondary study has mapped the research landscape of how commit messages have been used in corrective software maintenance. We present a systematic mapping study of 97 primary sources published between 2004 and May 2025, where we examine the goals, potential utilization of source code artifacts along with commit messages, methodologies, stakeholders, and the key findings about their influence on corrective maintenance. Our analysis reveals a growing interest in the usage of commit messages to perform corrective maintenance tasks, in particular for bug analysis and bug fix identification goals. Surprisingly few studies address other themes such as automated program repair, security development practices, etc. We find that the software artifacts most used in combination with commit messages are commit "diffs" and that repository mining, together with natural language processing (NLP) and artificial intelligence/machine learning (AI/ML) are the methodological foundations of studies in this field. Among stakeholders considered in previous studies, developers play the most important role in shaping corrective maintenance practices. Key findings in previous studies about commit messages establish their significant role in corrective maintenance, due to the fact that they carry crucial information helpful for stakeholders to understand and improve the code base through the software evolution process. Often, though, commit messages lack important information and are not enough to convey the intent of code changes to future readers. Therefore, developers should be aware of commit message contextual richness while committing code changes in VCS. (10.1145/nnnnnnn.nnnnnnn)
  DOI : 10.1145/nnnnnnn.nnnnnnn
• Beyond MERLIN: A Multi-Scale Self-Supervised Framework for Despeckling
  
  • Basso-Lacroix Stella
  • Bultingaire Thomas
  • Denis Loïc
  • Gasnier Nicolas
  • Desroches Damien
  • Tupin Florence
  , 2026. Synthetic aperture radar (SAR) images are severely affected by speckle. Numerous deep learning techniques have been developed to restore image quality. Among them, the self-supervised MERLIN strategy achieves state-of-the-art performance using only single-look complex (SLC) data for training, without imposing any particular constraint on the network architecture. While the method produces excellent results when properly trained, training the network can be challenging. The approach proposed in this study extends the MERLIN strategy by exploiting several resolution levels to increase the robustness and performance of the despeckling method. Low-resolution images are generated using a multilook approach, which mitigates speckle fluctuations and simplifies subsequent processing. The knowledge obtained from models trained at low resolution is then distilled into a higher-resolution network to effectively guide the training phase.
• Generalization of InSAR2InSAR to Sentinel-1 Multi-Channel Data: PolSAR and PolInSAR
  
  • Geara Carla
  • Gelas Colette
  • De Vitry Louis
  • Colin Elise
  • Tupin Florence
  , 2026. Despeckling SAR images is essential for improving data quality and facilitating their use in Earth observation applications. While single-channel SAR despeckling has been extensively studied, multi-channel SAR despeckling remains a challenging task. In this paper, we extend InSAR2InSAR, a semi-supervised InSAR parameter estimation method, to PolSAR and PolInSAR Sentinel-1 acquisitions. We also adapt a previously proposed uncertainty estimation principle, originally introduced for the single-channel despeckling filter MERLIN, to the covariance matrix domain in order to assess the quality of our filter by quantifying the uncertainty of its estimations.
• Convergence rates of Sum-of-Hermitian-Squares Hierarchies for the Pauli algebra
  
  • Almasi Ali
  • Bugár Dávid
  • Rouzé Cambyse
  • Brown Peter
  , 2026. Moment/Sum-of-Hermitian-Squares relaxations for noncommutative polynomial optimization problems have become an important tool for analyzing problems within quantum theory. Despite their widespread success, little is known about their rate of convergence and, consequently, their accuracy. In this work, we develop explicit convergence rates for relaxations of noncommutative polynomial optimization problems generated from the Pauli algebra -- covering applications to the ground state energy problem for n-qubit systems. In particular, we show that the rate of convergence can be bounded in terms of the smallest roots of a family of orthogonal polynomials known as Krawtchouk polynomials. Our result represents the first quantitative analysis of the rate of convergence for relaxations of noncommutative polynomial optimization problems.
• Learning with Importance Weighted Variational Inference
  
  • Daudel Kamélia
  • Roueff François
  , 2026. Several variational bounds involving importance weighting ideas generalize the Evidence Lower BOund (ELBO) for marginal likelihood optimization, such as the Importance-weighted Auto-Encoder (IWAE), Variational Rényi (VR) and VR-IWAE bounds. Yet, it remains unclear how the joint choice of bound and gradient estimator impacts the behavior of the resulting variational inference (VI) algorithms. This paper provides a unified theoretical comparison of reparameterized (REP) and doubly-reparameterized (DREP) gradient estimators tied to the IWAE, VR and VR-IWAE bounds. Through asymptotic analyses of the Signal-to-Noise Ratio as the number of Monter Carlo samples $N$ goes to infinity, we identify a bias-variance tradeoff in these gradient estimators and we formally justify the superiority of DREP over REP in importance-weighted VI. An additional asymptotic analysis for challenging regimes, where both $N$ and the Kullback-Leibler divergence between the variational and posterior densities go to infinity, indicates that importance-weighted VI gradient estimators point in a well-founded direction even when the variational approximation deteriorates. Together, these complementary results characterize the optimization trajectory in importance-weighted VI from poor initialization to final convergence. Importantly, our proof techniques establish general theoretical tools for the study of sample means ratios whose scope extend beyond VI and constitute an independent contribution to the field of Monte Carlo methods.
• Explaining CLIP Zero-shot Predictions Through Concepts
  
  • Ozdemir Onat
  • Christensen Anders
  • Alaniz Stephan
  • Akata Zeynep
  • Akbas Emre
  , 2026. Large-scale vision-language models such as CLIP have achieved remarkable success in zero-shot image recognition, yet their predictions remain largely opaque to human understanding. In contrast, Concept Bottleneck Models provide interpretable intermediate representations by reasoning through human-defined concepts, but they rely on concept supervision and lack the ability to generalize to unseen classes. We introduce EZPC that bridges these two paradigms by explaining CLIP's zero-shot predictions through human-understandable concepts. Our method projects CLIP's joint image-text embeddings into a concept space learned from language descriptions, enabling faithful and transparent explanations without additional supervision. The model learns this projection via a combination of alignment and reconstruction objectives, ensuring that concept activations preserve CLIP's semantic structure while remaining interpretable. Extensive experiments on five benchmark datasets, CIFAR-100, CUB-200-2011, Places365, ImageNet-100, and ImageNet-1k, demonstrate that our approach maintains CLIP's strong zero-shot classification accuracy while providing meaningful concept-level explanations. By grounding open-vocabulary predictions in explicit semantic concepts, our method offers a principled step toward interpretable and trustworthy vision-language models. Code is available at https://github.com/oonat/ezpc.
• On the explainability of max-plus neural networks
  
  • Enaieh Ikhlas
  • Fercoq Olivier
  • Ángel García
  , 2026. We investigate the explanability properties of the recently proposed linear-min-max neural networks. At initialization, they can be interpreted as k-medoids with the infinity norm as a distance. Then, they are trained using subgradient descent to better fit the data. The model has been shown to be a universal approximator. Yet, we can trace the decision process because a single most activated neuron is responsible for the value of the output. Using this property, we designed a pixel fragility measure that determines whether changes to a single pixel may be responsible to a change in the classification output. Experiments on the PneumoniaMnist dataset show that this explanation for the output of the neural network compares favorably to SHAP and Integrated Gradient.
• FINER: MLLMs Hallucinate under Fine-grained Negative Queries
  
  • Xiao Rui
  • Kim Sanghwan
  • Xian Yongqin
  • Akata Zeynep
  • Alaniz Stephan
  , 2026. Multimodal large language models (MLLMs) struggle with hallucinations, particularly with fine-grained queries, a challenge underrepresented by existing benchmarks that focus on coarse image-related questions. We introduce FIne-grained NEgative queRies (FINER), alongside two benchmarks: FINER-CompreCap and FINER-DOCCI. Using FINER, we analyze hallucinations across four settings: multi-object, multi-attribute, multi-relation, and "what" questions. Our benchmarks reveal that MLLMs hallucinate when fine-grained mismatches co-occur with genuinely present elements in the image. To address this, we propose FINER-Tuning, leveraging Direct Preference Optimization (DPO) on FINER-inspired data. Finetuning four frontier MLLMs with FINER-Tuning yields up to 24.2% gains (InternVL3.5-14B) on hallucinations from our benchmarks, while simultaneously improving performance on eight existing hallucination suites, and enhancing general multimodal capabilities across six benchmarks. Code, benchmark, and models are available at https://explainableml.github.io/finer-project/.
• What can we do in a symmetry-constrained perspective? The importance of the total charge's status in quantum reference frame frameworks
  
  • Doat Guilhem
  • Vanrietvelde Augustin
  , 2025. The study of quantum reference frames has received renewed interest over the last years, leading to the parallel development of non-equivalent frameworks by different com- munities. We clarify the differences between these frameworks. At the mathematical level, they mainly differ in the kind of symmetry (either weak or strong) employed to constrain the system. We show that this mathematical difference corresponds to a fundamental physical question: whether the global charge associated to the symmetry group is acces- sible to symmetry-constrained observers. In this context, we formulate a definition of a perspective in terms of operational capacities, or lack thereof. Turning to consequences of adopting either approach, we discuss how adopting the weak approach induces an ambi- guity in the momenta included in each perspective and bars from defining reversible QRF transformations. We then review and analyze the existing arguments motivating each approach, and show how they bear upon the problem of charge accessibility. Finally, we introduce a simple operational scenario in which upholding two reasonable physical pos- tulates leads to the conclusion that internal observers could measure the global charge by 1/ performing a relativized interference measurement and 2/ classically communicating.
• Using Locally Learnt Word Representations for better Textual Anomaly Detection
  
  • Breidenstein Alicia
  • Labeau Matthieu
  , 2024, pp.82-91. <div><p>The literature on general purpose textual Anomaly Detection is quite sparse, as most textual anomaly detection methods are implemented as out of domain detection in the context of pre-established classification tasks. Notably, in a field where pre-trained representations and models are of common use, the impact of the pre-training data on a task that lacks supervision has not been studied. In this paper, we use the simple setting of k-classes out anomaly detection and search for the best pairing of representation and classifier. We show that well-chosen embeddings allow a simple anomaly detection baseline such as OC-SVM to achieve similar results and even outperform deep state-of-the-art models.</p></div> (10.18653/v1/2024.insights-1.11)
  DOI : 10.18653/v1/2024.insights-1.11
• Trust-based attack detection model for connected cars using a Subjective Logic based framework
  
  • Ismail Ahmad
  • Fadlallah Ahmad
  • Bassi Francesca
  • Khatoun Rida
  , 2026. Traditional Voting Classifiers (Hard and Soft voting) assign fixed weights to individual models. In many cases, a machine learning model performs differently depending on the predicted class and the context. Hard and soft voting are static in nature which often prevents the full potential of ensemble methods from being realized. In contrast, Subjective Logic offers a probabilistic framework that effectively accounts for information uncertainty and the trustworthiness of information sources. This paper introduces a novel subjective logic based binary ensemble classifier that takes conflict between models and individual model performance into account to modify model weights in real time and improve the ensemble predictions.
• DELICATE: Diachronic Entity LInking using Classes And Temporal Evidence
  
  • Santini Cristian
  • Barzaghi Sebastian
  • Sernani Paolo
  • Frontoni Emanuele
  • Alam Mehwish
  Journal on Computing and Cultural Heritage, Association for Computing Machinery, 2026. In spite of the remarkable advancements in the field of Natural Language Processing, the task of Entity Linking (EL) remains challenging in the field of humanities due to complex document typologies, lack of domain-specific datasets and models, and long-tail entities, i.e., entities under-represented in Knowledge Bases (KBs). The goal of this paper is to address these issues with two main contributions. The first contribution is DELICATE, a novel neuro-symbolic method for EL on historical Italian which combines a BERT-based encoder with contextual information from Wikidata to select appropriate KB entities using temporal plausibility and entity type consistency. The second contribution is ENEIDE, a multi-domain EL corpus in historical Italian semi-automatically extracted from two annotated editions spanning from the 19th to the 20th century and including literary and political texts. Results show how DELICATE outperforms other EL models in historical Italian even if compared with larger architectures with billions of parameters. Moreover, further analyses reveal how DELICATE confidence scores and features sensitivity provide results which are more explainable and interpretable than purely neural methods.