ICAART 2026 Abstracts

Area 1 - Artificial Intelligence

Full Papers

Paper Nr:	33
Title:	π-NoCCHIO: An Architecture for Context-Aware Normative Reinforcement Learning
Authors:	Benoît Alcaraz
Abstract:	Normative reinforcement learning is a timely topic, as many systems deployed in the real-world are now using Artificial Intelligence (AI). As these systems now evolve among us, we need to ensure that they follow settled norms and conventions in order to avoid causing harm to people they interact with. This paper contributes to the field of normative reinforcement learning by presenting π-NoCCHIO, a novel hybrid architecture for normative reinforcement learning. This architecture is inspired by the combination of the Jiminy architecture with the AJAR framework. Its goal is to use the reasoning capacity of Jiminy to provide a signal that can be used to train a model-free normative RL agent, while still allowing for manual alterations of the norms.
Download

Paper Nr:	35
Title:	Combining Formal Argumentation and Reinforcement Learning: An Hybrid Approach to Machine Ethics
Authors:	Benoît Alcaraz, Rémy Chaput, Olivier Boissier and Christopher Leturc
Abstract:	Our society is increasingly confronted with socio-technical systems embedding Artificial Intelligence techniques, raising questions from the public, academia, and industry about their impact on humans. Systems that are designed to include ethical considerations mostly use top-down formalization or bottom-up learning, each having their respective advantages and drawbacks. We adopt instead a hybrid approach that combines advantages of both. Our proposed hybrid Multi-Agent Reinforcement Learning framework consists of learning agents that learn a task-oriented behavior while respecting some moral values, such as ecology or equity. To ensure that learning agents respect these moral values, we introduce symbolic moral judging agents to the system. They consider judgment functions, which are designed on formal argumentation to evaluate, but also to justify, their judgments about the learning agents’ behaviors, with respect to moral values. The learning agents’ rewards are then defined as an aggregation of the different judgments provided by the judging agents. Finally, this proposal is evaluated on an energy distribution problem within a simulated smart grid. Results show that the learning agents seek to increase each moral value.
Download

Paper Nr:	54
Title:	Construal Level Theory in Explainable AI: Effects of Explanation Formats on Reliance and Performance
Authors:	Daniel Eisenhardt and Hüseyin Hussein Keke
Abstract:	As Artificial Intelligence (AI)-based decision support systems proliferate in critical domains, ensuring appropriate user reliance is essential for effective human-AI collaboration. This need is particularly acute given that most advanced AI systems function as black boxes, where internal decision-making processes remain opaque to users, potentially undermining reliance despite high performance. Explainable AI (XAI) aims to address this challenge, yet understanding how explanation formats influence users’ information processing is critical for designing effective reliance-calibrating systems. This study examines how different XAI presentation formats affect user cognition and reliance. Drawing on construal level theory, we investigated how explanation formats influence information processing orientation in a news classification task. Results demonstrate that visual explanations promote global processing, while textual explanations encourage local processing. This creates a paradox where global processing decreases reliance yet improves performance. These findings provide valuable guidance for designing XAI systems that balance appropriate reliance calibration with enhanced decision quality.
Download

Paper Nr:	61
Title:	Quantum RNNs and LSTMs through Entangling and Disentangling Power of Unitary Transformations
Authors:	Ammar Daskin
Abstract:	In this paper, we present a framework for modeling quantum recurrent neural networks (RNNs) and their enhanced version, long short-term memory (LSTM) networks using the core ideas presented by Linden et al. (2009), where the entangling and disentangling power of unitary transformations is investigated. In particular, we interpret entangling and disentangling power as information retention and forgetting mechanisms in LSTMs. Thus, entanglement emerges as a key component of the optimization (training) process. We believe that, by leveraging prior knowledge of the entangling power of unitaries, the proposed quantum-classical framework can guide the design of better-parameterized quantum circuits for various real-world applications.
Download

Paper Nr:	67
Title:	Human-LLM Synergy in Context-Aware Adaptive Architecture for Scalable Drone Swarm Operation
Authors:	Ahmed R. Sadik, Muhammad Ashfaq, Niko Mäkitalo and Tommi Mikkonen
Abstract:	The deployment of autonomous drone swarms in disaster response missions necessitates the development of flexible, scalable, and robust coordination systems. Traditional fixed architectures struggle to cope with dynamic and unpredictable environments, leading to inefficiencies in energy consumption and connectivity. This paper addresses this gap by proposing an adaptive architecture for drone swarms, leveraging a Large Language Model (LLM) to dynamically select the optimal architecture-centralized, hierarchical, or holonic-based on real-time mission parameters such as task complexity, swarm size, and communication stability. Our system addresses the challenges of scalability, adaptability, and robustness, ensuring efficient energy consumption and maintaining connectivity under varying conditions. Extensive simulations demonstrate that our adaptive architecture outperforms traditional static models in terms of scalability, energy efficiency, and connectivity. These results highlight the potential of our approach to provide a scalable, adaptable, and resilient solution for real-world disaster response scenarios.
Download

Paper Nr:	70
Title:	Leveraging Local Invariants and Graph Neural Networks for Enhanced Anomaly Detection in Distributed Systems
Authors:	Nader Belhadj, Mohamed Amine Mezghich, Lassaad Latrach and Ridha Ghayoula
Abstract:	The security and reliability of distributed systems, particularly microservices, remain challenging due to their scale and dynamic nature. Traditional anomaly detection methods relying on handcrafted rules or thresholds often fail to generalize to evolving workloads. This paper proposes a framework that augments Graph Neural Networks (GNNs) with local invariants-stable node-level properties such as latency bounds or success ratios. By combining relational graph learning with invariant-based constraints, the approach improves both detection accuracy and interpretability. We validate the method on synthetic microservice benchmarks and public intrusion detection datasets (KDD Cup 1999, CICIDS2017, UNSW-NB15). Results show that integrating invariants consistently enhances recall and reduces false negatives compared to classical baselines and GNN-only models. These findings highlight the promise of invariant-augmented GNNs for anomaly detection, while motivating further evaluation on large-scale, real-world deployments.
Download

Paper Nr:	76
Title:	CORA: A Context-Driven Recommendation System Based on Multi-Dimensional User Clustering and Belief-Based Similarity Aggregation
Authors:	Jihene Latrech, Zahra Kodia, Nadia Ben Azzouna and Lamjed Ben Said
Abstract:	We present CORA, a context-driven recommendation system that integrates multi-dimensional contextual modeling and belief-based similarity fusion. The proposed approach consists of four key modules: (1) user clustering in each contextual dimension, (2) computation of dimension-specific similarity scores using tailored statistical metrics (Jensen-Shannon, Hellinger, or Wasserstein), (3) aggregation of similarities using Dempster-Shafer Theory to integrate and manage the heterogeneity of contextual data, encompassing emotional, demographic, and spatio-temporal dimensions, and (4) context-driven collaborative filtering based on the fused similarity matrix. By capturing subtle and heterogeneous contextual signals, CORA provides more personalized and contextually-driven recommendations. Experimental results on the LDOS-CoMoDa dataset demonstrate that CORA achieved an RMSE of 0.8918, outperforming baseline methods, and a competitive MAE of 0.8616. These results confirm the effectiveness of our model in delivering accurate and contextually relevant recommendations.
Download

Paper Nr:	80
Title:	Anomaly Detection with Quantum SVR in the NISQ Era: Limits of Robustness to Noise and Adversarial Attacks
Authors:	Kilian Tscharke, Maximilian Wendlinger, Sebastian Issel and Pascal Debus
Abstract:	Anomaly Detection (AD) is critical in data analysis, particularly within the domain of IT security. In this study, we explore the potential of Quantum Machine Learning for application to AD with special focus on the robustness to noise and adversarial attacks. We build upon previous work on Quantum Support Vector Regression (QSVR) for semisupervised AD by conducting a comprehensive benchmark on IBM quantum hardware using eleven datasets. Our results demonstrate that QSVR achieves strong classification performance and even outperforms the noiseless simulation on two of these datasets. Moreover, we investigate the influence of – in the NISQ-era inevitable – quantum noise on the performance of the QSVR. Our findings reveal that the model exhibits robustness to depolarizing, phase damping, phase flip, and bit flip noise, while amplitude damping and miscalibration noise prove to be more disruptive. Finally, we explore the domain of Quantum Adversarial Machine Learning by demonstrating that QSVR is highly vulnerable to adversarial attacks, with neither quantum noise nor adversarial training improving the model’s robustness against such attacks.
Download

Paper Nr:	89
Title:	Formal Analysis of Hopfield Networks through 0-1 Integer Linear Programming and SMT Solving
Authors:	Sahar Alzahrani, Sven Schewe and Xiaowei Huang
Abstract:	Hopfield Networks (HNs) are recurrent neural networks known for associative memory. They can store and recall patterns, but their behavior—like guaranteed convergence and stability—has been difficult to formally prove. This paper presents a novel method employing Satisfiability Modulo Theories (SMT) and 0-1 Integer Linear Programming (0-1ILP) to analyze HNs. We encode neuron states as binary variables and state transitions as linear constraints, the approach can formally verify all fixed points (stable states), detect cycles (repeating states), and calculate the maximum stabilization time (steps to stability) for synchronous and asynchronous updates. The technique provides a way to assess both global robustness (entire state space analysis) and local robustness (stability against small changes, like a single pixel flip). We evaluated symmetric zerodiagonal HNs with 25-784 neurons in letter images (1–4 patterns) and MNIST data (1–10 patterns). For 25 to 100 neurons, the Z3-based solver detects all fixed points and confirms the absence of cycles in 1 to 10 minutes, stabilizing in 1 to 4 steps. For larger networks (e.g., 784 neurons), analysis is limited by the exponentially large state space (2784 states). Recall drops for 3+ patterns (e.g., 10−33% for 36 neurons), with sensitivity increasing: 1 state change destabilizes 36 neurons, 7 for 100 neurons. This approach provides a powerful tool for constraint-based neural network verification, enhancing associative memory system reliability.
Download

Paper Nr:	91
Title:	Continual Learning by Regularization in Row Space of Weight Matrix of Previous Task for Deep Neural Network
Authors:	Honoka Yamashita, Takio Kurita and Masaki Onishi
Abstract:	When a pre-trained deep neural network is incrementally trained on a new task, its performance on previously learned tasks often degrades, a phenomenon known as catastrophic forgetting. We proposed a method to mitigate this performance degradation by introducing a regularizer in the row space of the weight matrix of the previous task. This regularization enables the learning of the new task to be guided toward the components of the feature vector in the null space that do not affect the previous task, while preserving the components in the row space that affect the previous task. We validated the effectiveness of the proposed method through two sets of experiments. The first experiment involved training a model on FashionMNIST following pretraining on MNIST. In the second experiment, we attached a multilayer perceptron to a ResNet pretrained on ImageNet and incrementally trained it on CIFAR-100, adding 20 classes at each step. For the ablation study, we also compared the proposed method with a simple regularization of the feature vector. Our results demonstrate that the proposed approach effectively preserves the performance of previously learned tasks while adapting to new tasks.
Download

Paper Nr:	92
Title:	VideoGAN-Based Trajectory Proposal for Automated Vehicles
Authors:	Annajoyce Mariani, Kira Maag and Hanno Gottschalk
Abstract:	Being able to generate realistic trajectory options is at the core of increasing the degree of automation of road vehicles. Existing model-driven, rule-based, and classical learning-based methods, however, often struggle to effectively capture the complex and multimodal distributions of future traffic behavior. In this paper, we investigate whether a generative adversarial network (GAN) trained on videos of bird’s-eye view (BEV) traffic scenarios can overcome these limitations by generating statistically accurate trajectories that correctly capture spatial relationships between the agents. To this end, we propose a pipeline that uses low-resolution BEV occupancy grid videos as training data for a video GAN model. From the generated videos of traffic scenarios we extract abstract trajectory data using single-frame object detection and frame-to-frame object matching. We particularly adopt a videoGAN architecture for its favorable trade-off between training efficiency and inference speed compared to diffusion models. Our approach achieves strong results within 100 GPU hours of training, with inference times under 20ms for scenes up to 20s long. Experiments on the Waymo Open Motion Dataset demonstrate that the generated trajectories align with real distributions of spatial and dynamic parameters. These findings highlight video-based GANs as a lightweight and scalable tool for realistic trajectory modeling, with potential to support downstream tasks such as prediction, planning, and simulation in automated driving.
Download

Paper Nr:	99
Title:	The MACB Problem: Definitions, Variants, and a PDDL+ Approach
Authors:	Dani G. Papamaximou, Rongge Guo, Francesco Percassi and Mauro Vallati
Abstract:	Modular Autonomous Customised Bus Systems (MACB) promise to significantly enhance public transport attractiveness and accessibility, leading to improved quality of life and reduced emissions. Compared to traditional bus systems, MACB provides a problem that poses a new set of challenges, ranging from the allocation of vehicles to the optimisation of routes and recharges. The MACB problem is attracting increasing interest within the transport research community, but it presents characteristics and dimensions that lend themselves well to approaches based on planning and combinatorial search. In this paper, with the aim of bridging the gap between different research communities, we provide a crisp definition of the MACB problem and present a planning-based approach to solve a specific variant of the problem, together with a set of benchmarks to foster research on this topic.
Download

Paper Nr:	100
Title:	Think before You Fetch: Smarter Retrieval for Smarter LLMs with Enhanced Table Understanding and Long-Form Reasoning
Authors:	Sreenivasan Mohandas and Tiasha Majumdar
Abstract:	Retrieval-augmented generation (RAG) pipelines often rely on static document chunking, which can fragment related evidence across chunk boundaries and lead to incomplete context for large language models (LLMs). Prior methods-stronger retrievers or larger context windows-address retrieval scale but not the core issue of chunk-boundary dependency. We propose Dynamic Context-based Retrieval (DCR), a framework that dynamically expands or contracts retrieved context beyond fixed chunk boundaries based on marginal information gain and LLM-based semantic validation. DCR incrementally assembles semantically coherent evidence until a context sufficiency criterion is met, combining statistical relevance with model-guided reasoning. Across long-document QA and table reasoning benchmarks, DCR improves factual consistency by up to 17.8% over dense and hybrid retrievers while reducing hallucinations. These findings demonstrate that mitigating chunk-boundary dependency is critical for reliable, context-complete reasoning in retrieval-augmented LLMs.
Download

Paper Nr:	101
Title:	Characterization and Detection of Incompleteness and Ambiguity in Multi-Turn Interactions with LLMs
Authors:	Riya Naik, Ashwin Srinivasan, Swati Agarwal and Estrid He
Abstract:	Natural language interaction with computers has been transformed by Large Language Models (LLMs), which now serve as modern-day oracles capable of answering a wide range of queries. Unlike the single-turn interaction with the Delphic oracle, LLMs support multi-turn dialogues where additional context can improve responses. This paper focuses on identifying incompleteness and ambiguity in user queries during multi-turn interactions with an LLM. Using a simple tagged message exchange model between senders and receivers, we define these properties based on the dialogue sequence. While these definitions help categorize datasets, they cannot be used directly to detect incompleteness or ambiguity. To bridge this gap, we explore the use of Embedding- and Text-based models as detectors. Our experiments on benchmark datasets show that: (a) answer correctness correlates strongly with the presence of incompleteness or ambiguity; (b) we can expect datasets with a high proportion of such questions to have longer multi-turn interactions; (c) effective detectors can be built using only the question and its context. These findings suggest that our proposed approach offers a useful mechanism for characterising datasets, and that trained detectors can be used to automatically identify queries that need to be reformulated before presenting to an LLM.
Download

Paper Nr:	109
Title:	Finding Minimal-Size Sorting Networks Using Deep Q-Learning
Authors:	Theodor Burcă and Mădălina Răschip
Abstract:	Discovering minimal-size sorting networks is a classic combinatorial challenge. This paper investigates the viability of deep reinforcement learning for this task by formulating network construction as a sequential decision-making problem learned by a Deep Q-Network (DQN) agent. We conducted a comparative analysis of a Double DQN agent, which utilizes separate policy and target networks, against a baseline Classic DQN agent that uses only a single policy network for both action selection and value estimation. Both are enhanced with an action masking technique to improve exploration. Experimental results for up to n = 10 inputs showed that our framework successfully generated valid networks that, after pruning, were near-optimal. Counterintuitively, for larger n, the theoretically less stable Classic DQN often discovered solutions that pruned to superior or equal quality compared to its Double DQN counterpart. This finding suggests a complex interplay between algorithmic stability and exploration dynamics in sparse-reward combinatorial problems, warranting further investigation.
Download

Paper Nr:	113
Title:	Interpretable Text Classification Applied to the Detection of LLM-Generated Creative Writing
Authors:	Minerva Suvanto, Andrea McGlinchey, Mattias Wahde and Peter J. Barclay
Abstract:	We consider the problem of distinguishing human-written creative fiction (excerpts from novels) from similar text generated by an LLM. Our results show that, while human observers perform poorly (near chance levels) on this binary classification task, a variety of machine-learning models achieve accuracy in the range 0.93 - 0.98 over a previously unseen test set, even using only short samples and single-token (unigram) features. We therefore employ an inherently interpretable (linear) classifier (with a test accuracy of 0.98), in order to elucidate the underlying reasons for this high accuracy. In our analysis, we identify specific unigram features indicative of LLM-generated text, one of the most important being that the LLM tends to use a larger variety of synonyms, thereby skewing the probability distributions in a manner that is easy to detect for a machine learning classifier, yet very difficult for a human observer. Four additional explanation categories were also identified, namely, temporal drift, Americanisms, foreign language usage, and colloquialisms. As identification of the AI-generated text depends on a constellation of such features, the classification appears robust, and therefore not easy to circumvent by malicious actors intent on misrepresenting AI-generated text as human work.
Download

Paper Nr:	114
Title:	Sensitivity and Ontology-Guided Counterfactual Explanation Generation for Battery State of Charge Estimation
Authors:	Slimane Arbaoui, Ali Ayadi, Ahmed Samet, Tedjani Mesbahi and Romuald Boné
Abstract:	Deep learning models have demonstrated exceptional performance in estimating battery parameters such as State Of Charge (SOC) and State Of Health (SOH) in electric vehicles. However, their black-box nature hinders interpretability, which is critical in safety-sensitive applications. In this work, we introduce a Sensitivity and Ontology-Guided Counterfactual Explanation method (SOCEG) for generating counterfactual explanations in multivariate time series regression tasks. SOCEG combines a gradient-based technique to identify key data segments for modification with an ontology-driven validator to enforce domain constraints and preserve feature correlations, ensuring plausibility and realism. Unlike conventional methods, it does not require an initial population and progressively refines counterfactual quality across iterations. We evaluate SOCEG on SOC estimation for Lithium Iron Phosphate (LFP) cells and benchmark it against four baselines, including GENO-TOPSIS and NSGA-II. Results demonstrate that SOCEG consistently surpasses existing approaches in dissimilarity, implausibility, instability, and runtime, while providing actionable insights into feature modifications, thereby enhancing the model interpretability.
Download

Paper Nr:	120
Title:	Analysis of Human Behavioral Transition Patterns and Cognitive Stance Estimation in Human-Agent Collision Avoidance Experiment Tasks
Authors:	Norifumi Watanabe and Kensuke Miyamoto
Abstract:	This study focuses on understanding the underlying “cognitive stance” that gives rise to diverse strategic choices observed in human cooperative behavior. We aim to enable robots to realize flexible coordination strategies. We adopt a learning-based approach in which multiple behavioral strategies are individually acquired in environments that assume fixed partner strategies. We then introduce a meta-strategy framework that allows robots to flexibly select among these strategies based on observed contexts. To estimate a human partner’s strategic tendency, we extract behavioral transition patterns from human subjects and define a similarity score based on subsequence matching length and frequency. Using this score, we compare inter-individual behavioral patterns and visualize their similarities using heatmaps. The results suggest that robots can infer a partner’s strategic stance and adapt their meta-strategy accordingly. We discuss the significance of this approach for designing robots capable of passive, context-sensitive cooperation, and argue that even seemingly assertive behaviors may reflect a deeper respect for a partner’s intentions, thereby qualifying as desirable cooperative actions.
Download

Paper Nr:	129
Title:	Dual Process Dreamer: Fast and Slow Decision-Making with World Models
Authors:	Tobias Lømo, Adel Baselizadeh, Kai Olav Ellefsen and Jim Torresen
Abstract:	Most robot systems are based on a single decision-making process. This process needs to balance time, energy, and accuracy in every situation. However, according to ”dual process theory” (DPT) from cognitive psychology, this is not how humans work. Depending on the situation, we have the ability to switch between two thinking methods, a fast system 1 (S1) and a slower system 2 (S2). In this paper, we propose a novel approach to a dual process architecture for robots and agents. Our method, called Dual Process Dreamer (DPDreamer), is a combination of a reinforcement learning policy network, a planning algorithm, and a learned world model. The world model allows the parts of DPDreamer to work together and create a more integrated system compared to previous proposals of DPT systems. DPDreamer was tested in a puzzle game called Sokoban, and by balancing the use of S1 and S2, DPDreamer managed a success rate similar to S2 while using S1 most of the time, showing the benefit of using a more adaptable system.
Download

Paper Nr:	133
Title:	Biased Exploration Q-Learning: A Simple Method for Embedding Knowledge into Reinforcement Learning
Authors:	Takahisa Imagawa and Shuichi Enokida
Abstract:	The cumulative regret of Q-learning, one of the representative methods in reinforcement learning, has been analyzed; however, most of them assume that the agent learns from scratch. If the agent has knowledge about the learning domain, its learning will be more efficient, as suggested in transfer learning research. Therefore, we modify an existing Q-learning method and propose Biased Exploration Q-learning (BEQ), which assumes that the agent can acquire domain knowledge in advance. We analyze BEQ theoretically and clarify the conditions under which pruning of suboptimal actions is possible and show that the upper bound of the regret of BEQ is significantly improved compared to the Q-learning. Furthermore, our experiments show that BEQ outperforms Potential-Based Reward Shaping (PBRS), a common method for introducing domain knowledge. Moreover, we apply BEQ in deep reinforcement learning and demonstrate its effectiveness in Atari games with simple domain knowledge.
Download

Paper Nr:	138
Title:	Quantum Architecture Search for Solving Quantum Machine Learning Tasks
Authors:	Michael Kölle, Simon Salfer, Tobias Rohe, Philipp Altmann and Claudia Linnhoff-Popien
Abstract:	Quantum computing uses quantum mechanics to address computational problems in ways that differ fundamentally from classical approaches. While current quantum hardware remains error-prone and limited in scale, Variational Quantum Circuits offer a noise-resilient framework suitable for today’s devices. The performance of these circuits strongly depends on the underlying architecture of their parameterized quantum components. Identifying efficient, hardware-compatible quantum circuit architectures-known as Quantum Architecture Search (QAS)-is therefore essential. Manual QAS is complex and error-prone, motivating efforts to automate it. Among various automated strategies, Reinforcement Learning (RL) remains underexplored, particularly in Quantum Machine Learning contexts. This work introduces RL-QAS, a framework that applies RL to discover effective circuit architectures for classification tasks. We evaluate RL-QAS using the Iris and binary MNIST datasets. The agent autonomously discovers low-complexity circuit designs that achieve high test accuracy. Our results show that RL is a viable approach for automated architecture search in quantum machine learning. However, applying RL-QAS to more complex tasks will require further refinement of the search strategy and performance evaluation mechanisms.
Download

Paper Nr:	139
Title:	Local Visibility Roadmaps in Continous Multi-Agent Path Finding
Authors:	Kristýna Janovská and Pavel Surynek
Abstract:	We address the problem of smooth continuous multi-agent path finding (SC-MAPF), where agents move in a continuous environment along smooth curves in continuous time. The paths for individual agents are searched in local visibility roadmaps that consider obstacles and other agents. A new algorithm is proposed in this work, which combines the previous algorithm CE-CBS for high-level search with any-angle path finding for local single-agent planning, where agents are to avoid obstacles, both static and other agents. The new algorithm is tested and compared to CE-CBS, the algorithm building individual paths using RRT∗, and SMT-CCBS algorithm, an algorithm for MAPFR. Experimental results show various advantages of the new algorithm over CE-CBS and SMT-CCBS in comparable settings.
Download

Paper Nr:	142
Title:	Am I Sure? Leveraging Expert Uncertainty for Reliable Inference in Smart Manufacturing
Authors:	Lukas Lodes and Alexander Schiendorfer
Abstract:	In supervised learning applications in smart manufacturing, such as predictive quality or predictive maintenance, not all instances are equally easy to annotate. Some instances are more difficult to label than others, e.g. due to inherent uncertainty in the sensory data or disagreement between different annotators. Overconfident predictions resulting from treating uncertain labels as certain can lead to costly decisions, such as scrapping parts that should have been retained or vice versa. In areas with similar problems, approaches were developed to work with and not against noisy labels by aggregating annotations of a committee of annotators, However, this is economically infeasible in the manufacturing industry. We therefore adapt these multi-annotator approaches to require only one single annotation with uncertainty indication. We demonstrate that this adaption retains the main advantages from approaches relying on multiple annotations and leads to more controlled model uncertainty. The results on the CIFAR10-H and ZAE Bayern ELPV datasets show that the use of soft labels outperforms other methods, including hard labels and loss scaling when it comes to accuracy and control of high-certainty model predictions.
Download

Paper Nr:	152
Title:	XDeepNN: An Explainable AI Framework for Identifying Adversarial Attacks
Authors:	Negar Zarei, Juba Agoun and Louenas Bounia
Abstract:	Deep neural networks (DNNs) are increasingly embedded in surveillance and command-and-control pipelines, yet small, often imperceptible adversarial perturbations can cause catastrophic failures. We present XDEEPNN, a three-phase defense framework that (i) synthesizes a budget-stratified threat suite (FGSM, PGD, and C&W under ℓ∞ budgets), (ii) detects perturbations using a steganalysis-inspired residual CNN (SRM frontend + learnable TLU + depth-separable residual groups), and (iii) generates pixel-level rationales via SHAP, which are pooled into compact signatures for a LightGBM meta-classifier. Evaluated on the Military Aircraft Detection (MAD) corpus and a synthetic Deep-Fake Aircraft (DFA) extension, XDEEPNN improves AUC from 0.81 to 0.94 over robust-training baselines while sustaining real-time throughput (≥40 fps at 720p) and delivering analyst-readable heat maps. We explicitly define our scope (white-box, ℓ∞-bounded attacks), justify key design choices (SRM expansion, TLU thresholding, SHAP pooling), and ensure reproducibility by releasing code, attack scripts, and the DFA dataset to facilitate public validation. Finally, we discuss extensions to stronger and physical attacks, and outline safeguards to prevent artifact leakage.
Download

Paper Nr:	166
Title:	Hierarchical Attention Networks for Multi-Scale Financial Volatility Forecasting
Authors:	Mihai Bogdan Deaconu and Ioan Daniel Pop
Abstract:	This paper introduces a novel Hierarchical Attention Network (HAN) architecture for multi-scale financial volatility forecasting. The proposed framework integrates heterogeneous market information from high-frequency data using a comprehensive feature engineering pipeline that includes microstructural metrics, relational nearest-neighbor features, and multi-level temporal representations. The HAN-T employs scale-specific encoders for short, mid, and long-term sequences, fused through a hierarchical attention mechanism that dynamically weighs temporal contributions. Trained and evaluated on the Optiver Realized Volatility Prediction dataset using time-aware cross-validation, the model achieves superior predictive performance compared to classical econometric approaches (GARCH), tree-based models (LightGBM), and flat Transformer baselines. Ablation studies confirm that both the hierarchical structure and the attention-based fusion contribute significantly to accuracy and stability. The proposed model demonstrates that explicitly modeling the multi-scale nature of financial time series enhances forecasting precision, providing a promising direction for future work in deep volatility modeling and adaptive market analytics.
Download

Paper Nr:	170
Title:	Is the Necessity Relation a Real Support? Analysis of Weighted Bipolar Argumentation Frameworks with Necessities
Authors:	Marie-Christine Lagasquie-Schiex
Abstract:	The topics of this paper are weighted bipolar abstract argumentation frameworks using necessity relation (wBAFN) and the precise meaning of this necessity relation in abstract argumentation. In order to clarify such a relation, gradual semantics for these frameworks are defined through a translation process of wBAFN into weighted abstract argumentation frameworks with only attacks (wAF); then the formal properties of wBAFN are defined and studied. Our study clearly shows that the necessity relation is not a support relation. Indeed, the classical meaning of support relations corresponds to the fact that a support should “help” its target by reinforcing its value, whereas the impact of a necessity to its target is completely different.
Download

Paper Nr:	173
Title:	Exploring Police Unit Deployment Strategies through Agent-Based Simulation and Crime Urban Digital Shadows
Authors:	Juan Palma-Borda, Eduardo Guzmán and María-Victoria Belmonte
Abstract:	Crime simulation has become an essential tool for understanding criminal dynamics and evaluating prevention strategies. Data-driven Agent-Based Modeling and Simulation (DABMS) provides a powerful approach for this task. It captures both micro-level agent interactions and their aggregate effects on crime patterns while leveraging real-world data. Despite its potential, most existing crime simulations lack realistic spatial environments and rarely consider the operational strategies of law enforcement. In this work, we extend our previously introduced Urban Digital Shadow platform, i.e., an environment powered by DABMS and real-world data to explicitly model police unit deployment and its impact on urban criminality. Our approach incorporates police agents into the dynamic model, enabling the evaluation of deployment strategies, and allows for the empirical analysis of their effects on crime reduction and prevention. The proposed method is characterized by its efficiency and explainability. Using heuristic-based functions derived from historical data and well-known crime theories, the platform supports the rapid allocation of police units across thousands of deployment points. The results demonstrate the potential of our approach as a decision support tool for policymakers, criminologists, and law enforcement agencies. The platform enables the exploration of alternative deployment scenarios, contributing to more effective, data-driven urban crime prevention strategies.
Download

Paper Nr:	177
Title:	i-Check: An Idempotence-Driven Optimisation Framework for AI Agents in Enterprise Workflows
Authors:	Sahil Kale, Yash Nikam and Vijaykant Nadadur
Abstract:	AI agents have emerged as pivotal assets in high-volume customer-facing applications. However, to ensure their performance and evaluation are trustworthy, it is vital to study and optimise their ability to generate consistent, reproducible results for identical or similar inputs. By introducing input constraints called idempotent-driving constraints, we showcase how various agents commonly deployed in commercial applications can achieve optimal idempotent performance, i.e., enhanced repeatability and consistency across responses up to 90% of response stability. Based on improved idempotence, we develop a conceptual, memory-driven framework named i-Check that can detect redundant agent requests and minimise repeated LLM calls to save resources in enterprise workflows. i-Check supports multiple operational patterns across a diverse set of agents and can be seamlessly integrated into existing enterprise pipelines. In a representative implementation simulating a demo user with 100 requests per agent, i-Check achieved up to 40% token savings and 37% cost savings across commonly used agents, while enhancing reliability and repeatability of outputs. These findings underscore the practical value of idempotence as a design principle for trustworthy and cost-efficient AI agents in enterprise settings. Our code and framework are released publicly for open-source usage: https://github.com/knowledge-verse-ai/I-Check.
Download

Paper Nr:	181
Title:	Towards an Approach for Classifying Skin Lesions Using Convolutional Neural Networks
Authors:	Rodrigo Guedes de Souza, Fabricio Ourique, Analúcia Schiaffino Morales, Antonio Carlos Sobieranski and Alison R. Panisson
Abstract:	Early detection of skin cancer is a critical challenge, particularly in regions with limited access to dermatologists. This paper presents a deep learning approach using Convolutional Neural Networks (CNNs) to classify skin lesions as benign or malignant. We developed a preprocessing and augmentation pipeline tailored for medical images to enhance model generalization. Our model is based on a modified ResNet18 architecture with attention mechanisms, implemented using TensorFlow and Keras. Additionally, we explore explainabil-ity techniques to provide interpretable visual explanations that support the model’s predictions. Trained on the International Skin Imaging Collaboration (ISIC) dataset, the proposed approach achieved an accuracy of 99.42% on the test set and 88.12% on the validation set. These results highlight the potential of CNN-based methods in assisting early skin cancer detection, particularly in automated screening scenarios where expert evaluation may be limited.
Download

Paper Nr:	185
Title:	CALM-V: Causal Adaptive Learning for Multi-Agents via Verification Protocol
Authors:	Hariprasauth Ramamoorthy, Unnikrishnan Nair and Suresh Sundaram
Abstract:	Multi-agent systems (MAS) frequently interact within dynamic, evolving environments like e-commerce, healthcare, supply chains, and autonomous driving. Such settings create non-stationarity, as each agent’s local dynamics shift in response to peers’ adaptive policies and changing tasks, leading to a moving target problem and risk of catastrophic forgetting. Moreover, privacy constraints often yield a Partially Observable Markov Decision Process (POMDP) formulation, where limited observations and sparse rewards hamper effective exploration and convergence. We introduce CALM-V (Causal Adaptive Learning for Multi-agents via Verifiable Protocol), a framework that dynamically adapts to shifting distributions, mitigates forgetting through an encrypted replay buffer, and enables privacy-preserving peer knowledge sharing by introducing Meta-Adaptive Verifiable Consensus (MAVC) layer. We validate CALM-V on a multi-agent simulation using the M5-Accuracy retail dataset, with agents facing shuffled holiday calendars and SKU turnover to simulate evolving tasks. Empirical results show that CALM-V achieves an 18% higher cumulative reward, reduces forgetting by 60%, and adapts about three times faster than strong RL baselines while maintaining privacy leakage to below 0.02 bits per transition.
Download

Paper Nr:	192
Title:	Survey on Explainability-Weaponising Adversarial Attack Vectors against Deep Neural Networks and Artificial Intelligence
Authors:	Marek Pawlicki, Ryszard Choraś, Rafał Kozik and Michał Choraś
Abstract:	Adversarial machine learning has revealed the fragility of deep neural networks, while explainable artificial intelligence has been introduced to improve the transparency and trust of AI. It has recently been demonstrated, however, that xAI can be weaponised, enabling adversaries to amplify the effectiveness and efficiency of adversarial attacks. This paper presents the first systematic survey dedicated to xAI-weaponising adversarial attacks. The literature is synthesised across four adversarial goals: evasion, poisoning/backdoors, privacy/inference, and model extraction. A unified taxonomy is proposed that organises attack vectors according to adversarial goals, operational roles of xAI, and attacker capabilities. The bibliographic methodology follows PRISMA guidelines, with structured queries applied to IEEE Xplore, ACM Digital Library, SpringerLink, ScienceDirect, and Google Scholar, complemented by snowballing. The date range was set to 2020-2025. The findings indicate that evasion attacks dominate current literature, while poisoning and extraction attacks remain comparatively underexplored. Open challenges and research directions are identified. This survey reframes xAI from a purely diagnostic tool to a security-critical interface and provides a foundation for principled defence.
Download

Paper Nr:	204
Title:	Automatic Detection and Remediation of Faults in Wi-Fi Networks: An Artificial Intelligence-Based Approach
Authors:	Raul Sorin Frandes and Ioan Daniel Pop
Abstract:	The growing complexity of Wi-Fi networks necessitates automated diagnostic tools for non-technical users. This paper presents WiFixiFi, a system that addresses this need through an end-to-end, explainable Artificial Intelligence (AI) pipeline for automated fault diagnosis and remediation. The system’s architecture integrates three specialized models: a hybrid CNN-LSTM for time-series fault detection, a Graph Neural Network (GNN) for context-aware root cause analysis, and a regression model to predict the effectiveness of remediation actions. Evaluated on a custom-generated dataset, the pipeline demonstrates high performance, achieving 98.4% accuracy in fault detection, 88.1% in root cause analysis, and a R2 score of 0.997 for its self-healing predictions. Integrated Explainable AI (XAI) techniques provide transparency into the model’s reasoning, validating the approach as a powerful and interpretable solution for expanding Wi-Fi troubleshooting.
Download

Paper Nr:	212
Title:	When Prompts Become Payloads: A Framework for Mitigating SQL Injection Attacks in Large Language Model-Driven Applications
Authors:	Farzad Nourmohammadzadeh Motlagh, Mehrdad Hajizadeh, Mehryar Majd, Pejman Najafi, Feng Cheng and Christoph Meinel
Abstract:	Natural language interfaces to structured databases are becoming increasingly common, largely due to advances in large language models (LLMs) that enable users to query data using conversational input rather than formal query languages such as SQL. While this paradigm significantly improves usability and accessibility, it introduces new security risks, particularly the amplification of SQL injection vulnerabilities through the prompt-to-SQL translation process. Malicious users can exploit these mechanisms by crafting adversarial prompts that manipulate model behavior and generate unsafe queries. In this work, we propose a multi-layered security framework designed to detect and mitigate LLM-mediated SQL injection attacks. The framework integrates a front-end security shield for prompt sanitization, an advanced threat detection model for behavioral and semantic anomaly identification, and a signature-based control layer for known attack patterns. We evaluate the proposed framework under diverse and realistic attack scenarios, including prompt injection, obfuscated SQL payloads, and context-manipulation attacks. To ensure robustness, we generate and curate a comprehensive benchmark dataset of adversarial prompts and assess performance across a fine-tuned LLM configuration. Experimental results demonstrate that the proposed approach achieves high detection accuracy while maintaining low false-positive rates, significantly improving the secure deployment of LLM-powered database applications.
Download

Paper Nr:	220
Title:	Making Robots Play by the Rules: The ROS 2 CLIPS-Executive
Authors:	Tarik Viehmann, Daniel Swoboda, Samridhi Kalra, Himanshu Grover and Gerhard Lakemeyer
Abstract:	CLIPS is a rule-based programming language for building knowledge-driven applications, well suited for the complex task of coordinating autonomous robots. Inspired by the CLIPS-Executive originally developed for the lesser known Fawkes robotics framework, we present an Integration of CLIPS into the ROS ecosystem. Additionally, we show the flexibility of CLIPS by describing a PDDL-based planning framework integration.
Download

Paper Nr:	225
Title:	A Hybrid Acoustic Model for Automatic Speech Recognition Based on Discriminative Learning
Authors:	Mukesh Kumar Rohil and Ishant Gupta
Abstract:	The much researched probabilistic speech recognition systems decompose the problem of automatic speech recognition into language modeling and acoustic modeling. Once these two models are defined and their parameters are learnt, a Hidden Markov Model can be built to make predictions on new audio signals. Our work proposes a novel acoustic model inspired by a significant advantage of fitting a single one-dimensional discrete probability distribution over fitting multiple multi-dimensional continuous distributions. Mathematically if y is a multidimensional continuous random variable and x is a discrete random variable, then modeling P(x\|y) using a classification algorithm and P(x) using proportions can be used to model P(y\|x) as P(x\|y) / P(x) up to a constant factor. This is as opposed to fitting probability densities e.g. multivariate Gaussians to each of the labeled portions of the dataset. Since the proposed approach does not assume a (multidimensional) parametric probability distribution for P(y\|x), it has two main advantages, first reduced training time, and second, since P(x\|y) can be modeled quite powerfully using complex models such as neural networks, it is arguably a more accurate fit to the training data. The proposed model has advantages in the terms of speed, accuracy and memory requirements.
Download

Paper Nr:	230
Title:	Extending Temporal Case-Based Reasoning for Action-Conditioned Time Series Prediction from Mixed Asynchronous Data: Application to Data-Driven Medical Simulation
Authors:	Hugo Boisaubert, Lucas Vincent, Corinne Lejus-Bourdeau and Christine Sinoquet
Abstract:	We present TCBR2S2 (Temporal Case-Based Reasoning for Reactive Simulation Scenarios), a framework designed to evolve digital systems in response to external stimuli. The problem is formulated as a case-based reasoning task where (i) past real systems with evolutions resembling the recent history of the simulated system are identified, and (ii) the near future of the simulation, represented as a multivariate time series, is predicted from the most similar observed cases. TCBR2S2 integrates two key components: (i) an action-conditioned temporal pattern recognition method and (ii) a prediction function for simulating system evolution. A downgraded-mode prediction mechanism ensures simulation continuity when no actions are available to guide time series generation. An automated similarity thresholding heuristic supports candidate rejection. Extensive experiments were conducted on a medical use case. Six TCBR2S2 instantiations produced realistic time series dynamics; three were compared with an advanced state-of-the-art machine learning model, achieving comparable realism at lower computational cost and without hyperparameter tuning. By avoiding model design and training, TCBR2S2 can be readily applied to scenarios with event-annotated time series, enabling on-the-fly, data-driven predictions for multivariate temporal data.
Download

Paper Nr:	252
Title:	NRLPSO: A Reinforcement-Learning PSO with Nash-Consistent Scheduling for Random Forest Hyperparameter Optimization in Intrusion Detection
Authors:	Nasreddine Hamdi, Akram Belazi, Safya Belghith and Héctor Migallón
Abstract:	Balancing exploration and exploitation remains a core challenge in swarm optimization, where standard PSO often suffers from premature convergence and limited adaptability to dynamic search landscapes. This work introduces NRLPSO, a particle swarm optimizer that couples reinforcement learning with an evolutionary game-theoretic scheduler to adapt search dynamics online. The method treats EXPLORE, EXPLOIT, REFINE as strategies within a single population evolutionary game; a Nash-consistent mixed policy over operators is updated via replicator dynamics using operator-conditioned payoffs that blend diversity gains and fitness improvements through a smoothed, diversity-aware coefficient. The exploit operator is governed by Q-learning with per-dimension teacher selection and baseline-corrected rewards, while the scheduler relies on an O(Nd) diversity proxy, keeping per-iteration complexity comparable to standard PSO. Evaluations across nine CEC-2017 benchmarks and a cybersecurity case study on Random Forest hyperparameter tuning for intrusion detection show that NRLPSO achieves faster convergence, better final fitness, and improved robustness, with statistically significant gains over strong baselines. These results demonstrate a practical, principled approach to overcoming PSO’s exploration–exploitation limitations through game-theoretic scheduling in swarm intelligence.
Download

Paper Nr:	253
Title:	Integrating Large Language Models with the COMET Method for Automated Multi-Criteria Decision-Making
Authors:	Andrii Shekhovtsov, Michał Gudowicz and Wojciech Sałabun
Abstract:	Large Language Models (LLMs) are often used for decision aid by users, and they are increasingly adopted as expert decision-makers in Multi-Criteria Decision-Making (MCDM) contexts, yet their reliability and consistency remain insufficiently understood. This study examines the potential of integrating LLMs with the Characteristic Objects Method (COMET) to enhance the robustness of AI-assisted decision support. Using LLaMA 3.1 8B as the expert component, we propose a structured LLM-COMET framework for MCDA in which the model performs structured pairwise comparisons to construct a decision model automatically and compare it to LLM-only decision-making. The influence of the temperature parameter on the stability and consistency of LLM-based judgments is analyzed through a comprehensive sensitivity analysis study. Experimental results, based on a real-world decision problem, reveal that higher temperature values substantially degrade decision quality for both LLM-COMET and LLM-only approaches. However, the proposed LLM-COMET approach demonstrates greater robustness and stability than direct LLM-only rankings. Furthermore, notable differences between the two approaches are present, even under very similar settings, highlighting the importance of structured decision frameworks for improving the reliability and interpretability of LLM-assisted decision-making.
Download

Paper Nr:	254
Title:	TRACT: A Transformer and Statistical Framework for Anomaly Detection in Multivariate Non-Stationary Time Series
Authors:	Ibrahim Traore, Imen Megdiche, Jérôme Marquet-Doléac and Lotfi Chaari
Abstract:	Physiological and inertial signals analysis supports numerous healthcare applications, including disease detection, rehabilitation, and treatment. Advancements in signal processing enable the representation of most stationary and non-stationary phenomena using mathematical expressions. These representations provide valuable insights and help in identifying distinctive patterns of interest. In this paper, we propose TRACT, a deep learning and statistical framework designed to detect anomalies in non-stationary environments. It comprises two main components, a transformer-based reconstruction model that captures signal patterns through multi-resolution attention, extending the standard attention mechanism in transformer architecture. During inference, reconstruction errors are computed by comparing observed signals with their reconstructed versions. Statistical modeling is applied to these errors, with parameters estimated directly from the data. TRACT adapts to varying data rates across datasets without imposing strict distribution assumptions, resulting in enhanced robustness and accuracy in anomaly detection for multivariate non-stationary time series. We evaluate TRACT on 12 real-world multivariate time series datasets from diverse domains, demonstrating its performance in anomaly detection tasks with various constraints and its ability to provide early warnings for anomalous events.
Download

Paper Nr:	256
Title:	Sentiment Analysis for Fake News Detection Based on Deep Learning
Authors:	Rana Jarraya, Souhir Bouaziz, Boudour Ammar and Seifeddine Mechti
Abstract:	This study presents a model for fake news detection that integrates sentiment analysis to enhance its effectiveness. The proposed approach SAFN-BERT Sentiment Analysis–Driven Fake News Detection with BERT combines contextual and emotional features of news from articles, extracted through deep learning models such as transformer architectures like BERT. Sentiment analysis is performed using a lexicon-based method with TextBlob to capture the polarity and subjectivity of the text. The model's performance was validated using two datasets: ISOT and GossipCop. Experimental results demonstrate the superiority of the BERT-based model, which achieved remarkable accuracies of 99.34% on the ISOT dataset and 97.54% on the GossipCop dataset. These outcomes underscore the robustness of BERT, complemented by lexicon-based sentiment analysis with TextBlob, in capturing the nuanced textual and emotional dimensions of news content.
Download

Paper Nr:	264
Title:	Dynamic Membrane Time Constant in Spiking Neural Networks: A Stable and Trainable Solution for Event-Based Learning
Authors:	Ahmed Akremi, Hatem Boujemaa, Adel Ben Mnaouer and Omar Tayan
Abstract:	Spiking Neural Networks (SNNs) have emerged as a biologically inspired paradigm for processing spatiotem-poral data with improved energy efficiency and temporal precision. In this work, we introduce the Logarithmic Leaky Integrate-and-Fire (LogLIF) neuron, a novel spiking neuron model that employs a logarithmic reparameterization of the membrane time constant to enable stable and trainable temporal dynamics. We benchmark LogLIF against two widely used models-the classical Leaky Integrate-and-Fire (LIF) and the Parametric LIF (PLIF)-across four representative datasets: MNIST with latency encoding, Neuromorphic-MNIST (N-MNIST), the Spiking Heidelberg Digits (SHD), and the DVS Gesture dataset. For each dataset, we evaluate the same model architecture and parameters, analyzing performance in terms of training loss, test accuracy, and average spike rate. Experimental results show that LogLIF achieves more stable convergence, lower spike activity, and consistent accuracy improvements on temporal datasets, while PLIF retains advantages on frame-based tasks. These findings demonstrate the effectiveness of LogLIF as a robust and biologically plausible solution for event-based learning, and highlight the importance of neuron model design in advancing SNN performance.
Download

Paper Nr:	274
Title:	Multi-Attribute Bias Mitigation via Evolutionary Model Merging and Unlearning
Authors:	Yuka Seki, Ryohei Orihara, Yasuyuki Tahara, Akihiko Ohsuga and Yuichi Sei
Abstract:	The propagation of societal biases and stereotypes by Large Language Models (LLMs) presents a critical challenge to their responsible deployment, often leading to the generation of unfair or harmful content. Several debiasing methods have been proposed, such as dataset reconstruction, output filtering, and unlearning for single attributes. However, these methods often face several challenges. These include the degradation of general language capabilities and the amplification of other biases when mitigating a single attribute. To address these challenges, we propose a method that simultaneously mitigates bias across multiple attributes. Our approach involves individually applying unlearning to suppress stereotypical associations for each attribute and then integrating these specialized models via evolutionary model merging. We demonstrate the effectiveness of our method on a publicly available, pre-trained LLM by targeting four distinct sensitive attributes: gender, profession, religion, and race. Our experimental results confirm that our approach successfully mitigates biases across all targeted attributes. Moreover, fluency and downstream task performance remain largely unaffected, indicating the preservation of the model’s core linguistic competence. This work highlights a promising direction for balancing fairness and utility in LLMs.
Download

Paper Nr:	275
Title:	A Badminton Optimal Shot Prediction Method Based on Deep Reinforcement Learning and Game Trees
Authors:	Tomoki Minooka, Yasuyuki Tahara, Akihiko Ohsuga and Yuichi Sei
Abstract:	In recent years, research on quantitatively evaluating and predicting players’ tactical behaviors in badminton has been actively conducted. Among these approaches, methods that represent rally progressions as game trees and predict the next shot have attracted increasing attention. However, conventional game trees are constructed based on observed rallies, which limits their ability to handle unseen rally developments and fails to sufficiently reflect how much each shot contributes to rally success. To address these limitations, this study proposes a novel method that quantifies the contribution of each shot to rally victory using deep reinforcement learning and incorporates the resulting value as a reward in the game tree. In addition, we expand the game tree by generating pseudo rallies, enabling the model to handle unseen situations. Experimental evaluations using real match data demonstrate that the proposed method outperforms conventional game tree–based and existing prediction models in shot prediction accuracy, while flexibly presenting multiple promising shot candidates. This approach is expected to contribute to tactical analysis and decision-making support in badminton matches.
Download

Paper Nr:	290
Title:	Towards Domain-Robust Activity Recognition Using Textual Representations of Binary Sensor Events
Authors:	Ali Ncibi
Abstract:	Language-based representations have recently emerged as a promising approach for cross-domain Human Activity Recognition (HAR) in smart homes, where binary sensor streams are verbalized into natural-language descriptions and processed by pretrained encoders. However, prior work has typically fixed both the textualization scheme and the text embedding model, leaving open how linguistic design choices influence transferability. This paper presents a comprehensive factorial analysis of textualization and embedding strategies for language-based HAR. We systematically vary (i) how sensor event windows are expressed-across seven existing and novel sequential and summarized textualizations-and (ii) how they are embedded using lexical (TF-IDF), static (Word2Vec), and contextual (SBERT) encoders. Experiments on four public smart-home datasets under consistent in-domain and cross-domain transfer conditions reveal that textualization design, not encoder complexity, governs performance. Sequential, event-ordered sentences maximize in-domain accuracy, while single-sentence, schema-based summaries-such as the proposed Compound Sensor Summary (CSS)-generalize best across homes. Clause-level ablations further show that event descriptions drive recognition, whereas explicit timing information can reduce robustness by overfitting to home-specific schedules. Overall, our findings establish a reproducible framework for analyzing and designing language-based representations in HAR, demonstrating that linguistic structure-rather than deep contextualization-is the primary determinant of domain robustness in smart-home activity recognition.

Paper Nr:	291
Title:	ResSET: A Hybrid Residual Squeeze-Excitation Transformer Network for Sleep Apnea Detection Using Single-Lead ECG Signal
Authors:	Duc Thien Pham and Roman Mouček
Abstract:	Sleep apnea (SA), a disorder characterized by repeated breathing interruptions during sleep, can lead to severe health issues if left untreated. Conventional diagnostic methods like polysomnography (PSG) are thorough but often impractical due to their high cost, discomfort, and complexity. Currently, the development of effective, non-invasive, and accessible diagnostic techniques is of great interest. The electrocardiogram (ECG) plays a crucial role in diagnosing SA due to its capability to detect abnormal heart activity. This study introduces ResSET, a hybrid Residual Squeeze-Excitation Transformer network for automatic SA detection using a single-lead ECG signal. The proposed model achieved 96.25% accuracy for per-segment and 100% accuracy for per-recording classification on the PhysioNet Apnea-ECG dataset, and 99.02% accuracy for the UCD St. Vincent’s University Hospital Sleep Apnea Database (UCDDB), outperforming existing state-of-the-art models in accurately identifying SA from single-lead ECG signal.
Download

Paper Nr:	297
Title:	Explaining Online Debate Evolution under Bipolar Gradual Argumentation Semantics
Authors:	Caren Al Anaissy and Nicolas Maudet
Abstract:	Online debates allow for online collective discussions essential for forming opinions, decisions, and policies within society. Computational argumentation plays an important role in structuring online debates and inferring conclusions in these debates. It allows to represent the arguments exchanged by the participants, the interactions between the arguments and the participants’ votes on each argument. It proposes argumentation semantics that can be used to infer a debate’s outcome by evaluating the exchanged arguments’ strengths. However, in massive online debates, where a large number of arguments is exchanged, it becomes difficult for a participant to navigate through the debate and to understand the reasoning and computation behind the semantics. In this paper, we address this issue by generating abductive explanations (i.e. sufficient reasons) for the debates’ evolution of outcomes computed by bipolar gradual semantics. We define intuitive strategies and heuristics to produce explanations addressing the questions: “why is a debate issue’s final weight higher/lower than its initial weight under a specific semantics?”. We illustrate our methodology, compare and analyze the different heuristics and strategies with respect to the size of their corresponding generated explanations, by conducting several experiments on a real-world dataset.
Download

Paper Nr:	300
Title:	Machine Unlearning Using Forgetting Neural Networks
Authors:	Amartya Hatua, Trung T. Nguyen, Filip Cano and Andrew H. Sung
Abstract:	Modern computer systems store vast amounts of personal data, enabling advances in AI and ML but risking user privacy and trust. For privacy reasons, it is sometimes desired for an ML model to forget part of the data it was trained on. In this paper, we introduce a novel unlearning approach based on Forgetting Neural Networks (FNNs), a neuroscience-inspired architecture that explicitly encodes forgetting through multiplicative decay factors. While FNNs had previously been studied as a theoretical construct, we provide the first concrete implementation and demonstrate their effectiveness for targeted unlearning. We propose several variants with per-neuron forgetting factors, including rank-based assignments guided by activation levels, and evaluate them on MNIST and Fashion-MNIST benchmarks. Our method systematically removes information associated with forget sets while preserving performance on retained data. Membership inference attacks confirm the effectiveness of FNN-based unlearning in erasing information about the training data from the neural network. These results establish FNNs as a promising foundation for efficient and interpretable unlearning.
Download

Paper Nr:	310
Title:	Semantic-Guided Deep Metric Learning for Lung Nodule Retrieval: A Radiology-Aware Framework to Bridge the Semantic Gap in CBMIR
Authors:	Mahbouba Hattab and Ahmed Maalel
Abstract:	Content-based Medical Image Retrieval (CBMIR) plays an increasing role in assisting clinical decision-making by identifying visually and semantically related cases from large imaging repositories. A persistent obstacle, however, is the semantic gap between the visual patterns captured by algorithms and the diagnostic concepts interpreted by radiologists. This work proposes a semantic metric learning framework based on a Siamese Triplet Network designed for the retrieval of lung nodules in CT scans. Unlike conventional triplet formation, the proposed approach employs radiological annotations-specifically malignancy scores and morphological descriptors-to guide semantic triplet generation, ensuring clinical relevance in the learned representations. The framework further integrates batch-hard mining to refine discrimination among embeddings and a scalable FAISS retrieval backend for efficient similarity search on large datasets. Experiments conducted on the LIDC-IDRI dataset demonstrate superior retrieval accuracy over radiomics-based models, CNN classifiers, and recent self-supervised methods such as SimCLR and Vision Transformers. Beyond quantitative gains, qualitative analysis confirms that the learned embeddings align with radiologists’ reasoning, underlining the potential of semantics-guided metric learning to enhance both the interpretability and reliability of CBMIR systems.
Download

Paper Nr:	338
Title:	HSSF-ABIDE: A Semi-Supervised Federated Framework with Heterogeneous Clients for Multi-Site fMRI
Authors:	Ameny Ihkaf, Faten Chaieb and Ali Ben Abbes
Abstract:	In this work, we adapt and extend the Heterogeneous Semi-Supervised Federated Learning (HSSF) paradigm to a multi-site autism spectrum disorder (ASD) classification using ABIDE data. Each client trains locally with scarce labels and abundant unlabeled samples, while a server aggregates knowledge via logit-based distillation on a small public set. The method couples supervised cross-entropy with temperature-scaled distillation and an unsupervised consistency objective gated by a local confidence threshold. We split the ABIDE dataset into three scanner-defined clients to reflect real-world acquisition heterogeneity. Under a federated learning setting, each client trains a distinct local backbone (ResNet-18, EfficientNet, or MobileNetV2), enabling heterogeneous client models while preserving privacy. Our approach reaches a global accuracy of 0.96 after 300 communication rounds-close to centralized training (0.98) while preserving data locality. At the client level, consistent gains were observed over isolated training indicating effective cross-site knowledge transfer despite model and data heterogeneity. We discuss communication costs, sensitivity to the pseudo-label threshold, and fairness across clients, and we outline directions for personalized federated learning under multimodal extensions.
Download

Paper Nr:	339
Title:	A Modular LLM-Based Cognitive Architecture for Sensitive Contexts: A Case Study on Local Home Automation
Authors:	Mathias Barrois and Wejden Abdallah
Abstract:	Technology is becoming increasingly embedded in everyday life, streamlining tasks and enhancing overall quality of life. Smart assistants now play a central role in both professional and personal contexts, whether as home automation systems or conversational agents such as generative AI, offering support, advice, and problem-solving capabilities. Their utility is evident, and their influence on daily activities will continue to expand in the coming years. The growing popularity of home control systems reflects the long-standing aspiration of having a household assistant capable of managing diverse tasks. Yet, most existing solutions still rely on rigid manual configuration, requiring users to define precise commands for effective operation. In parallel, large language models (LLMs) are advancing rapidly, enabling systems to understand and respond to naturally formulated queries with high accuracy, without human fine-tuning. This progress raises a key question: how can we ensure privacy while benefiting from a local intelligent assistant that leverages advanced LLMs and Retrieval-Augmented Generation (RAG)? This paper introduces AdaSyn, a modular cognitive architecture bridging the gap between language-based agents and data-driven actions in sensitive contexts. To demonstrate this approach, we present Dolores, a privacy-focused, fully local smart home assistant that uses open-source LLMs to communicate naturally and securely, without Internet reliance.
Download

Paper Nr:	342
Title:	Federated Learning–Based Semi-Supervised IDS for Medical IoT
Authors:	Ahlem Harhad, Cyril Drocourt, David Durand, Guillaume Muller, Kamal Singh, Abdoulaye Sene and Gil Utard
Abstract:	The rapid adoption of Internet of Things (IoT) technologies in the medical field has introduced new challenges in securing sensitive data against increasingly sophisticated cyber threats. To address the limitations of intrusion detection systems (IDS) that rely on centralized machine learning models, federated learning has emerged as a more privacy-conscious alternative, allowing distributed devices to collaboratively train models without sharing raw data. In this work, we propose a federated, semi-supervised intrusion detection system that couples unsupervised client-side representation learning using a variational autoencoder (VAE); specifically a β-VAE, with server-side classification learning using a small portion of labeled data. Beyond representation learning, we also use a VAE to generate synthetic data and oversample rare attacks. We evaluate across two IoT intrusion detection datasets: the BoT-IoT dataset and the WUSTL-EHMS-2020 Medical IoT dataset, and release simulation code to enable reproducible evaluation and fair comparisons. To explain performance gains, we compare the client-side latent space to that of a standard autoencoder (AE). On the BoT-IoT, the latent space is better structured, showing higher silhouette scores and clearer 3D cluster separations, which yields stronger detection. This improved representation supports stronger detection performance on both datasets while preserving privacy on connected medical devices.
Download

Paper Nr:	359
Title:	A Stream Reasoning Framework for Thermal Image-Based Anomaly Detection in Lithium-Ion Batteries
Authors:	Marwa Zitouni, Sayara Hasanova, Franco Giustozzi, Ahmed Samet and Tedjani Mesbahi
Abstract:	Battery degradation presents a critical challenge to the performance, safety, and lifespan of lithium ion batteries, particularly in electric vehicles and energy storage systems. Research efforts are progressively directed toward the development of advanced diagnostic frameworks capable of monitoring and early detection of degradation phenomena, particularly those observable through thermal behavior. To address this issue, this paper proposes a knowledge based stream reasoning framework that continuously analyzes thermal imaging data to detect degradation related thermal anomalies. Our approach integrates thermal image preprocessing techniques, including segmentation, with a domain ontology that captures spatial, temporal, and thermal relationships inherent to battery operation. By analyzing thermal data as a continuous stream, our approach can identify abnormal heating patterns such as irregular heat diffusion and localized hotspots that indicate potential early stages of degradation. This continuous, ontology guided analysis enables detection of potential failures, enhancing battery monitoring and predictive maintenance capabilities. To assess the effectiveness of the proposed framework, tests were conducted using thermal images acquired during battery aging experiments conducted on a battery cycler.
Download

Paper Nr:	362
Title:	A Novel Graph Neural Network Approach for Social Network Fake News Identification in Arabic
Authors:	Fériel Trabelsi Ben Fraj and Yathreb Mannai
Abstract:	Due to the profound changes in the distribution of information caused by social networks, people are increasingly choosing alternative sources for information searches and access over traditional media. However, the widespread use of social networks encourages bad habits, especially the dissemination of false information. In this paper, we provide an innovative and comprehensive graph neural network approach to deal with such a challenge, specifically for the Arabic language. We believe that the various forms of information (textual and/or multimedia material, social context, dynamics, etc.) and the interactions between them are particularly advantageous in determining the trustworthiness of the news. So, we propose the use of a heterogeneous spatio-temporal graph neural network. Such a network is fed by a huge and complex heterogeneous graph consisting of three types of vertices: authors, posts, and words, as well as two key edges: ’writes’ and ’includes’. AraBERT embedding vectors, contextual, behavioral, and social data all enhance the graph structure. We assess our method on two different datasets and get excellent results. The F-score is 0.947 for a private dataset (TunFake) and 0.933 for the AraFacts dataset. It outperforms other competitive versions with fewer features.
Download

Paper Nr:	371
Title:	LLM Guided Low Code/No Code Synthesis: Visual Scripting for Games
Authors:	Ciprian Paduraru, Razvan Mutu and Alin Stefanescu
Abstract:	This paper investigates LLM-assisted construction of executable visual programs in low-code/no-code (LCNC) game development (illustrated with Unreal Engine 5). LCNC denotes visual and template-driven development that minimizes hand-written code. The method treats prompt-to-graph synthesis as a structured process. The generator emits a two-part artifact consisting of a numbered guide and plain text blocks that serialize the node graphs. The validator then checks wiring and cross-references in a fixed order. A fixer applies local edits that preserve identifiers. A deprecation redirect map, extracted from engine metadata, ensures compatibility across versions. The stages iterate until there are no issues or a small iteration cap is reached, stopping the loop. Evaluation combines human ratings of guide quality and structural validity with an editor compile check on human-authored prompts and paraphrastic variants. Compared with generator-only baselines, the pipeline achieves higher structural validity and a higher compilation success rate, with limited additional latency. Most gains arise from resolving missing connections, multi-driven inputs, and minor naming drift. While the requirements of our approach reflect feedback from the game development industry, the design transfers to other LCNC node-graph editors that expose similar constraints and version metadata. Code, prompts, and validator rules are released publicly.
Download

Paper Nr:	372
Title:	Toward Secure and Compliant AI: Organizational Standards and Protocols for NLP Model Lifecycle Management
Authors:	Sunil Arora and John Hastings
Abstract:	Natural Language Processing (NLP) systems are increasingly used in sensitive domains such as healthcare, finance, and government, where they handle large volumes of personal and regulated data. However, these systems introduce distinct risks related to security, privacy, and regulatory compliance that are not fully addressed by existing AI governance frameworks. This paper introduces the Secure and Compliant NLP Lifecy-cle Management Framework (SC-NLP-LMF), a comprehensive six-phase model designed to ensure the secure operation of NLP systems from development to retirement. The framework, developed through a systematic PRISMA-based review of 45 peer-reviewed and regulatory sources, aligns with leading standards, including NIST AI RMF, ISO/IEC 42001:2023, the EU AI Act, and MITRE ATLAS. It integrates established methods for bias detection, privacy protection (differential privacy, federated learning), secure deployment, explainability, and secure model decommissioning. A healthcare case study illustrates how SC-NLP-LMF detects emerging terminology drift (e.g., COVID-related language) and guides compliant model updates. The framework offers organizations a practical, lifecycle-wide structure for developing, deploying, and maintaining secure and accountable NLP systems in high-risk environments.

Paper Nr:	389
Title:	Coaching How to Search
Authors:	Vassilis Markos and Loizos Michael
Abstract:	Searching is commonly addressed by either directly programming a machine to do so or letting it explore the search state space and craft a path “on its own”, as in Reinforcement Learning. Both approaches place most of the computational and / or cognitive load to one end, be it the programmer or the machine learner. An alternative way, balancing load between the two learning ends, is to have a human or machine coach possesing explicit knowledge on the underlying search problem to provide adequate advice upon a machine learner’s request for further knowledge. In this paper, building on previous work on Machine Coaching, specialized in the particular context of searching, a novel learning paradigm is presented, investigating its efficacy and conformity. Potential next steps and real-world applications of the proposed paradigm are also discussed.
Download

Paper Nr:	393
Title:	Linguistic Analysis of Bias and Emotion in Indian Health News: A Multi-Phased Approach Using BERTopic and Fine-Tuned DistilBERT
Authors:	M. Aravind, Kesar Tripathi and K. M. Kavitha
Abstract:	This study addresses the challenge of linguistic bias and potentially misleading framing in Indian health news by proposing a large-scale analytical framework for examining emotional and thematic patterns in media discourse. We analyze linguistic features, emotional framing, and topic-level variations using a multi-phased Natural Language Processing pipeline. The foundation for this work is an initially collected corpus of over 50,000 health-related news articles in English scraped from major Indian news outlets, which was systematically curated to a high-quality subset of approximately 4,000 articles for downstream analysis. Our methodology integrates BERTopic for semantically coherent thematic modeling, enabling contextual grouping of articles prior to linguistic examination. Emotion analysis is performed using a fine-tuned DistilBERT transformer model to categorize articles into four core emotions, achieving a classification accuracy of 86.29% substantially outperforming the traditional Machine Learning baseline. Rather than directly predicting misinformation, the extracted emotional and subjectivity signals are analyzed as continuous linguistic indicators associated with biased or potentially misleading health narratives. Weakly supervised keyword-based cues are used to initialize coarse-grained categories, enabling scalable exploratory analysis without manual annotation. The results indicate that emotionally skewed language and elevated subjectivity are non-randomly distributed across health topics, suggesting that linguistic bias and potential misinformation follow quantifiable and detectable patterns. This work demonstrates the utility of integrated NLP techniques for systematically analyzing health-related “infodemics” in public communication.
Download

Paper Nr:	404
Title:	MACO: A Strategic Board Game Environment for Advanced AI Research
Authors:	Ali Bakhiet Elimam and Raúl Montoliu
Abstract:	This paper presents Multi-Action Connect (MACO), a strategic board-game environment for AI research on multi-action planning in adversarial settings. MACO is a deterministic, fully observable game played on an 8×8 grid with a connect-6 objective, where each turn comprises three sequential placements, and the rules include a limited-use explode piece that clears adjacent cells when placed. The environment is implemented as a configurable, extensible framework intended to isolate within-turn sequential decision making while remaining tactically rich. To demonstrate MACO’s utility as a research testbed, we integrate representative agents spanning baselines (Random, One-Step Look Ahead, One-Turn Look Ahead), tree search (MCTS and Bridge-Burning MCTS), and evolutionary planning (Genetic Algorithm and Online Evolution), and evaluate them in AI-versus-AI tournaments under fixed per-turn time budgets. Results indicate that methods tailored to within-turn multi-action sequencing (Online Evolution, 74.4%; Bridge-Burning MCTS, 70.3%; and Genetic Algorithm, 69.5%) substantially outperform baselines and standard MCTS (31.2%). Beyond games, MACO provides a controlled setting for studying sequential planning under combinatorial action spaces.
Download

Paper Nr:	415
Title:	Trustworthy Output Neurons for Regression: Applications to Facial Age Estimation and Time-Series Forecasting
Authors:	Yassine Hmidy
Abstract:	Neural networks for regression tasks require a linear output in order to predict the requested real value. In this paper we discuss the use of a non-linear aggregation function as an output neuron for regression tasks. This non-linear aggregation function is based on the computation of Choquet integrals with respect to a parametric set function. We have proven that this aggregation method provides a reliability index through its interval-valued outputs, where the width of the interval quantifies the uncertainty associated with each prediction. Our objective is to ensure that this property is preserved even when our non-linear aggregation function is used as the output layer in complex neural network architectures. We introduce this new aggregation function to replace the conventional linear output layer in networks such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs). To demonstrate the effectiveness of this method, we conducted experiments on two distinct regression tasks involving images and time series data.
Download

Paper Nr:	417
Title:	JewelCVSS: A Domain-Tuned LLM for Automated Vulnerability Scoring
Authors:	Roberto Lorusso, Antonio Maci, Alessandro Santorsola, Pietro Spalluto and Stefano Valcada
Abstract:	Accurately predicting the Common Vulnerability Scoring System (CVSS) score from Common Vulnerability and Exposures (CVE) descriptions is a cornerstone of effective vulnerability management, enabling organizations to prioritize remediation efforts and reduce exposure to emerging threats. However, existing approaches, ranging from rule-based systems to conventional Machine Learning (ML), often struggle with the semantic complexity and variability of vulnerability descriptions, leading to suboptimal risk assessments. For this reason, recent studies have investigated the use of Large Language Models (LLMs) for this task, yet they often exhibit limitations in accurately predicting specific CVSS components due to the fact that LLMs are frequently employed as zero-shot classifiers, relying solely on prompt engineering without explicit adaptation. Simply providing a few-shot prompt at inference time is insufficient to capture the structured semantics of the CVSS framework. While some studies on specialized domains have demonstrated the benefits of fine-tuning, these studies have not evaluated its impact on a broad, general-purpose setting. To bridge this gap, this paper presents JewelCVSS, a fine-tuned LLM based on Gemma3, specifically designed for CVSS component-aware prediction. By leveraging input standardization by transforming CVEs description in standard MITRE format and employing quantized low-rank adaptation (QLoRA) for efficient, low-resource fine-tuning, JewelCVSS achieves promising performance in both overall score and individual metric prediction. Specifically, using a state-of-the-art benchmark consisting of real data, it reaches a mean CVSS vector classification accuracy of ∼82%, with a mean absolute error in predicting the corresponding score equal to 0.9, outperforming both general-purpose LLMs and established ML baselines.
Download

Paper Nr:	425
Title:	A Unified Pipeline for 2D Face Synthesis, Restoration, and Mask‑Guided Editing with Generative Image Models
Authors:	Ali Raad Abdulkareem, Marwa Jabberi, Islem Jarraya, Tarek M. Hamdani and Adel M. Alimi
Abstract:	Text-to-image models can synthesize photorealistic faces from natural language, but outputs often suffer from local artifacts, limited effective resolution, and fragile attribute control. We propose a modular, model-agnostic cascade for high-quality 2D face generation that couples a Face Generator (DALL·E 2, Imagen, SDXL) with a Face Restoration stage (CodeFormer) and optional mask-guided refinement for region-selective edits. The pipeline improves perceptual naturalness, stabilizes geometry, and preserves identity while enabling targeted corrections (e.g., eyewear, teeth, jewelry). We evaluate with a unified protocol spanning geometric stability (normalized landmark error), no-reference IQA (BRISQUE, NIQE, PIQE), fidelity–diversity (PRDC: precision, recall, density, coverage), and perceptual/identity measures (LPIPS change; ArcFace cosine similarity). Restoration consistently improves no-reference perceptual quality and local fidelity for DALL·E 2 and Imagen, with model-dependent effects for SDXL that motivate per-model tuning; identity and facial geometry are largely preserved. Under our configuration, Imagen’s BRISQUE improves by 22.9% and NIQE by 7.7% on average; geometric drift remains ≤ 0.0039 NRMSE across models; post-enhancement ArcFace cosine similarity is ≥ 0.828, indicating preserved identity. These results establish a practical, extensible pipeline for text-driven 2D face synthesis and enhancement and provide guidance on integrating restoration with modern generative models while highlighting per-model tuning considerations.

Paper Nr:	427
Title:	Bridging the Maghrebi Dialect Gap: Multi-Dialect Named Entity Recognition for North African Arabic Using Dialect-Aware Active Learning and Cross-Dialectal Transfer
Authors:	Hassen Mahdhaoui, Abdelkarim Mars, Manar Joundy Hazar, Alaa Abid Muslam Abid Ali, Salah Zrigui and Mounir Zrigui
Abstract:	Named Entity Recognition (NER) for Maghrebi Arabic dialects presents formidable challenges due to substantial linguistic variations across North African varieties (Tunisian, Moroccan, Algerian, Libyan), pervasive French-Arabic code-switching phenomena, Berber substrate influences, absence of orthographic standardization, and severe scarcity of annotated dialectal corpora. While Modern Standard Arabic (MSA) NER has achieved performance exceeding 90% F1-score, these systems demonstrate catastrophic degradation when applied to dialectal text, with F1-scores dropping to 55-65%. In this work, we present a comprehensive multi-dialect NER framework specifically architected for Maghrebi Arabic varieties. Our contributions comprise: (1) a Maghrebi dialect detection module achieving 89.7% classification accuracy, (2) a novel neural architecture with Common Maghrebi Features Layer handling French/Berber code-switching, (3) dialect-aware active learning reducing annotation costs by 43%, (4) effective cross-dialectal transfer learning (13.4% improvement for low-resource Libyan), and (5) integration of large language models for auto-correction. We construct and release the first large-scale multi-dialect Maghrebi NER benchmark comprising 330,000 annotated sentences. Our model achieves macro-averaged F1-score of 87.34% across all Maghrebi dialects, substantially outperforming MSA-only baselines (61.2% F1), standard multi-dialect approaches (80.9% F1), and existing dialectal Arabic NER systems. This work establishes the first comprehensive foundation for Maghrebi Arabic NER with direct applications to social media analytics, government digital services, and business intelligence.
Download

Paper Nr:	431
Title:	BiScaleFormer: Modeling Trend and Seasonality with Multi-Scale Attention for Time Series Forecasting
Authors:	Kexin Peng and Hitoshi Iima
Abstract:	This paper proposes a dual-channel multi-scale Transformer model for time series forecasting, BiScaleFormer. The model explicitly decomposes input sequences into trend and seasonal components, which are then processed through independent multi-scale attention pathways to adaptively capture long-term and short-term temporal patterns, respectively. The outputs of the two branches are subsequently integrated via a configurable fusion module, enabling the model to flexibly adapt to diverse temporal behaviors. Experimental results on the Electricity Transformer Temperature benchmark dataset demonstrate that BiScaleFormer outperforms existing state-of-the-art methods in terms of both forecasting accuracy and model interpretability, showcasing the potential of dual-scale attention mechanisms for building robust and modular forecasting systems.
Download

Paper Nr:	435
Title:	Progressive Diffusion Model for UAVs Image Generation
Authors:	Hèdi Fkih, Abdelaziz Kallel and Zied Chtourou
Abstract:	Monitoring systems today rely heavily on deep learning models for sky surveillance. These networks are generally trained on images in the visible spectrum. Difficulties arise in low-light conditions, in bad weather and at night, necessitating the capture of images in the thermal domain. However, the recognition of small objects that can be found in the sky, such as Mini/Micro drones, is best achieved with visible images. To address these challenges, we focus on the translation of thermal images into visible images. In this paper, we propose a new diffusion model specifically designed for efficient translation tasks with limited datasets, entitled Progressive Diffusion Model (PDM). Our approach is based on a progressive noise removal strategy. The forward process progressively adds noise to each image of the combined dataset at different resolutions. In this way, the model is exposed to a greater diversity of noise patterns, even with a limited dataset. The reverse process consists of an iterative refinement step. It starts with pure noise at the lowest resolution and progressively removes noise in stages, depending on the corresponding sampled thermal image. Then, the generated image is upsampled to become the input for the next refinement step, making the process faster and more efficient. Our model achieved state-of-the-art translation results on a customized Mini/Micro drones dataset.
Download

Paper Nr:	458
Title:	Explainable Anomaly Detection in Industrial IoT via CNN–LSTM and SHAP-Based Rule Induction
Authors:	Wajih Abdallah and Sami Mnasri
Abstract:	Industrial Internet of Things (IIoT) settings generate vast amounts of multidimensional sensor data, where timely and reliable anomaly detection is critical for predictive maintenance and operational safety. Although deep learning techniques achieve high detection performance, the black box nature of these techniques limits the understanding and trustworthiness of human operators working in mission critical applications. We address this issue by proposing an explainable deep learning framework that incorporates a convolutional– recurrent backbone for spatio-temporal feature learning, with a layer for explanations based on SHAP and rule induction for human-interpretable decision support. The experimental results from two widely used industry datasets, SWaT and WADI, show a detection performance improvement of 6–7% F1 score, with a reduction in anomaly detection delay of more than 40% compared with LSTM and autoencoder-based approaches. The explainability module maintains fidelity above 90%, achieving explanation latencies under 100 ms, an acceptable limit for a real-time application. Furthermore, localized feature attributions are consistent with instantiating system ground-truth variables such as tank levels and pump flows, underlining the operational significance of the explanations. In summary, the findings demonstrate that the quantifiable explainability feature incorporated into the deep learning framework not only improves detection performance, but also improves the trustworthiness of algorithmic data analysis and deployability for anomaly detection applications in IIoT environments.
Download

Paper Nr:	477
Title:	A Learning-Augmented Ant Colony Optimization with Graph Neural Network Algorithm for Multi-Objective Optimization
Authors:	Majdi Nciri, Imen Ben Mansour, Ines Alaya and Moncef Tagina
Abstract:	The application of metaheuristics and machine learning (ML) techniques for addressing multi-objective problems has emerged as a prominent area of research in recent years. This article proposes the integration of machine learning methodologies into the Ant Colony Optimization framework to tackle the multi-objective multi-dimensional knapsack problem. Specifically, we employ characteristics that describe each item in the problem. These features are then utilized by a Graph Neural Network to estimate the probabilities of items being part of the optimal solution. These predictions are subsequently incorporated into the ant algorithm to direct the search process towards the most promising regions of the search space, thereby enhancing the likelihood of obtaining high-quality solutions. The performance and effectiveness of the novel approach are demonstrated through a series of comparisons with state-of-the-art algorithms. Experimental results indicate that the incorporation of the solution prediction method significantly enhances the Ant Colony Optimization performance in terms of both convergence and solution diversity.
Download

Paper Nr:	486
Title:	Learning from Interpretable Goals: Dynamic Weight Balancing for Multi-Objective Reinforcement Learning
Authors:	Simon Schwan, Willie Szollmann and Sabine Glesner
Abstract:	Reward engineering remains a central challenge in reinforcement learning (RL), particularly when multiple, potentially conflicting objectives must be balanced. Often, this is addressed through a weighted sum of reward components, yet this approach introduces several difficulties: extensive manual tuning of weights, lack of semantic interpretability, and limited generality of methods that rely on algorithm-specific extensions. In this work, we address these challenges by proposing a goal-based reward formulation that replaces opaque weight tuning with interpretable goal definitions. These goals are learned through a curriculum-based training scheme that eliminates the need for manual weight tuning by dynamically adapting weights while remaining agnostic to the underlying RL algorithm. We evaluate our approach on five control environments from the Farama Gymnasium and show that it achieves comparable or superior success rates to the original tasks. Moreover, our method eliminates the need for extensive reward weighting and introduces an interpretable training framework for multi-objective reinforcement learning.
Download

Paper Nr:	487
Title:	Analysing Automated Monitoring of Industrial Water Treatment Plants: From IoT Sensing to Computer Vision Approaches
Authors:	Rajarshi Biswas, Om Khairate and Agnetha Flore
Abstract:	In this work, we survey different techniques for solving challenges encountered in industrial waste-water treatment plants through vision-based monitoring. Our focus lies on three core applications: inflow anomaly detection, aeration monitoring using bubble segmentation in industrial water treatment containers, and camera-based water-level estimation in sedimentation tanks. We perform a survey grounded in real-world implementation challenges, that is, lighting and reflection variation, water clarity, limited annotation availability, and domain shift. While our emphasis is on computer vision, we also acknowledge complementary IoT-based chemical sensing and above- or underwater visual inspection approaches for industrial compliance, as well as recent developments in segmentation, marker detection, and geometric reasoning for water-level measurement. To demonstrate the practical implications of these techniques, we implement three representative pipelines: background subtraction with geometric constraints for inflow anomaly detection, aeration monitoring through instance based bubble segmentation using YOLO, and checkerboard-based calibration for water-level estimation in sedimentation tanks. We report the results obtained and conclude with a discussion of the observed trade-offs in performance, data availability, and future directions for integrating IoT and vision-based systems into a unified, edge-deployable monitoring framework for industrial water treatment processes.
Download

Paper Nr:	496
Title:	How Can We Effectively Identify Optimal Hyperparameters to Improve Model Reliability in Medical Applications?
Authors:	Syrine Ben Ahmed, Amani Elaoud and Imene Ben Hafaiedh
Abstract:	Machine learning is increasingly critical for solving real-world problems, particularly in sensitive domains such as healthcare. The effectiveness of these models depends not only on the chosen algorithm but also on fine-tuning them according to each dataset and task. A major challenge is selecting hyperparameters, which strongly influence model performance and generalization. Optimizing models for medical datasets requires a careful, context-aware hyperparameter tuning approach rather than simply choosing an algorithm. This work presents a guided strategy to identify the most effective hyperparameters for each dataset and task, applied to three well-known machine learning models: Support Vector Machines (SVM), K-Nearest Neighbors (KNN), and Random Forests (RF). We investigate hyperparameter optimization techniques including Grid Search, Random Search, Genetic Algorithm, Bayesian Optimization, Optuna, and the Slime Mould Algorithm (SMA). Model performance is evaluated using accuracy, precision, recall, F1-Score, specificity, and AUC-ROC to capture medical decision-making needs. Results demonstrate how optimization strategies impact effectiveness, particularly under challenges like class imbalance, and provide practical guidance for selecting suitable tuning methods to improve model reliability in healthcare applications.
Download

Paper Nr:	500
Title:	Sliced-Wasserstein Distribution Alignment Loss Improves the Ultra-Low-Bit Quantization of Large Language Models
Authors:	Deyu Cao, Yixin Yin and Samin Aref
Abstract:	The benefits of most large language models come with steep and often hidden economic and environmental costs due to their resource usage inefficiency during deployment. Model quantization improves energy and memory efficiency through representing model parameters by lower-precision values. However, compression below 4-bits often distorts activation distributions and degrades performance. We address this challenge by introducing a sliced Wasserstein loss function for distribution-aware calibration in ultra-low-bit post-training quantization. The proposed loss aligns the output distributions of full-precision and quantized models under random linear projections, complementing standard mean-squared error loss without adding any computational overhead during inference. Our proposed loss function can be incorporated with any post-training quantization framework that has a retraining component. We demonstrate the performance gains of our proposed model by incorporating it with two frontier methods known as OmniQuant and TesseraQ. Compared to these two baselines, the proposed loss consistently improves both perplexity and downstream task accuracy across multiple ultra-low-bit settings. Our proposed loss function recovers 4.12-20.37% of the OmniQuant’s lost accuracy on the language model LLaMA-2-7B, 0.93–7.65% on OPT-6.7B, and 2.26–6.20% on LLaMA-2-13B. TesseraQ’s accuracy degradation is recovered by 3.63-7.63% in relative terms when augmented by our proposed loss function. Taken together, these results demonstrate that distributional alignment provides a simple yet effective performance boost that can push the limits of frontier quantization methods. Our method is available on GitHub to facilitate future progress in ultra-low-bit quantization.
Download

Paper Nr:	511
Title:	Multimodal Sentiment Analysis: A Survey
Authors:	Mehrez Hosni, Hamza Gharsellaoui and Sadok Bouamama
Abstract:	Multimodal Sentiment Analysis (MSA) has emerged as a rapidly growing field that seeks to interpret human emotions by integrating information from text, audio, and visual modalities. This survey provides a comprehensive overview of the evolution of sentiment analysis from unimodal to multimodal approaches, emphasizing the role of deep learning and fusion strategies in enhancing performance. We review state of the art methods, including attention based fusion and multimodal transformers, and highlight benchmark datasets such as CMU-MOSI, CMU-MOSEI, MELD, and CMU-MOSEAS, which have played a pivotal role in advancing research. The paper also analyzes the strengths and limitations of existing fusion techniques, ranging from early and late fusion to hybrid and attention driven frameworks. Persistent challenges such as cross modal alignment, data imbalance, computational complexity, and the handling of figurative language are discussed in detail. Finally, we outline promising future directions, including lightweight and interpretable architectures, cross lingual and domain generalizable models, and the integration of underexplored modalities. By synthesizing methodologies, datasets, and challenges, this survey aims to guide researchers toward more robust, efficient, and impactful sentiment aware systems.
Download

Paper Nr:	516
Title:	From Attention to Frequency: Integration of Vision Transformer and FFT-ReLU for Enhanced Image Deblurring
Authors:	Syed Mumtahin Mahmud, Mahdi Mohd Hossain Noki, Prothito Shovon Majumder, Abdul Mohaimen Al Radi, Md. Haider Ali and Md. Mosaddek Khan
Abstract:	Image deblurring is vital in computer vision, aiming to recover sharp images from blurry ones caused by motion or camera shake. While deep learning approaches such as CNNs and Vision Transformers (ViTs) have advanced this field, they often struggle with complex or high-resolution blur and computational demands. We propose a new dual-domain architecture that unifies Vision Transformers with a frequency-domain FFT-ReLU module, explicitly bridging spatial attention modeling and frequency sparsity. In this structure, the ViT backbone captures local and global dependencies, while the FFT-ReLU component enforces frequency-domain sparsity to suppress blur-related artifacts and preserve fine details. Extensive experiments on benchmark datasets demonstrate that this architecture achieves superior PSNR, SSIM, and perceptual quality compared to state-of-the-art models. Quantitative metrics, qualitative comparisons, and human preference evaluations confirm its effectiveness, establishing a practical and generalizable paradigm for real-world image restoration. Notably, the frequency-domain module introduces negligible memory overhead and scales efficiently to high-resolution images, making the approach practical for real-world deployment. The source code is available at https://github.com/Dip-to/Vision-Transformer-and-FFT-ReLU-Fusion-for-Enhanced-Image-Deblurring.
Download

Paper Nr:	519
Title:	Smart Road Allocation through Auction-Based Mechanisms for Congestion Prevention
Authors:	Chava Schweitzer, Rina Azoulay, Esther David and Rica Gonen
Abstract:	In the modern world, traffic management faces significant challenges, as growing urbanization and rising vehicle ownership place unprecedented pressure on road networks, leading to persistent congestion and inefficiencies. To address this, we aim to enable vehicles with higher urgency or greater willingness to pay to access the most valuable routes and time slots, while guiding other traffic toward alternative paths or off-peak periods. We present a network-wide traffic management mechanism that allocates complete, uncongested routes via an online combinatorial auction across space–time road segments, achieving incentive compatibility up to ε—so that any potential gain from misreporting is bounded by ε, and truthful reporting remains essentially optimal for all users. Simulation results on an OSM-derived urban network show that our mechanism consistently achieves higher social welfare and more efficient capacity utilization than the BG baseline, fixed-price allocation, and free entry (where segments close once capacity is reached). The approach maintains fairness and service quality, requires no dedicated infrastructure or pre-declared tolls, and ensures predictable travel times by pre-allocating complete routes subject to capacity.
Download

Paper Nr:	532
Title:	REFINE: Robust Evaluation Framework for IDS under Concept Drift in Dynamic Environments
Authors:	Gabriele Nicolò Costa, Alessandra De Paola, Salvatore Drago, Pierluca Ferraro and Giuseppe Lo Re
Abstract:	Most machine learning-based Intrusion Detection Systems (IDSs) are designed for stationary environments, where data distributions is assumed to remain constant over time. However, modern network environments are dynamic, and this can lead to significant changes in the observed environment since the training phase, causing degradation in IDS performance. Consequently, increasing attention has been given to online learning techniques designed to address such phenomenon, known as concept drift. Designing such adaptive systems is a far from trivial task, due to a multitude of factors, such as experimental biases as well as the lack of real-world labeled datasets with precise drift annotations. Moreover, the evaluation of such systems still lacks a standardized methodology, and critical aspects are often inconsistently addressed, making comparisons between approaches particularly difficult. To address these challenges, this work proposes REFINE, a Robust Evaluation Framework for IDS under concept drift in dynamic environments. REFINE combines a Concept Drift Stream Generator (CDSG), which produces realistic datasets from real network traffic with controlled drift characteristics, and a robust online evaluation pipeline that mitigates experimental biases. Results demonstrate that REFINE enables accurate, unbiased evaluation and comparison of online IDSs, providing critical insights into their adaptation and detection capabilities across various drift scenarios.
Download

Paper Nr:	536
Title:	Quantum King-Ring Domination in Chess: A QAOA Approach
Authors:	Gerhard Stenzel, Michael Kölle, Tobias Rohe, Julian Hager, Leo Sünkel, Maximilian Zorn and Claudia Linnhoff-Popien
Abstract:	The Quantum Approximate Optimization Algorithm (QAOA) is extensively benchmarked on synthetic random instances such as MaxCut, TSP, and SAT problems, but these lack semantic structure and human interpretability, offering limited insight into performance on real-world problems with meaningful constraints. We introduce Quantum King-Ring Domination (QKRD), a NISQ-scale benchmark derived from chess tactical positions that provides 5,000 structured instances with one-hot constraints, spatial locality, and 10–40 qubit scale. Using QKRD, we systematically evaluate QAOA design choices and find that constraint-preserving mixers (XY, domain-wall) converge approximately 13 steps faster than standard mixers (p < 10−7, d ≈ 0.5)while eliminating penalty tuning, warm-start strategies reduce convergence by 45 steps (p < 10−127, d = 3.35) with energy improvements exceeding d = 8, and CVaR optimization yields an informative negative result with worse energy (p < 10−40, d = 1.21) and no coverage benefit. Intrinsic validation shows QAOA outperforms greedy heuristics by 12.6% and random selection by 80.1%. Our results demonstrate that structured benchmarks reveal advantages of problem-informed QAOA techniques obscured in random instances. We release all code, data, and experimental artifacts for reproducible NISQ algorithm research.
Download

Paper Nr:	552
Title:	LLM-Based Risk Scenario Generation for Hardware Products: A Quantitative Validation against Real-World User Reviews
Authors:	Arisa Morozumi and Hisashi Hayashi
Abstract:	Conventional risk analyses for hardware products, such as Failure Mode and Effect Analysis (FMEA), are fundamentally function-centric and often fail to anticipate critical risks related to user experience degradation. This study addresses this gap by introducing a novel ”value-centric” risk analysis framework that leverages Large Language Models (LLMs). We propose a systematic methodology that steers an LLM’s reasoning process using a hierarchical model of customer value. To validate the framework’s efficacy, we conducted a rigorous quantitative evaluation, benchmarking the LLM’s generated scenarios against a ground truth of 153 real-world consumer reviews. By employing Manhattan distance to measure the discrepancy, our findings show that the value-centric approach significantly mitigates the LLM’s inherent biases. It successfully replicates the structure of real-world issues with superior fidelity across different product concepts-from those with physical defects to those with experiential value flaws. This work establishes a practical, value-centric methodology for preemptive risk management and reframes the LLM as a strategic partner capable of foreseeing customer dissatisfaction at the earliest stages of development.
Download

Paper Nr:	564
Title:	Neural Network-Based Learners for Effort-Aware Software Defect Prediction
Authors:	Camelia-Petrina Nadejde, Camelia Serban and Andreea Vescan
Abstract:	Predicting software defects is vital for maintaining reliable systems and is a significant issue for the industry. Automating the defect detection process in software can greatly decrease errors, reduce development time, and lower costs. This paper has two main objectives: to replicate and generalize the results of a previous study and to offer new insights by extending the study using different prediction models. A controlled replication is conducted to evaluate five classical machine-learning learners (CART, RF, SVM, UBagSVM, UBstSVM) and four ranking strategies (Prob, CBS+, Prob/LOC, EA-Z) across 59 project pairs drawn from the PROMISE and AEEEM repositories. Results demonstrate high replication fidelity: EA-Z and Prob/LOC consistently yield the best trade-off between Recall@20 and Popt, confirming the stability of the original findings while slightly improving robustness under resampling. To explore non-linear generalization, two neural network-based learners (NN-MLP and NN-RBFN) are further assessed under the same effort-aware setting. NN-RBFN significantly surpasses NN-MLP in both Recall@20 and Popt-particularly with the CBS+ and Prob/LOC strategies-highlighting the value of radial basis representations for capturing complex metric–defect interactions. Overall, the study provides empirical evidence that neural learners can enhance EA-SDP performance without altering the relative ranking hierarchy observed in traditional models.
Download

Paper Nr:	569
Title:	NeuroSync-Agent: A Real-Time Multimodal Fusion and LLM Reasoning Framework for Cognitive-State Inference
Authors:	Basma Jalloul, Bassem Bouaziz, Siwar Chaabene and Walid Mahdi
Abstract:	Real-Time interpretation of physiological signals requires both reliable multimodal fusion and intelligent reasoning. Existing systems often lack validated synchronization or do not incorporate AI-driven interpretation, limiting their use in continuous cognitive-state assessment. We introduce NeuroSync-Agent, a lightweight framework that integrates sub-millisecond synchronization of electroencephalography (EEG) and heart rate variability (HRV) with a multimodal large language model (LLM) reasoning module. NeuroSync’s synchronization pipeline combines interpolation, temporal binning, and drift correction to align heterogeneous streams with an average RMSE of 0.10 ms and 76 ms latency, providing a stable foundation for downstream AI analysis. On top of this fused representation, the proposed LLM agent consumes structured EEG-HRV summaries and generates transparent, human-readable assessments, offering interpretable cognitive-state inference and decision-support insights. We validate the agent using a controlled synthetic dataset with EEG-HRV patterns representative of typical Healthy and MCI-like conditions, providing a neutral testbed for assessing diagnostic reasoning. Results show consistent classification behavior and physiologically coherent explanations with an accuracy of 92%, highlighting the potential of combining multimodal fusion with LLM-based reasoning. By bridging low-level physiological alignment with high-level AI interpretation, NeuroSync-Agent provides an extensible architecture for future intelligent monitoring systems and real-time cognitive assessment applications.
Download

Paper Nr:	570
Title:	Sura.ai: Multi-Agent Infrastructure Recovery with LLM-Powered Autonomous Remediation
Authors:	Ananya Arvind, Shruthi Sathya Narayanan and Saishriya Sathya Narayanan
Abstract:	Infrastructure failures like the 2024 CrowdStrike incident in July demonstrate the critical need for autonomous recovery systems that can detect, diagnose, and remediate outages without the dependency on human intervention. We present Sura.ai, a multi-agent system using four cooperating Fetch.ai uAgents orchestrated via the Agentverse Mailbox to provide autonomous infrastructure recovery. Our system integrates LLM-powered root cause analysis through Claude Sonnet 3.5, enabling intelligent decision-making beyond simple rule-based automation. We conducted comprehensive testing across four disaster scenarios including simulated faulty updates, CPU spikes, and cascading failures. Sura.ai achieved a 97.6% action success rate (41/42 incidents resolved) with intelligent alert deduplication across all tested scenarios. Our work demonstrates the feasibility of LLM-based multi-agent orchestration for critical infrastructure resilience and introduces practical patterns for AgentOps implementation in the growing agentic cybersecurity space.
Download

Paper Nr:	577
Title:	When Specialized Beats General: Embedding-Based vs. Large Language Model Classification for the MEPA Pedagogical Ontology
Authors:	Léonard Noth, Sandy Ingram El Helou, Joris Felder, Morgane Nissille and Bernadette Charlier
Abstract:	Accurately classifying user inputs into pedagogical ontologies is essential for conversational Artificial Intelligence (AI) systems in educational technology. The MEPA ontology comprises seven specialised pedagogical concepts designed to represent personal learning experiences, and presents unique classification challenges due to its domain-specific nature and conceptual nuances. This paper evaluates the ability of state-of-the-art large language models to perform this specialised classification task effectively. Multiple LLM variants were tested, including GPT-5 (Standard, Mini, and Nano) and the Claude model families. All were configured with extended reasoning capabilities and optimised prompts. Despite these optimisations, LLMs achieved suboptimal classification performance on our test dataset of 350 expert-labelled samples. In response, we developed a specialised classification system that combines OpenAI’s Text Embedding 3-Large model with a feedforward neural network that was trained on 14,000 balanced samples. Our custom model achieved 94% accuracy, outperforming the best LLM baseline by 12%. We analyse the factors contributing to the limitations of LLMs in highly specialised ontological contexts and discuss the practical implications for educational technology systems that require precise alignment with pedagogical frameworks. Our findings suggest that domain-specific models can significantly outperform general-purpose LLMs for specialised classification tasks.
Download

Paper Nr:	597
Title:	Enhancing Speaker Naming through Multimodal Fusion of Text and Audio Cues
Authors:	Mohamed Lazhar Bellagha and Mounir Zrigui
Abstract:	This study addresses the problem of automatic speaker naming, which involves the identification of speakers based on their true identities. Conventional methods that rely on textual information and linguistic rules to link names with adjacent speaker segments face significant limitations, particularly when applied to spoken text. In this paper, we propose a speaker identification approach that leverages both the audio signal and its transcription to name the speakers. Our approach consists of two modules. The first module, named ”Names assignment”, analyzes the lexical and semantic context when a name is mentioned, determining whether it corresponds to the current speaker segment, the next speaker segment, or the previous speaker segment. To achieve this, we introduce an attention-over-attention (AOA) neural network for names assignment. This model employs an attention mechanism to assess the relevance of each word with respect to a given name. During the evaluation phase, our model yields satisfactory results, achieving an F1 rate of 87.92% and an accuracy rate of 88.69%. The second module, ”Names Propagation”, propagates names to other segments based on acoustic similarity while integrating speaker role information into the naming process. Incorporating this information reduces the identification error rate from 28% to 15.3%.
Download

Paper Nr:	598
Title:	Quantitatively Audited Multi-View XAI for Medical Image Classification: Application to MGMT Promoter Methylation
Authors:	Manel Mili, Abderrahman Ben Abdeljelil, Antoine Manzanera, Asma Ben Abdallah, Jose Javier Otero and Mohamed Hedi Bedoui
Abstract:	The deployment of artificial intelligence (AI) in safety-critical domains such as healthcare requires not only high predictive performance but also quantitatively audited explanations to support trust and human oversight. We present a backbone-agnostic, multi-view explainable AI (XAI) framework for complex image classification, instantiated on the prediction of O6-methylguanine–DNA methyltransferase (MGMT) promoter methylation from hematoxylin–eosin (H&E) slides. A lightweight MobileNetV2-based classifier with shared attention mechanisms and NdLinear projections is coupled with three complementary explanation modalities: (i) SmoothGrad-CAM saliency maps that highlight discriminative tissue regions; (ii) Vision Transformer-conditioned textual rationales grounded in image embeddings; and (iii) pixel-space counterfactuals that synthesize minimal perturbations required to flip model decisions. We quantitatively audit these explanations using sensitivity measures, jointly assessing predictive performance and explanation behavior. On a held-out glioblastoma test cohort, the framework attains 94% patient-level F1-score, and it preserves non-trivial discrimination after fine-tuning on an external grade II–III cohort (83% F1-score). By tightly coupling prediction, multi-view explanations and their quantitative audit, this work contributes to the XAI field and offers a practical blueprint for trustworthy biomarker prediction in computational pathology.
Download

Paper Nr:	600
Title:	TrustBoot: A Trust Bootstrapping Framework for Semi-Supervised Malware Detection
Authors:	Andrea Augello, Alessandra De Paola and Giuseppe Lo Re
Abstract:	Today, malware detection represents one of the most critical cybersecurity challenges due to the rapid evolution of threats. One of the most promising approach is the adoption of machine learning (ML) detection methods, nevertheless, their design is not trivial due to the scarcity of up-to-date labeled data. In order to keep up with emerging malware variants, ML-based detection systems must be frequently updated and retrained using recent samples. However, the manual process of feature engineering and expert labeling and analysis is time-consuming and costly, making it impractical for frequent updates. This work presents TrustBoot, a semi-supervised framework for detecting malicious software, that exploits the exact knowledge only about a small set of trusted applications, and is capable of processing a larger set of unlabeling applications. To achieve this goal, TrustBoot adopts a visual encoding of binary executable, that eases the detection of anomalies, which are related to the presence of malware. Experiments on large Android malware datasets demonstrate that the proposed pipeline achieves competitive detection performance, matching or exceeding fully supervised approaches while substantially reducing the need for manual intervention for the dataset curation and overcoming the reliance on labeled malicious data.
Download

Paper Nr:	621
Title:	Region-Focused CNN–ViT Hybrid for Medical Image Classification
Authors:	Yosra Didi, Ahlam Walha and Ali Wali
Abstract:	Image classification remains a central challenge in computer vision, particularly in medical domains where diagnostic accuracy and robust feature representation are critical. Recent advances in deep learning have highlighted the complementary strengths of Convolutional Neural Networks in capturing localized spatial patterns and Vision Transformers in modeling global contextual dependencies. In this work, we propose a hybrid CNN–ViT architecture enhanced with cross-attention mechanisms for medical image classification. The framework begins by extracting complementary features through parallel CNN and ViT backbones, then employs bidirectional cross-attention to fuse these representations. A dedicated segmentation branch generates precise region-of-interest masks, which guide ROI-focused feature extraction to concentrate on diagnostically relevant areas. These segmented regions are encoded and combined with the attended features for final classification. Experimental results on chest X-ray datasets demonstrate that our approach outperforms standalone CNN and ViT baselines, achieving superior accuracy while providing interpretable segmentation maps. The framework’s progressive design-extracting, fusing, segmenting, and classifying-enables focused analysis of critical regions, making it particularly suitable for medical imaging applications where precise localization and classification are paramount.
Download

Paper Nr:	631
Title:	MTLD-HSD: A Hierarchical Multi-Task Three-Stream Model for Explainable Hate Speech Detection in Low-Resource Tunisian Arabic
Authors:	Abdelkarim Mars and Samia Ben Ismail
Abstract:	Detecting hate speech in low-resource Arabic dialects remains challenging due to code-switching, nonstandardized spelling, and Arabizi (Latin-script Arabic). We present MTLD-HSD, a multi-task learning framework for Tunisian Arabic hate speech detection with four contributions: (1) a three-stream architecture combining dialectal, semantic, and affective encoders fused via graph attention; (2) joint optimization of hate classification, toxicity intensity regression, and target identification; (3) the largest Tunisian hate speech corpus (32,726 samples); and (4) integrated explainability through attention visualization and SHAP analysis.Our dialectal encoder achieves 94% Arabizi normalization accuracy via seq2seq modeling. The complete system achieves 96% weighted F1-score, outperforming AraBERTv2 by 7% absolute and showing particular strength on implicit hate (+12%) and minority classes (93% F1 on hate). Zero-shot transfer to Moroccan and Algerian Arabic yields 78–81% F1, demonstrating cross-dialect generalization. Ablation studies confirm each component’s contribution, and bias audits show consistent performance across demographic groups.
Download

Paper Nr:	634
Title:	Agent-as-a-Graph: Knowledge Graph-Based Tool and Agent Retrieval for LLM Multi-Agent Systems
Authors:	Faheem Nizar, Elias Lumer, Anmol Gulati, Pradeep Honaganahalli Basavaraju and Vamse Kumar Subbiah
Abstract:	Recent advances in Large Language Model Multi-Agent Systems enable scalable orchestration and retrieval of specialized, parallelized subagents, each equipped with hundreds or thousands of Model Context Protocol (MCP) servers and tools. However, existing agent, MCP, and retrieval methods typically match queries against a single agent description, obscuring fine-grained tool capabilities of each agent, resulting in suboptimal agent selection. We introduce Agent-as-a-Graph retrieval, a knowledge graph retrieval augmented generation approach that represents both tools and their parent agents as nodes and edges in a knowledge graph. During retrieval, i) relevant agents and tool nodes are first retrieved through vector search, ii) we apply a type-specific weighted reciprocal rank fusion (wRRF) for reranking tools and agents, and iii) parent agents are traversed in the knowledge graph for the final set of agents. We evaluate Agent-as-a-Graph retrieval on the LiveMCPBench benchmark, achieving 14.9% and 14.6% improvements in Recall@5 and nDCG@5 over prior state-of-the-art retrievers, and 2.4% improvements in wRRF optimizations.
Download

Paper Nr:	644
Title:	Reinforcement Learning-Guided Deep Neural Networks with Particle Swarm Optimization for Early Sepsis Prediction in Intensive Care Units
Authors:	Yassin Ben Youssef, Asma Amdouni and Ghaith Manita
Abstract:	Sepsis remains a critical cause of mortality in intensive care units (ICUs), where early and accurate prediction is essential to improving patient outcomes. Existing machine learning approaches often rely on suboptimal hyperparameter configurations and typically fail to balance multiple clinical objectives simultaneously. This work presents a novel framework that integrates deep neural networks with Particle Swarm Optimization (PSO) for automated hyperparameter tuning, combined with Reinforcement Learning (RL) for dynamic refinement of prediction thresholds. The PSO component optimizes a weighted composite objective function that combines classification accuracy, sensitivity, and specificity, while the RL agent adapts decision thresholds based on patient-specific risk trajectories. Evaluated on a comprehensive sepsis dataset comprising 110,204 ICU patient records, the optimized deep learning model achieves 96.8% accuracy, 95.3% sensitivity, and 97.2% specificity, outperforming five state-of-the-art baseline methods across all metrics. The framework further incorporates SHAP and LIME to generate clinically interpretable explanations at both global and patient levels. Experimental results indicate that PSO reduces false alarms by 42% compared to grid search while identifying near-optimal neural architectures 8.3 times faster. Taken together, these contributions constitute one of the first comprehensive integrations of metaheuristic optimization, reinforcement learning, and explainable deep learning for sepsis prediction, yielding a practical, interpretable, and highly accurate clinical decision support tool for ICU settings.
Download

Paper Nr:	646
Title:	Adaptive Hybrid Genetic Algorithm with Q-Learning for Constrained Project Portfolio Optimization
Authors:	Zied Ben Youssef, Yosr Slama and Ghaith Manita
Abstract:	Project portfolio optimization requires selecting a subset of candidate projects that maximizes expected returns while satisfying budget and risk constraints and accounting for interaction effects such as synergies. This paper proposes QL-HGA, an adaptive hybrid genetic algorithm with Q-learning and opposition-based diversity enhancement, for constrained project portfolio optimization. In QL-HGA, a Q-learning agent dynamically selects crossover and mutation operators according to the current population state, opposition-based learning is invoked to restore diversity when convergence stalls, a greedy repair mechanism enforces feasibility, and a local search operator refines promising offspring. The formulation captures pairwise synergy costs, risk penalties, and budget limits within a binary optimization framework. Computational experiments on benchmark portfolios with varying sizes compare QL-HGA against several state-of-the-art metaheuristics. The results indicate that QL-HGA achieves superior mean fitness with competitive stability across configurations, and non-parametric tests with Vargha–Delaney effect sizes point to at least a moderate practical advantage over all competitors.
Download

Paper Nr:	650
Title:	Text to Process Model: An Agentic Pipeline for BPMN Generation
Authors:	Amal Tarifa and Imed Kouki
Abstract:	The direct application of Large Language Models (LLMs) to generate Business Process Model and Notation (BPMN) diagrams is often undermined by a lack of determinism, leading to unverifiable outputs. To overcome this limitation, we introduce an agentic approach that transforms BPMN generation from a probabilistic task into a verifiable synthesis pipeline. Our work begins by decomposing the generation process into discrete, inspectable stages, validated through ablation studies to ensure the necessity of each module. We further introduce an adaptable intermediate representation that serves as a schema-validated contract between stages. Finally, we provide an empirical evaluation using process descriptions from the PMo Dataset. By enforcing logical integrity at intermediate steps through systematic prompt engineering techniques, our framework consistently produces structurally sound and valid BPMN diagrams.
Download

Paper Nr:	652
Title:	Ontology-Guided Hierarchical Preference Elicitation with Bayesian Active Querying
Authors:	Anton Agafonov, Andrew Ponomarev and Alexander Smirnov
Abstract:	Preference elicitation is a central component of modern decision support systems, particularly in settings where user needs must be inferred from sparse and uncertain signals. Psychological and decision-theoretic studies emphasize that preferences are often constructed during interaction rather than fixed in advance, making the elicitation process a critical determinant of decision quality. Recent advances in large language models (LLMs) have expanded the possibilities for natural-language interaction, yet semantic interpretation alone is insufficient in domains where the decision space is structured through ontologies or hierarchical taxonomies. In such settings, effective elicitation must account not only for semantic proximity between the query and candidate categories, but also for their positions and relationships within the hierarchy. This paper introduces a framework that operates on the results of semantic retrieval and iteratively refines user intent by actively querying informative branching points of an ontology. Bayesian belief tracking is employed to guide query selection at hierarchical decision nodes, enabling principled exploration-exploitation trade-offs during refinement. Starting from an initial set of semantically relevant categories, the method identifies informative decision nodes, queries user feedback, and progressively narrows the search space. Experiments on a multilingual product taxonomy show that the proposed ontology-guided refinement approach yields more stable and interpretable results than purely semantic retrieval baselines.
Download

Paper Nr:	655
Title:	Hidden in Plain Pixels: Deep Learning and Dynamic Encryption for Secure Image and Text Steganography
Authors:	Naman Sharma, Christoph Schommer and Aria Nourbakhsh
Abstract:	This paper presents a dual-module steganographic framework that conceals either full RGB images or textual messages inside cover images while preserving imperceptibility and enabling reliable recovery. The imagein-image module employs encoder–decoder networks based on the U-Net family to disperse the secret across multi-scale features; a lightweight CNN decoder reconstructs the hidden image. The text-in-image module converts input text to ASCII, maps it to 8-bit binary, pads as needed, and reshapes it into a 256 × 256 × 3 payload image prior to embedding and recovery. To strengthen confidentiality without sacrificing capacity, we apply a dynamic XOR scheme combined with pixel-block shuffling using per-sample seeds before embedding. We study U-Net, U-Net++, and VNet encoders and examine encoder/decoder loss pairings to characterize trade-offs between stego quality and payload fidelity. Evaluation on 256×256 RGB data reports MSE, PSNR, and SSIM for image fidelity, as well as text accuracy and BLEU for language reconstruction. Results indicate robust imperceptibility, dependable decoding, and a favorable fidelity–robustness balance with U-Net++ while maintaining key-based confidentiality.
Download

Paper Nr:	669
Title:	A Knowledge-Based Book Recommender through Enhanced Retrieval-Augmented Generation
Authors:	Vlada Khomenko, Christoph Schommer and Salima Lamsiyah
Abstract:	Recommender systems (RS) address choice overload and reduce friction in discovering new items by providing users with suggestions that match their interests. Traditional RS, which rely on user-item interaction history or structured metadata, often struggle with cold-start scenarios and limited transparency. This paper addresses these limitations by proposing a knowledge-based book recommendation system that operates on unstructured natural-language item descriptions. Leveraging the capabilities of large language models (LLMs), the system employs a retrieval-augmented generation (RAG) architecture with a hybrid semantic-lexical retrieval setup. It enhances retrieval through query expansion, topic filtering, and cross-encoder re-ranking, and integrates chain-of-thought (CoT) prompting for more predictable, transparent, and structured generation. The system is comprehensively evaluated across retrieval, generation, and user-centered metrics. Results demonstrate that each proposed enhancement contributes incrementally to the system performance, while user feedback confirms its usefulness and the value of transparent natural-language explanations. Overall, this work demonstrates the feasibility and effectiveness of a cold-start-capable, RAG-based recommendation approach that leverages unstructured data, providing a foundation for broader applications across other domains. A demo of the system is available (https://github.com/VladaK16/book-recommender.git).
Download

Paper Nr:	670
Title:	Agentic Health Analytics for Personalized Insights from Wearable Data
Authors:	Hamza Haruna Mohammed, Gabriel Kiss, J. Artur Serrano and Frank Lindseth
Abstract:	Wearable devices generate large volumes of physiological time-series data, yet transforming these signals into actionable and trustworthy health insights remains challenging due to fragmented analytics pipelines and opaque language-model–based recommendations. This paper presents Agentic Health Analytics (AHA), a modular analytics and reasoning pipeline for converting heterogeneous wearable data into explainable, personalized well-being insights. The core contribution is a structured agentic reasoning layer that explicitly decomposes wearable analytics into data-grounded facts, physiological rationale, and actionable micro-habits prior to natural-language generation. By separating numerical signal processing, interpretation, and text generation, the approach preserves traceability from raw data to recommendations, mitigates hallucinations, and supports safety-aware personalization, limitations commonly observed in end-to-end LLM-based solutions. The pipeline integrates multi-scale time-series analysis with constrained large language model decoding to ensure alignment between computed trends and generated narratives. We evaluate the approach on real-world, multi-device wearable datasets and compare it against traditional time-series models, end-to-end LLMs, and retrieval-augmented generation. Results demonstrate improved numerical grounding, higher trend consistency, and zero safety violations. The proposed approach provides a transparent and extensible analytics and reasoning component for personalized wearable health applications.
Download

Paper Nr:	681
Title:	Graph Neural Networks for Credit Propensity
Authors:	Lucas de Moraes Pinto Pereira, Guilherme S. Rodrigues and Raul Yukihiro Matsushita
Abstract:	Credit propensity modeling, understood as estimating the probability that a customer will contract a specific credit product within a given business horizon, is a core component of retail banking decisioning. In practice, banks usually rely on flexible tabular models trained on individual features, which ignore how customers are connected to each other. Graph Neural Networks (GNNs) can exploit this relational structure, but in real credit settings it is still unclear how to construct the graph and which architectural choices actually matter. In this work we compare strong tabular baselines with a family of GNN models that incorporate two types of relational information: household and kinship links between customers, and peer-to-peer payment flows. Using proprietary data from a large Brazilian retail bank, we evaluate multiple backbones (GCN, GIN, GINE, R-GCN, R-GAT) and readouts. Our results show that turning on a simple homogeneous GNN already yields a substantial and consistent lift over the tabular Random Forest baseline, with comparable or lower training cost. Modeling relation types brings smaller but systematic gains, while the impact of adding directed transaction edges depends strongly on the backbone. Overall, the study provides evidence that graph-based models can substantially improve credit propensity prediction in practice and clarifies which graph and architecture choices are most beneficial in a banking environment.
Download

Paper Nr:	686
Title:	FGAN-AD-U: Unsupervised Multi-Scale Graph-Transformer Reconstruction for Industrial Anomaly Detection
Authors:	Sri Vishnu Gopinathan, Akkasha Latif and Bushara A. R.
Abstract:	One of the major challenges in modern manufacturing is to achieve reliable and automated in-line quality inspection, and this can be done by accurately detecting very subtle defects appearing in high-resolution industrial images without depending on annotated anomalies. This paper introduces FGAN-AD-U, a fully unsupervised anomaly detection framework that jointly models the structural and contextual geometry of normal industrial images. The proposed method integrates multi-scale relational graph construction with a dual-path GNN-Transformer reconstruction module, enabling simultaneous capture of fine-grained local relationships and long-range global dependencies. Patch embeddings extracted from ResNet-50 FPN layers are organized into a hierarchical graph pyramid, where a 3-layer GCN encodes relational consistency and a Transformer encoder models global contextual patterns. Anomalies are detected through reconstruction inconsistency between original and fused representations, eliminating the need for labels, masks, or handcrafted priors. Extensive experiments on two challenging benchmarks, MVTec AD and Texture-A, D-demonstrate that FGAN-AD-U achieves state-of-the-art performance, with 99.4% image-level AUROC and 99.2% pixel-level AUROC on MVTec AD, and 96.1% AUROC on Texture-AD. Ablation studies highlight the complementary contributions of multi-scale graph reasoning and global Transformer modeling. Overall, FGAN-AD-U provides a robust, generalizable, and scalable solution for real-world industrial inspection tasks.
Download

Paper Nr:	695
Title:	MPSync: A Parallel Framework for Set-Wise Clustering and Classification on Large-Scale and Multi-Port Data Streams
Authors:	Aarti, Jagat Sesh Challa, Mridul Chandak, Utkarsh Darolia, Kanav Kapoor, Navneet Goyal and Poonam Goyal
Abstract:	The exponential growth of data generated from IoT, social media, and real-time analytics necessitates scalable and efficient techniques for processing large-scale data streams in real-time. Traditional set-wise clustering and classification algorithms, which group sets of data objects based on distribution patterns, struggle to manage large-scale and multi-port data streams effectively in dynamic environments. To address this, we propose MPSYNC, a novel parallel framework for both set-wise clustering and classification over distributed memory architectures. It is designed for effectively handling multi-port data streams by distributing them across multiple processors. Unlike existing set-wise clustering and classification methods, MPSYNC incorporates a temporal feature through a Tilted-Time Window Framework (TTWF), enabling users to select a specific time window across multiple windows to provide the requested clustering or classification outputs. Experimental results demonstrate that MPSYNC exhibits substantial improvements in processing speed and scalability compared to state-of-the-art set-wise methods, rendering it highly effective for real-time analytics on massive and dynamic data streams.

Paper Nr:	701
Title:	Privacy Risk Predictions Based on Fundamental Understanding of Personal Data and an Evolving Threat Landscape
Authors:	Haoran Niu and K. Suzanne Barber
Abstract:	It is difficult for individuals and organizations to protect personal information without a fundamental understanding of relative privacy risks. By analyzing over 5,000 empirical identity theft and fraud cases, this research identifies which types of personal data are exposed, how frequently such exposures occur, and what the consequences of those exposures are. We construct an Identity Ecosystem graph-a foundational, graph-based model in which nodes represent personally identifiable information (PII) attributes and edges represent empirical disclosure relationships between them (e.g., one PII attribute is exposed due to the exposure of another). Leveraging this graph structure, we develop a privacy risk prediction framework that uses graph theory and graph neural networks to estimate the likelihood of further disclosures when certain PII attributes are compromised. The results show that our approach effectively addresses the core question: Can the disclosure of a given identity attribute possibly lead to the disclosure of another attribute? The code for the privacy risk prediction framework is available at: https://github.com/niu-haoran/ Privacy-Risk-Predictions-and-UTCID-Identity-Ecosystem.git.
Download

Paper Nr:	702
Title:	RNM-TD3: N:M Semi-Structured Sparse Reinforcement Learning from Scratch
Authors:	Isam Vrce, Andreas Kassler and Gökçe Aydos
Abstract:	Sparsity is a well-studied technique for compressing deep neural networks (DNNs) without compromising performance. In deep reinforcement learning (DRL), neural networks with up to 5% of their original weights can still be trained with minimal performance loss compared to their dense counterparts. However, most existing methods rely on unstructured fine-grained sparsity, which limits hardware acceleration opportunities due to irregular computation patterns. Structured coarse-grained sparsity enables hardware acceleration, yet typically degrades performance and increases pruning complexity. In this work, we present, to the best of our knowledge, the first study on N:M structured sparsity in RL, which balances compression, performance, and hardware efficiency. Our framework enforces row-wise N:M sparsity throughout training for all networks in off-policy RL (TD3), maintaining compatibility with accelerators that support N:M sparse matrix operations. Experiments on continuous-control benchmarks show that RNM-TD3, our N:M sparse agent, outperforms its dense counterpart at 50%-75% sparsity (e.g., 2:4 and 1:4), achieving up to a 14% increase in performance at 2:4 sparsity on the Ant environment. RNM-TD3 remains competitive even at 87.5% sparsity (1:8), while enabling potential training speedups.
Download

Paper Nr:	703
Title:	AMP–SRD: An Agentic Multi-Pipeline Framework for Characterizing Semantic, Structural, and Performance Regressions Introduced by Code Changes
Authors:	Sabrine Boussema, Brahim Hnich, Ali Ben Mrad and Mohamed Wiem Mkaouer
Abstract:	Software is no longer written by humans alone. Developers and intelligent agents now co-author the same codebases, accelerating delivery but also multiplying subtle risks. Small edits can silently shift semantics, break implicit contracts at call sites, or erode performance long before benchmarks or users reveal the damage. Traditional validation in CI/CD-while effective for syntax and unit checks-struggles to capture semantic drifts and latent slowdowns without explicit oracles. This paper presents AMP-SRD, an agentic, multi-pipeline framework that observes code as it evolves. By juxtaposing what a program used to do with what it now does, AMP-SRD characterizes regressions along three axes: semantic behavior, call-site compatibility, and performance stability. Beyond flagging issues, it produces explainable narratives and actionable repair hints. We evaluate AMP-SRD on a curated, commit-level dataset and report level-wise metrics, showing that a semantic-first, cooperative-agent approach can detect regressions early, without relying on preexisting test suites or handcrafted specifications. In line with common early-development practice, we assume a stable functional specification S0. The code revisions under analysis are therefore expected to preserve the same intended behaviour. Any observed semantic, structural, or performance deviation is interpreted as a regression rather than an intentional requirement evolution.
Download

Paper Nr:	706
Title:	Handling Uncertainty in Waste Streams: Possibilistic Aggregation of Deep Experts
Authors:	Samar Daou, Jihen Frikha Elleuch, Mouna Zouari Mehdi, Dorra Sellami, Salwa Sahnoun, Ahmed Fakhfakh and Khaled Elleuch
Abstract:	Automatic classification of waste materials is a critical component of modern intelligent recycling systems. However, plastic waste stream images exhibit high variability, visual ambiguity, and frequent label uncertainty, making conventional probabilistic deep classifiers unreliable. In this work, we introduce a possibilistic deep ensemble designed to handle uncertainty by replacing standard softmax-based aggregation with a possibility-driven fusion scheme. Our approach combines homogeneous deep experts whose outputs are transformed into normalized possibility distributions, then aggregated using a compromise possibilistic fusion to preserve plausibility without probability dilution. Additionally, we enrich the training dataset using a StyleGAN-based synthetic generator, improving robustness to real-world variations and rare cases. Experiments on a multi-class plastic waste dataset show consistent improvements over single model, especially in ambiguous and under-represented categories such as LDPE. The proposed method achieves an accuracy of 98.93%, along with higher macro-F1, demonstrating the relevance of possibility theory for uncertainty handling in waste classification.
Download

Paper Nr:	707
Title:	An Operationalization of Contestation for Decision-Making Systems
Authors:	Christodoulos Ioannou and Loizos Michael
Abstract:	Ensuring transparency and accountability in automated decision-making systems remains a central challenge, particularly when such systems rely on opaque or black-box models. While contestability has emerged as a promising paradigm for addressing this challenge, it is often discussed at a conceptual level without a clear path to implementation. In this work, we operationalize contestability by providing explicit answers to the foundational design questions that a contestable system must address: who the parties in the process are, what their epistemic states entail, how they may interact, what communication language structures their exchange, and what actions may follow the contestation. We instantiate these components within a prioritized argumentation framework grounded in predicate logic, enabling structured justification, challenge, and revision of system decisions. Through a loan-application use case and a prototype implementation, we demonstrate how this operationalization supports transparent, auditable, and adaptive AI decision-making. A software system built on this operationalization allows real human users to engage in contesting dialogues and provides the infrastructure for future empirical evaluation.
Download

Paper Nr:	710
Title:	Distinguishing Grammatical Errors from Code-Switching in Multilingual Learner Texts
Authors:	Helmi Baazaoui, Lilia Cheniti Belcadhi, David Laiymani and Christophe Guyeux
Abstract:	Grammatical Error Correction (GEC) is essential in educational NLP but struggles with learner writing containing code-switching (CSW) mixing languages within a sentence. Traditional monolingual GEC models often mislabel valid multilingual spans as errors and lack pedagogical feedback. We propose a novel three-stage curriculum learning framework that distinguishes grammatical errors from code-switches at the token level. Starting with synthetic CSW-GEC data generated via linguistically informed span injection, we create two human-annotated English–French and English–Arabic benchmarks. Fine-tuning multilingual Transformers (XLM-R, mDeBERTa) with staged supervision improves fluency, error sensitivity, and robustness. Results show curriculum models outperform baselines, with mDeBERTa achieving top F0.5 scores. Error analysis highlights challenges at GE–CSW boundaries, guiding future dataset expansion and finer error classification. This work supports Intelligent Tutoring Systems that correct without penalizing code-switching, enabling inclusive language learning.
Download

Paper Nr:	713
Title:	Optimizing Prompts Efficiently with Iterative Determinantal Point Processes
Authors:	Nahid Abdollahi, Sahar Vahdati, Mehdi Eftekhari and Jens Lehmann
Abstract:	The performance of large language models relies heavily on effective prompt engineering. Hand-crafted methods, such as Plan-and-Solve, have improved reasoning capabilities but often remain inefficient at ensuring diversity in generated prompts, thereby limiting their overall effectiveness. PromptBreeder (PB) addressed this limitation by introducing a self-referential self-improvement mechanism that evolves prompts using a binary tournament genetic algorithm. This approach enables PB to iteratively explore the prompt space, optimizing both diversity and performance. Similarly, GEPA focuses on evolving and refining prompts through a genetic algorithm. While PB and GEPA have achieved significant advancements in prompt generation, they introduce a new challenge in the form of increased computational overhead and algorithmic complexity, which makes them less practical for many real-world applications. To address these issues, we propose DPPROMPT (DPP + Prompt), a method that optimizes prompts using a local search strategy. It employs Determinantal Point Processes (DPPs) to select diverse and high-quality prompts, effectively balancing performance and diversity without relying on complex self-referential mechanisms. We evaluated DPPROMPT on ten benchmarks, namely MultiArith, SingleEq, AddSub, SVAMP, SQA, CSQA, AQuA-RAT, GSM8K, MATH500, and GPQA, using three LLMs: Mistral, Llama, and Qwen. The results demonstrate strong reasoning performance and flexibility across different model architectures. By achieving results comparable to PromptBreeder while significantly reducing computational requirements, DPPROMPT provides a practical and efficient solution for automated prompt generation.

Paper Nr:	716
Title:	Artificial Intelligence-Based Multimodal Alzheimer Prediction System (AI-MAPS): A System Overview
Authors:	Achref Kassaoui, Imen Abdennadher and Ismael Bouassida Rodriguez
Abstract:	Alzheimer’s disease (AD) affects over 50 million people worldwide, making early detection and personalized risk assessment crucial. We present AI-MAPS (Artificial Intelligence-based Multimodal Alzheimer Prediction System), which integrates ensemble classification, Cox survival modeling, and a CN→MCI transformation model to predict disease progression. Using data from 2,430 patients and 16,421 longitudinal visits in ADNI, the system extracts 40+ multimodal features and achieves 94.7% classification accuracy, with concordance indices of 0.783 (CN→MCI) and 0.843 (MCI→AD). A web-based Gradio interface provides clinicians with actionable insights—current diagnosis, 5-year progression probabilities, hazard ratios, and Kaplan–Meier curves—without programming. AI-MAPS offers calibrated, interpretable, and clinically usable predictions, simulating complete CN→MCI→AD trajectories and supporting informed decision-making in AD care.
Download

Short Papers

Paper Nr:	18
Title:	Evaluation of Communicative Health Literacy and Textual Coherence from Unlabeled Texts: Sentence-Level Segmentation and Meta-Clustering Based Approach
Authors:	Mouheb Mehdoui, Amel Fraisse, Widad Mustafa El Hadi and Mounir Zrigui
Abstract:	This paper addresses the challenge of measuring communicative health literacy in unstructured user-generated content, such as social media posts, where traditional readability metrics often fail to capture the dynamic complexity of the language. A novel method is proposed that combines sentence-level readability analysis with meta-clustering to evaluate how readability evolves throughout a text. Two key metrics are introduced: the Readability Variance Score (RVS), which measures the consistency of sentence-level readability, and the Readability Transition Score (RTS), which assesses the smoothness of transitions between sentences. These metrics facilitate the classification of texts into six communicative health literacy profiles, ranging from “Fluent Communicator” to “Struggling Reader.” The study analyzes a large-scale dataset of 1 million English posts, showing that while most users generate moderately readable content, their discourse often lacks structural coherence. These findings highlight the need for personalized interventions to enhance health literacy. This approach contributes to health communication research by providing fine-grained insights into textual coherence and promoting improved e-accessibility in English digital health environments.
Download

Paper Nr:	23
Title:	AI-Driven PCB Assembly Defect Detection Using Hybrid Deep Learning Architectures
Authors:	Mustafa Yasin Erten
Abstract:	This study proposes a deep learning–based image classification system to assess the assembly accuracy of electronic components on printed circuit boards (PCBs) and to automatically detect errors in manufacturing processes. In this context, Vision Transformer (ViT), Xception, ResNet-50, MobileNetV2, and a custom-designed CNN architecture were implemented for defect detection. A hybrid architecture was also created by combining the custom CNN structure with the Xception and MobileNetV2 models; in this hybrid design, multi-level feature fusion strategies were applied to enhance the model’s overall accuracy and generalization capability. The results indicate that, in particular, the ensemble architecture and the ViT architecture exhibit high accuracy and stability, achieving performance superior to that of conventional models. This paper demonstrates the effectiveness of AI-based systems for the early and automatic detection of PCB manufacturing defects and contributes to industrial quality control processes. The findings demonstrate not only high classification performance but also highlight the broader potential of artificial intelligence technologies to transform automated quality inspection and reliability assurance in modern electronic manufacturing.
Download

Paper Nr:	28
Title:	Evaluating Customer Service with Large Language Models: A Case Study
Authors:	Matheus Menezes, Ilana Santos and Lucas Pinheiro
Abstract:	Customer service is a primary activity for companies that want to remain competitive. Through customer interaction, companies can extract and analyze data that help to capture insights and ultimately improve their activities. These customer data can vary from structured information, such as quantitative ratings of products and services, to unstructured data, including customer feedback, chat logs, and media interactions. The latter can be particularly challenging due to the complexities of natural language processing. The progress of generative artificial intelligence, particularly the Large Language Models (LLMs), has enabled the development of highly contextual applications to support companies in decision-making. However, the efficacy of LLM depends on factors such as data quality, prompt engineering, and fine-tuning strategies. To understand the advances and limitations of LLMs in customer services, we developed an open-source Llama 3 agent that assesses customer service representatives’ behavior in a multi-utility call center. The artificial agent analyzes transcribed phone recordings between the representatives and the customers and then answers a quality report with 18 binary yes or no questions. Based on 7,765 customer service analyses, our results indicate the agent’s potential to conduct the task, with high accuracy, precision, recall, and F1-score across most question categories. However, we observed that semantically complex questions and a low number of true negatives (imbalanced dataset) led to random predictions, particularly in the negative class. We address these challenges in detail, highlighting relevant metrics and potential solutions that similar tasks can adopt. Through this case study, we hope to provide insight into how companies can leverage generative solutions to support decision-making processes.
Download

Paper Nr:	39
Title:	From Synthetic Code to Real Threats: Multi-Label Vulnerability Detection with JavaBERT for Cyber Threat Intelligence
Authors:	Othmane Cherqi, Anass Sebbar, Barhim Anegdouil and Mohammed Boulmalf
Abstract:	Large code-language models hold promise for supporting cyber threat intelligence (CTI) by identifying vulnerabilities before they are weaponized. Yet, most evaluations rely on synthetic benchmarks (e.g., Juliet), far from the complexity of real-world exploits represented in datasets such as OWASP. In this work, we revisit the problem of domain shift between synthetic and operational data, reframing vulnerability detection as a multi-label Common Weakness Enumeration (CWE) classification task with an additional binary “vul-nerable” flag. Using JavaBERT, a lightweight transformer variant of GraphCodeBERT, we demonstrate that simple yet effective strategies i.e., multi-label stratified sampling, sliding-window pooling, and class-weighted loss-achieve strong performance: 0.96 macro-F1 on Juliet and 0.50 macro-F1 0.82 binary F1 on OWASP, competitive with graph-based models while retaining transformer simplicity. Beyond benchmarks, we show how our findings inform CTI workflows: identifying dataset leakage risks, quantifying limits of cross-dataset generalization, and highlighting gaps that CTI analysts must address when integrating AI into vulnerability intelligence pipelines. We argue that bridging synthetic and real-world code analysis is a key step toward AI-augmented OSINT and CTI platforms that can track vulnerabilities across ecosystems, anticipate exploitation patterns, and ultimately reduce attacker advantage.
Download

Paper Nr:	43
Title:	An Ensemble Deep Reinforcement Learning Model for Automated Cryptocurrency Trading
Authors:	Ameni Youssfi Nouira
Abstract:	The possibility of automated bitcoin trading to improve trading strategies in volatile markets is attracting a lot of interest. In order to improve Bitcoin trading performance, this study proposes an ensemble deep reinforcement learning (DRL) system that combines Soft Actor Critic (SAC) with Proximal Policy Optimization (PPO). Although SAC uses entropy regularization to enhance exploration and adaptability in volatile market situations, PPO guarantees stable policy updates. Our approach uses market data to guide decisions with the goal of minimizing risk and maximizing cumulative returns. Comparing the suggested ensemble model to individual DRL algorithms, experimental findings show how successful it is. These results demonstrate the potential of ensemble reinforcement learning in applications related to cryptocurrency trading.
Download

Paper Nr:	51
Title:	zkUIT: Private, Portable Diploma Verification via Set-Membership zkSNARKs and a Succinct On-Chain Verifier
Authors:	Khoa Tan Vo, Hong-Tri Nguyen and Tu-Anh Nguyen-Hoang
Abstract:	Academic credential verification remains fragmented, cost-intensive, and privacy-sensitive. Many blockchain approaches either expose metadata on-chain or rely on off-chain trust hubs that reintroduce single points of failure. We present zkUIT, a minimal on-chain verifier that realizes set-membership proofs via a zkSNARK over Poseidon-based Merkle commitments, with a single public input [root], no diploma attributes on-chain, and policy enforced off-chain for portability across EVM rollups. A reference implementation on Sepolia testnets (Ethereum Sepolia as L1 and zkSync Sepolia as L2) shows holder-side proving of 1.09 ± 0.02 s (mean ± sd, n=10) and L2 verification latency of ≈ 10.3±0.5 s with a per-call cost of ≈ 2.89×10−4 ETH. By contrast, L1 exhibited highly variable fees (peaking at ≈ 1.74×10−2 ETH) and intermittent failures (up to 50%) in our sampled windows. By minimizing on-chain state and decoupling policy, zkUIT achieves privacy-by-construction and predictable operations on L2 rollups. zkUIT complements rather than replaces institutional portals, while agent-mediated workflows can layer issuer allowlists, root freshness, and anomaly detection off-chain.
Download

Paper Nr:	52
Title:	Last Skills Standing: Adaptive Skill Discovery in Unsupervised Reinforcement Learning
Authors:	Gayathri Rangu and Shivashankar B. Nair
Abstract:	Unsupervised Skill Discovery using Reinforcement Learning (RL) often suffers from redundant and lowquality skill repertoires. This is largely due to static skill sets and the unavailability of online evaluation metrics during the training process. We propose Discriminator-based Pruning (DiP), an adaptive skill discovery framework that leverages discriminator confidence trends and temporal stability analyses to refine skill sets during training, combined with Skill-Fair Episode Budgeting (SFEB) to ensure fair training exposure. Our method, DiP + SFEB, adaptively identifies and removes redundant or under-performing skills during training itself. This led to substantially more compact and diverse repertoires. Results indicate that across diverse continuous-control benchmarks, DiP+SFEB improves skill discriminability by up to 54% relative to baselines, while reducing training episodes by an average of 23.8% (with savings as high as 47.2% in some environments), thus producing higher-quality skills and hastening learning convergence. This approach prevents mode collapse by stabilizing the discriminator and producing distinguishable, interpretable skills across environments of varying complexity. From the results, one can conclude that augmenting online skill-pruning with fair budgeting can significantly enhance the efficiency and diversity of unsupervised skill discovery in RL.
Download

Paper Nr:	58
Title:	Structured Relational Representations
Authors:	Arun Kumar and Paul Schrater
Abstract:	Invariant representations are core to representation learning, yet a central challenge remains: uncovering invariants that are stable and transferable without suppressing task-relevant signals. This raises fundamental questions, requiring further inquiry, about the appropriate level of abstraction at which such invariants should be defined and which aspects of a system they should characterize. Interpretation of the environment relies on abstract knowledge structures to make sense of the current state, which leads to interactions, essential drivers of learning and knowledge acquisition. Interpretation operates at the level of higher-order relational knowledge; hence, we propose that invariant structures must be where knowledge resides, specifically as partitions defined by the closure of relational paths within an abstract knowledge space. These partitions serve as the core invariant representations, forming the structural substrate where knowledge is stored and learning occurs. On the other hand, inter-partition connectors enable the deployment of these knowledge partitions encoding task-relevant transitions. Thus, invariant partitions provide the foundational primitives of structured representation. We formalize the computational foundations for structured relational representations of the invariant partitions based on closed semiring, a relational algebraic structure.
Download

Paper Nr:	62
Title:	Managing Critical Resources while Learning New Skills: A Transfer Approach for Hierarchical RL
Authors:	Thibault Roux, Filipo S. Perotto, Jean-Loup Farges and Gauthier Picard
Abstract:	Existing research in Deep RL often addresses either safety or transfer learning separately. However, the crucial challenge of preserving safety while transferring knowledge to learn new skills remains underexplored. This paper tackles the problem of skill acquisition within a sequential transfer learning context with a safety constraint: managing critical resources that should never be depleted during learning. We propose a novel approach integrating Hierarchical Reinforcement Learning for task decomposition, Transfer Learning to leverage a trustworthy previously trained source policy, and Elastic Weight Consolidation to protect safety-critical parameters. Our method is empirically validated in a 2D drone simulation involving navigation, resource management, and adaptation from a source task (navigation) to a target task (package delivery). Experimental results demonstrate that our method successfully acquires the new package delivery skill while preserving the critical battery management behavior learned in the source task throughout the transfer process. The complete source code is accessible here.
Download

Paper Nr:	63
Title:	M3DNet: Minimalist 3D Detection Backbone for Efficient and Accurate 3D Object Detection
Authors:	Thurimerla Prasanth, Ram Prasad Padhy and B. Sivaselvan
Abstract:	We present M3DNet, a minimalist and computationally efficient backbone architecture tailored for 3D object detection (3D-OD) task using LiDAR point cloud data. Unlike traditional backbones that rely heavily on dense convolutional stacks or computationally expensive attention mechanisms, M3DNet leverages depthwise separable convolutions to significantly reduce the number of parameters. The lightweight nature of the backbone makes it especially suitable for deployment in real-time applications such as autonomous vehicles. The backbone is constructed by integrating lightweight depthwise convolutional blocks in each stage of the feature extraction pipeline. The backbone preserves spatial detail while minimizing computational overhead. We used a dual kernel configuration in regular convolutional blocks to learn multi-level features. M3DNet is able to maintain strong feature representation capabilities necessary for accurate 3D localization and classification. We evaluate our method on the widely used KITTI 3D object detection benchmark. Experimental results demonstrate that M3DNet achieves competitive accuracy while reducing the model size by nearly 50 percent compared to conventional backbones. Our M3DNet is listed on the official KITTI leaderboard. We conducted several ablation experiments and reported the results to validate the efficiency of the backbone. The proposed M3DNet offers a strong alternative to heavier backbones by balancing accuracy and parameters, which makes it well-suited for scalable deployment of lightweight perception systems.
Download

Paper Nr:	78
Title:	Dynamic Weighting with Sinusoidal Modulation for Rough Inclusion Classifiers
Authors:	Piotr Artiemjew and Artur Samojluk
Abstract:	We introduce new weighted classification techniques based on weak variants of rough inclusions by enriching classical 8v1.1–8v1.5 algorithms with sinusoidal weight modulation. The proposed 9v1.1–9v1.4 classifiers dynamically adjust per-attribute contributions of training objects according to descriptor distances, while incorporating cyclic rewards and penalties. This modulation increases sensitivity to local data structure and provides more flexible weighting schemes than fixed monotone transformations. We evaluate the methods on five benchmark UCI data sets using Monte Carlo Cross Validation. The results demonstrate that our classifiers are at least comparable to earlier variants, and in several cases-such as Pima Indians Diabetes and Hepatitis-achieve improved balanced accuracy and better detection of positive cases. These findings indicate that sinusoidal modulation is a promising extension of rough inclusion–based classification, enhancing generalization while maintaining interpretability.
Download

Paper Nr:	79
Title:	A Metric Hybrid Planning Approach to Solving Pandemic Planning Problems with Simple SIR Models
Authors:	Ari Gestetner and Buser Say
Abstract:	A pandemic is the spread of a disease across large regions, and can have devastating costs to the society in terms of health, economic and social. As such, the study of effective pandemic mitigation strategies can yield significant positive impact on the society. A pandemic can be approximately described using a compartmental model, such as the Susceptible–Infected–Removed (SIR) model. In this paper, we extend the solution equations of a simple SIR model to a state transition model with lockdowns. We formalize a metric hybrid planning problem based on this state transition model, and solve it using a metric hybrid planner. We improve the runtime effectiveness of the metric hybrid planner with the addition of valid inequalities, and demonstrate the success of our approach both theoretically and experimentally under various challenging settings.
Download

Paper Nr:	81
Title:	A Comparative Benchmark of Machine Learning Models for Static Malware Analysis: From EMBER 2018 to the Challenges of 2024
Authors:	David-Cristian Horvath and Imre Zsigmond
Abstract:	The growth of malware poses a persistent threat to digital infrastructure, requiring the development of efficient detection systems. While machine learning (ML) is key to modern static analysis, the practical trade-offs between models are not always well-documented. This paper addresses this gap by presenting a benchmark of six ML models on the EMBER 2018 dataset, including three Gradient Boosted Decision Tree (GBDT) implementations (LightGBM, XGBoost, CatBoost), a Random Forest, an Extra Trees classifier, and a Multilayer Perceptron (MLP). We extend this analysis by performing hyperparameter optimization, revealing performance improvements. Our evaluation moves beyond standard classification metrics to include efficiency criteria: inference time, training time, and model size. To investigate the impact of concept drift, we conduct a comparative analysis using the newly released EMBER2024 dataset. By training models on its modern v3 feature set and evaluating them on both the standard test set and the specialized ”challenge set” of initially-undetected malware, we quantify the increased difficulty of the detection landscape. Our findings confirm the state-of-the-art performance of GBDT models but also highlight the performance degradation against evasive threats, providing a guide for researchers on the balance between predictive accuracy, computational cost, and model transparency in malware detection scenarios.
Download

Paper Nr:	82
Title:	Adaptive Retriever Weighting for Robust Retrieval-Augmented Generation
Authors:	Sreenivasan Mohandas and Jahnavi Chowdary Pothuri
Abstract:	Retrieval-Augmented Generation (RAG) enhances Large Language models (LLMs) by grounding their responses in external knowledge, thereby reducing hallucinations and improving factual accuracy. However, the effectiveness of RAG critically depends on the choice and combination of retrievers, which often exhibit complementary strengths across domains. Existing approaches often rely on static or heuristic fusion of multiple retrievers, limiting adaptability across domains and query types. To address this, we propose Adaptive Retriever Weighting (ARW)--a lightweight, optimizer-inspired framework that dynamically adjusts the contribution of each retriever based on retrieval feedback. ARW maintains moving averages of retriever reliability, analogous to the momentum and variance updates in adaptive optimizers such as Adam. We instantiate ARW with four diverse retrievers-BM25, Fusion Retriever (FR), Context-Aware Fusion Retriever (CAFR), and Memento-capturing sparse, dense, and memory-based retrieval paradigms. Extensive experiments on HotpotQA, Natural Questions, and MS MARCO show that ARW consistently outperforms each individual retriever (BM25 (lexical retriever), FR, CAFR, Memento) as well as adaptive baselines such as Dynamic Alpha Tuning (DAT), demonstrating that optimizer-inspired adaptive weighting provides a simple and effective approach to building robust, domain-general RAG systems.
Download

Paper Nr:	83
Title:	Explainable AI-Assisted Framework for Fast and Reliable Fault Isolation in 5G Mobile Networks
Authors:	Quang-Vinh Tran, Tuan-Bach Phan, Duc-Thinh Vu and Quang-Diep Pham
Abstract:	Viettel’s 5G networks imposes stringent requirements on both reliability and Quality of Service (QoS), which makes timely fault detection and localization a crucial challenge. To address this, we leverage operational log data collected from the network and develop a lightweight anomaly detection mechanism augmented with explainable AI (XAI) for rapid fault localization. In this framework, anomalies are first detected using unsupervised models, after which core network entities are systematically examined as potential root causes through interpretable scoring mechanisms. By combining machine learning with explainability, the proposed solution enables automated decision-making in real time, reduces reliance on manual troubleshooting, and provides actionable insights for network operators. Preliminary experiments conducted on large-scale 5G log data demonstrate encouraging results, showing that the framework not only achieves accurate and early fault isolation but also remains computationally efficient for deployment in production environments. These findings highlight the potential of AI-driven, self-healing management mechanisms to enhance service continuity and support adaptive, intelligent operations in next-generation 5G infrastructures.
Download

Paper Nr:	85
Title:	Modeling Medical Descriptive Norm Convergence and Violation via Nonlocal Transport and Agent Learning
Authors:	Chao Li, Olga Petruchik and Sergey Kovalchuk
Abstract:	This paper proposes a unified model to study the dynamics of convergence toward and violation of descriptive medical norms in distributed healthcare environments. The model innovatively integrates a partial differential equation (PDE)-based diffusion-migration opinion dynamics framework with an agent-based modeling approach grounded in subjective-objective Gaussian mixtures (SINP-OBJ). At its core, it dynamically adjusts the parameters (σ, µ) of the perception kernel function using the Wasserstein distance between SINP and OBJ. Crucially, the sign of µ (positive/negative) mathematically determines whether the system trends toward norm convergence (group aggregation) or triggers norm violation (deviation from the mainstream). Through mathematical validation, agent-based extended simulations, and experiments on a real vertigo clinical dataset, the model successfully verifies its mathematical rigor and demonstrates that both norm convergence and violation phenomena can be uniformly characterized and transitioned within a single framework. This provides a new perspective for understanding the propagation and evolution of norms in complex medical systems.
Download

Paper Nr:	93
Title:	On the Capacity of Deep Autoencoder-Based Normal Behaviour Models in Wind Turbine Condition Monitoring
Authors:	Hoang Tu Bui, Juan F. H. Albarracín, Kyro Keown and Conor Ryan
Abstract:	This study compares Deep Autencoder (AE)-based Normal Behaviour Models (NBM) for anomaly detection in Wind Turbine SCADA data. Using a proprietary industrial dataset, we evaluate performance under real-world challenges, like class imbalance and unseen anomalies. We conduct a systematic comparison across unsupervised, semi-supervised, and supervised approaches, and examine the impact of auxiliary loss functions. Our results show that unsupervised Vanilla AEs struggle to separate normal and abnormal data, different from what the NBM literature claims about the effectiveness of unsupervised setups, as we obtain a significant increase in the Area Under the Curve (AUC) via its classification head. We propose an Adversarial Robust AE (ARAE) to improve detection in data-scarce scenarios. In settings with limited abnormal data, ARAE maintains stable recall at a 1:4 abnormal-to-normal ratio, outperforming other models under severe class imbalance. Based on these results, we recommend a Bottleneck architecture for scenarios with abundant labelled data and ARAE for those with scarce abnormal examples.
Download

Paper Nr:	95
Title:	SACn: Soft Actor-Critic with n-step Returns
Authors:	Jakub Łyskawa, Jakub Lewandowski and Paweł Wawrzyński
Abstract:	Soft Actor-Critic (SAC) is widely used in practical applications and is now one of the most relevant offpolicy online model-free reinforcement learning (RL) methods. The technique of n-step returns is known to increase the convergence speed of RL algorithms compared to their 1-step returns-based versions. However, SAC is notoriously difficult to combine with n-step returns, since their usual combination introduces bias in off-policy algorithms due to the changes in action distribution. While this problem is solved by importance sampling, a method for estimating expected values of one distribution using samples from another distribution, importance sampling may result in numerical instability. In this work, we combine SAC with n-step returns in a way that overcomes this issue. We present an approach to applying numerically stable importance sampling with simplified hyperparameter selection. Furthermore, we analyze the entropy estimation approach of Soft Actor-Critic in the context of the n-step maximum entropy framework and formulate the τ-sampled entropy estimation to reduce the variance of the learning target. Finally, we formulate the Soft Actor-Critic with n-step returns (SACn) algorithm that we experimentally verify on MuJoCo simulated environments.
Download

Paper Nr:	105
Title:	Exploring Customization Approaches of Large Language Models for Effective Causal Reasoning in Occupational Incident Texts
Authors:	Manato Nakamura, Satoru Hayamizu, Masanori Hattori, Takafumi Fuseya, Hidetoshi Iwamatsu, Takahiro Tajiri, Seiya Sano and Kazunori Terada
Abstract:	Recent advances in large language models such as GPT-4 have enabled highly accurate multi-label classification. In the occupational incident text domain, their performance has even been reported to approach human-level performance. However, prior work still faces three obstacles to operational deployment: (i) ambiguous label ontologies, (ii) lack of systematic comparisons of LLM customization methods, and (iii) absence of evaluations for advanced adaptation techniques such as retrieval-augmented few-shot prompting and finetuning. In this study, we performed multi-label annotation of incident causes for 442 occupational incident reports from an electric power company, using an expert-refined set of 21 labels and three models—GPT-4.1, o3, and o4-mini. To comprehensively evaluate the impact of different adaptation strategies, we combined prompt engineering, retrieval-augmented 3- and 10-shot prompting, and supervised fine-tuning, yielding 24 experimental conditions. Experimental results showed that GPT-4.1, augmented with prompt engineering, retrieval-augmented 10-shot prompting, and fine-tuning, achieved an F1 score of 65.0%, compared with a nonLLM BERT baseline (F1 = 51.8%). Notably, the highest-F1 configurations consistently employed few-shot prompting, supervised fine-tuning, or both. These findings demonstrate that incorporating expert reasoning patterns into an LLM is a particularly effective strategy for improving multi-label classification accuracy in domain-specific texts
Download

Paper Nr:	112
Title:	A Lightweight Model for Accurate Multi-View Hand Pose Recognition
Authors:	Artur Kadyrzhanov, Sergio Esteban-Romero, Manuel Gil-Martín and Marco Raoul Marini
Abstract:	This paper introduces a lightweight architecture for multi-view hand pose recognition on multimodal fusion of images and landmarks. The proposed model employs a compact Convolutional Neural Network (CNN) to extract visual features from dual-view grayscale images, while a Multi-Layer Perceptron (MLP) processes the corresponding Leap Motion Controller 2 hand landmarks. The two modalities are fused to create an efficient yet discriminative representation. Compared to the Vision Transformer (ViT)+MLP baseline, which achieves an F1 score of 79.33±0.09 % with 8.95×107 parameters, our CNN+MLP model reaches a higher recognition accuracy of 85.36 ± 0.08 % while requiring only 2.13 × 105 parameters, which corresponds to an important reduction of the model size. Moreover, a landmarks-only variant using the MLP achieves 85.22 ± 0.08 % accuracy with just 6.46×104 parameters. These results, obtained on the Multi-view Leap2 Hand Pose Dataset under a Leave-One-Subject-Out Cross-Validation protocol, demonstrate that accurate multi-view hand pose recognition can be achieved with dramatically fewer parameters, enabling efficient deployment in resourceconstrained environments.
Download

Paper Nr:	116
Title:	Diffusion Models for Constrained Planning with Probabilistic Risk-Awareness Guarantees
Authors:	Albin Larsson Forsberg, Kenneth Lau, Alexandros Nikou, Aneta Vulgarakis Feljan and Jana Tumova
Abstract:	Diffusion models have shown great potential in generating trajectory plans for agents in environments with unknown dynamics. However, such models provide no safety guarantees. In this work, we focus on risk-aware planning with respect to safety constraints and introduce a probabilistically risk-aware variant of Diffuser (PRA-Diffuser). The diffusion model initially learns a distribution over trajectories that may or may not be unsafe. We then fine-tune this model to reduce the probability of sampling such unsafe trajectories. We analyze the proposed solution and introduce a provable lower bound on risk of safety violation leveraging concentration inequalities for conditional Value-at-Risk. Our approach can be applied to models that have been pre-trained, potentially from datasets containing unsafe trajectories. Our empirical results demonstrate that our approach significantly reduces unsafe trajectories generated by the diffusion model across multiple environments.
Download

Paper Nr:	118
Title:	Context-Aware Autoencoders for Anomaly Detection in Maritime Surveillance
Authors:	Divya Acharya, Pierre Bernabé, Antoine Chevrot, Helge Spieker, Arnaud Gotlieb and Bruno Legeard
Abstract:	The detection of anomalies is crucial to ensuring the safety and security of maritime vessel traffic surveillance. Although autoencoders are popular for anomaly detection, their effectiveness in identifying collective and contextual anomalies is limited, especially in the maritime domain, where anomalies depend on vessel-specific contexts derived from self-reported AIS messages. To address these limitations, we propose a novel solution: the context-aware autoencoder. By integrating context-specific thresholds, our method improves detection accuracy and reduces computational cost. We compare four context-aware autoencoder variants and a conventional autoencoder using a case study focused on fishing status anomalies in maritime surveillance. Results demonstrate the significant impact of context on reconstruction loss and anomaly detection. The context-aware autoencoder outperforms others in detecting anomalies in time series data. By incorporating context-specific thresholds and recognizing the importance of context in anomaly detection, our approach offers a promising solution to improve accuracy in maritime vessel traffic surveillance systems.
Download

Paper Nr:	119
Title:	Image Quality Recovery from Bicubic Upscaling on Symmetric Scales via Deep Neural Networks
Authors:	Nhat Tan Vo, Duc Bao Ngo and Tien Len Bui
Abstract:	Image super-resolution (SR) is a fundamental task in computer vision that seeks to reconstruct high-resolution (HR) images from low-resolution (LR) inputs. Conventional methods such as bicubic interpolation are widely used for upscaling but typically produce blurry results with noticeable artifacts, particularly at large scaling factors. Although recent deep learning approaches have achieved impressive results in symmetric multiscale SR, most assume direct LR-to-HR mapping and are not optimized for inputs that have already been upscaled by bicubic interpolation. In this work, we address the challenging problem of quality recovery from bicubic-upscaled images at symmetric scales. We propose SPADEUNET-SR, a novel U-Net–based architecture integrated with SPADE modulation and refined through an EDSR-inspired residual module. The model effectively suppresses interpolation artifacts, restores high-frequency textures, and improves perceptual sharpness. Extensive experiments on standard benchmarks demonstrate that our method achieves competitive or superior results across multiple integer and fractional scales, striking a balance between fidelity and perceptual realism, and offering a robust solution for practical image enhancement scenarios.
Download

Paper Nr:	122
Title:	Backrooms-Llama: A Specific Domain Small Language Model for Computational Storytelling
Authors:	Roxanne Silva Julia, Rita Maria Silva Julia, Marcelo Zanchetta do Nascimento and Elaine Ribeiro de Faria
Abstract:	The popularization of Large Language Models (LLMs) faces a significant limiting challenge: the high cost associated with their use, whether through expensive hardware, cloud processing or paid APIs. In this light, the main objective of this work is to propose an approach to generate smaller language models, which can be trained in more modest hardware and that are effective in generating texts related to specific domains. The case study chosen for that are The Backrooms, an online legend which describes an alternative reality containing different levels with very unique and rich descriptions. For this, a Backrooms dataset was created and used to produce Backrooms-Llama, a small language model (SLM) conceived to generate Backrooms level descriptions from the fine-tuning of Llama 1B. Backrooms-Llama’s performance was then evaluated against the famous and much larger DeepSeek-V3 by using chatGPT 4o-mini as judge and a group of human evaluators. Additionally, a brand new evaluation method, which evaluates generated pictures of the output of each competitor model was proposed. The results found are promising, showing that, even though DeepSeek performed better, Backrooms-Llama, operating in a much more modest architecture, also got mainly relevant evaluations and was able to learn how to generate Backrooms level descriptions.
Download

Paper Nr:	123
Title:	ADEP-DQ: A Framework for Continuous Data Quality Assessment in MOOC Deep Learning Pipelines
Authors:	Thu Nguyen, Hong-Tri Nguyen and Tu-Anh Nguyen-Hoang
Abstract:	Massive open online courses (MOOCs) broaden access to education but still face high dropout rates and limited individualized support, motivating learning analytics that can anticipate performance and enable timely interventions. However, the effectiveness of deep-learning (DL) systems is strongly constrained by input quality: MOOC data are heterogeneous, dynamic, and frequently incomplete, inconsistent, noisy, or stale. While DL is well-suited to modeling the temporal complexity of MOOC signals (e.g., sequential logs, multimodal content, evolving learner behavior), its predictive utility remains highly sensitive to data quality (DQ). In this work, we formulate these learning-analytics tasks as multi-class classification (early prediction of course quality and early prediction of learning outcomes). Yet most existing DQ practices depend on manual review or ad hoc diagnostics that do not scale. We introduce ADEP-DQ (Automated Data Evaluation Pipeline for Data Quality), a scalable framework for systematic DQ assessment and monitoring that operates directly on the DataFrame/Tensor representations used in DL pipelines. ADEP-DQ automatically measures four core DQ dimensions (completeness, consistency, accuracy, and timeliness) and supports continuous monitoring with reactive actions (e.g., alerting, resampling, selective retraining) to maintain the reliability of the data–model value chain. Its key innovation is a performance-oriented view that links DQ to predictive utility via Acc-DQ (Accuracy-Driven Data Quality), a proxy for input reliability inferred from model feedback and statistical learning, reducing dependence on fixed reference datasets. By integrating automated, multidimensional measurement with direct performance relevance, ADEP-DQ enables DQ-aware deep learning and continuous DQ feedback in education, improving robustness, reducing manual tuning, mitigating unnecessary data loss, and accelerating MLOps through continuous DQ governance.
Download

Paper Nr:	124
Title:	Revolutionizing Deepfake Video Detection: A Quantum-Inspired Multimodal Approach
Authors:	Atul Pandey, S. N. Bhakti, Bhawana Rudra and Rajesh Kumar Krishnan
Abstract:	The rapid advancement of deepfake generation techniques across audio and video modalities has created significant challenges for reliable detection. Conventional deep learning approaches achieve strong performance but are limited by high computational demands due to large parameter counts and model sizes. To overcome these limitations, this paper proposes a hybrid classical–quantum multimodal framework for deepfake video detection. The framework employs ResNet50 and WavLM as pretrained classical feature extractors for video and audio modalities, respectively, while a Quantum Long Short-Term Memory (QLSTM) network is used for temporal sequence modeling. Experimental evaluations demonstrate that the proposed approach achieves performance comparable to its classical counterpart while drastically reducing trainable parameters in the temporal module, from (1,083,392) in the LSTM baseline to only (10,792) in the QLSTM, reflecting a 99% reduction. In the overall model parameter, the hybrid quantum model contains 19.7M parameters with a trained size of 499.32 MB, compared to 120.8M parameters and 503.61 MB in the equivalent classical model. These results demonstrate that the proposed hybrid quantum framework reduces the trainable parameter without compromising detection performance, underscoring the promise of quantum-enhanced architectures for scalable and resource-efficient multimodal deepfake detection.
Download

Paper Nr:	126
Title:	A Multi-Agent LLM Framework with RAG for Shipbuilding Engineering Documents
Authors:	Boris Escalante Galvan, George Drakoulas, Bram Cals, Tayfun Kucukyilmaz, Matthijs Schakel, Dina Semchenko, Juri Kuzjatkin and Gil Pina Cabral
Abstract:	In many industries, compliance relies on interpreting complex rulebooks that contain dense prose, technical tables, and engineering equations. In the maritime sector, these standards are considered critical to ensure that vessels meet regulatory standards. Yet, using manual labor for information extraction is time-consuming and inefficient. The development of Large Language Models and Retrieval Augmented Generation (RAG) systems offers a promising solution for the automation of such tasks. However, naive RAG systems often fail to combine varied data cohesively. To address that, this study proposes a novel multi-agent RAG architecture for maritime engineering standards, decomposing the complex processes and delegating sub-tasks to specialized agents. The system employs a five-agent pipeline: i) chapter-suggester, that suggests relevant chapters from manuals ii) main retriever, which is responsible for retrieving large pieces of information through a vector database, iii) prompt-generator, which generates prompts for missing information, iv) organizer, which organizes the retrieved information, and v) response-giver, which generates a final response. Results showed that different LLMs demonstrated varying performance in extracting engineering equations from shipbuilding manuals, with some achieving strong accuracy. For engineering computations, the multi-agent system showed moderate accuracy on simpler tasks but substantial errors on complex calculations, indicating the need for validation and hybrid approaches.
Download

Paper Nr:	127
Title:	FindMe Reforged: Temporal Logic AI for Richer Videogame Scenarios
Authors:	Vadim Malvone, Aniello Murano, Vincenzo Pio Palma and Salvatore Romano
Abstract:	This work presents an enhanced version of FindMe, a framework for adaptive and formally verified decision-making in Non-Player Characters (NPCs) within Unreal Engine 5.4. Using Computation Tree Logic (CTL), the system not only verifies temporal properties but also synthesizes action policies in real time, addressing a key limitation of traditional model checkers restricted to validation. To improve scalability and responsiveness, heuristic search through A* is combined with systematic exploration via Breadth-First Search (BFS), balancing efficiency with formal guarantees. By reducing computational overhead while preserving strategic depth, the tool enables NPCs to adapt to unpredictable player behavior and complex scenarios. This contribution shifts the focus from static verification to real-time policy synthesis, offering a practical approach for embedding goal-driven and formally grounded AI in modern games.
Download

Paper Nr:	134
Title:	GB-KAN: Gradient Boosting with Interpretable Kolmogorov-Arnold Networks
Authors:	Janis Mohr and Jörg Frochte
Abstract:	Gradient boosting remains a dominant approach for tabular data, but its widespread reliance on decision trees limits interpretability and the smoothness of learned functions. We propose Gradient-Boosted Kolmogorov-Arnold Networks (GB-KANs), a boosting framework that replaces trees with shallow Kolmogorov-Arnold networks as base learners. By fitting KAN shape functions to pseudo-residuals in an additive, stagewise manner, GB-KANs inherit the predictive strength of boosting while yielding per-feature functions that are easy to inspect. Across real-world datasets, GB-KANs achieve accuracy competitive with established baselines and produce shape functions that are faithful, sparse, and stable. They also yield well-calibrated probabilities without post-hoc correction. These properties position GB-KANs as a promising approach when interpretability is essential.
Download

Paper Nr:	137
Title:	Investigating the Lottery Ticket Hypothesis for Variational Quantum Circuits
Authors:	Michael Kölle, Leonhard Klingert, Julian Schönberger, Philipp Altmann, Tobias Rohe and Claudia Linnhoff-Popien
Abstract:	Quantum computing is an emerging field in computer science that has seen considerable progress in recent years, especially in machine learning. By harnessing the principles of quantum physics, it can surpass the limitations of classical algorithms. However, variational quantum circuits (VQCs), which rely on adjustable parameters, often face the barren plateau phenomenon, hindering optimization. The Lottery Ticket Hypothesis (LTH) is a recent concept in classical machine learning (ML) that has led to notable improvements in parameter efficiency for neural networks. It states that within a large network, a smaller, more efficient subnetwork, or “winning ticket,” can achieve comparable performance. This work investigates whether applying LTH-inspired pruning to VQCs can serve as a strategy to mitigate optimization challenges like barren plateaus. Our experiments on the Iris and Wine datasets show that while the weak LTH holds for VQCs-identifying sparse, well-performing subnetworks-it does not consistently lead to improved trainability. However, we uncover a notable exception where pruning an over-parameterized VQC significantly boosts accuracy, suggesting that for some VQCs, sparsity can act as a powerful regularizer. For the strong LTH, we find a winning ticket in a binary VQC that achieves high accuracy with only 45% of its weights, without any training. These findings indicate that while LTH is not a universal solution for barren plateaus, it is a valuable tool for revealing the complex interplay between parameterization, optimization, and performance in VQCs.
Download

Paper Nr:	144
Title:	Developing Intelligent Chatbots for Telecom Security Support: A Comparative Study of Large Language Model Utilization Strategies
Authors:	Yuejun Guo, Qiang Tang, Hoang Trang and Cu D. Nguyen
Abstract:	Large language models (LLMs) have significantly advanced natural language processing across multiple domains with their remarkable capabilities in processing, understanding, and generating human-like text at scale. Different strategies have emerged for adapting these models to specialized domains, such as fine-tuning on domain-specific data and augmenting base models with external knowledge retrieval systems. In this paper, we focus on the telecom security domain and compare three strategies for leveraging LLMs: using unmodified base LLMs with prompt engineering, fine-tuning LLMs on domain-specific data, and enhancing base LLMs with Retrieval-Augmented Generation (RAG). Our experiments demonstrate that while fine-tuned LLMs (slightly) improves performance than base models, the RAG strategy outperforms fine-tuned LLMs while requiring significantly lower computational resources and offering greater flexibility. These findings suggest that for practical telecom security applications, RAG represents a more efficient and effective strategy than domain-specific fine-tuning.
Download

Paper Nr:	145
Title:	Geometric 2D Scene Graph Generation
Authors:	Christoph Jahn, Urs Waldmann and Bastian Goldlücke
Abstract:	In production processes for consumer products, assembly instructions are essential not only for planning but also for executing the production process. Likewise in robotics, it is crucial for an assembly robot to understand how components fit together and can be assembled. To facilitate these tasks, we contribute a method for constructing scene graphs to represent and characterize assembly relationships between components. Our approach does not rely on semantic data and is capable of handling a very small dataset. To realize this, the output of a Faster R-CNN model is used to create geometric representations, which are then processed by a transformer architecture to generate an adjacency matrix. This matrix serves as input to a Siamese network that uses message passing based on an attentional graph convolutional network (aGCN) architecture to characterize the connections between the components. We validate our method on a study dataset of toy model components which can be assembled into transportation vehicles.
Download

Paper Nr:	146
Title:	Alignment Challenges in Arabic Multimodal Sentiment Analysis
Authors:	Ayda Boukhalda, Hasna Chouikhi and Fethi Jarray
Abstract:	Aligning heterogeneous modalities is a critical step in multimodal sentiment analysis (MuSA), as it enables the effective integration of textual and visual information. While alignment techniques have been extensively studied in English and other widely used languages, Arabic MSA remains relatively underexplored. In this study, we conduct a comparative analysis of three prominent alignment methods-Co-discriminative learning (CODIS), Wasserstein Modality Alignment (WMA), and Mixture of Experts (Alt-MoE)-across three diverse Arabic multimodal datasets. The results indicate that WMA consistently achieves the highest performance, reaching 80.14% accuracy and 79.42% F1 score on the Scraped Arabic dataset, benefiting from robust semantic alignment, particularly on datasets with high variability. CODIS demonstrates strong performance on datasets with consistent multimodal patterns but is more sensitive to noise, while Alt-MoE shows adaptability to specific datasets but exhibits variability due to over-specialisation.
Download

Paper Nr:	156
Title:	Resource-Aware Handwritten Text Recognition: A Compact CNN-Augmentation Pipeline for Sustainable Edge Deployment
Authors:	Omar Haddad, Mohamed Nazih Omri and Montaha Nour Mchiri
Abstract:	Deploying handwritten text recognition (HTR) on edge devices is difficult due to writer variability, cursive ligatures, and scarce labeled data in real deployments. Although Transformer-based recognizers achieve strong accuracy, their computational footprint (tens of millions of parameters and high latency) often prevents low-power, privacy-preserving use cases. We present a reproducible, resource-aware pipeline that couples handwriting-oriented pre-processing with on-the-fly augmentation and a compact CNN backbone (1.3M parameters). Across EMNIST-Letters, IAM words, and a proprietary 3.4k-sample corpus, the proposed model improves accuracy by +6.8 points over a vanilla CNN and halves the character error rate (14.8% → 7.4%), while maintaining low inference latency (4.2 ms/sample on GPU and 18.5 ms/sample on ARM CPU). For word recognition, a hybrid CNN→TrOCR configuration further reduces IAM CER from 3.0% to 2.7%. These results highlight a practical accuracy–efficiency trade-off for sustainable on-device HTR.

Paper Nr:	169
Title:	A Constraint Optimization Approach for Flexible and Preference-Aware Medical Appointment Scheduling
Authors:	George Assaf, Sven Löffler and Petra Hofstedt
Abstract:	In medical appointment scheduling, patient preferences, such as desired appointment dates, times, weekdays, or physicians, should be respected whenever possible to maximize patient satisfaction. Moreover, it is crucial to offer alternative appointment proposals, allowing both patients and physicians to participate in the decision-making process. To address these needs, we present a constraint programming model that integrates soft constraints into the fundamental constraints of the medical appointment scheduling problem (MASP) to capture patient preferences and provide alternative schedules. The proposed model demonstrates the capability to produce good-quality schedules that balance operational requirements with patient-centered preferences. By exploring different search strategies, we demonstrate that our model produces solutions within an acceptable time frame.
Download

Paper Nr:	175
Title:	A Novel Approach for Emotion Recognition Using Hybrid Appearance–Geometric Features with Explainable AI
Authors:	Rim Afdhal, Ridha Ejbali and Mourad Zaied
Abstract:	Facial Expression Recognition (FER) is a fundamental component of affective computing, enabling machines to interpret human emotions through visual cues. The growing success of deep learning approaches have enabled automatic FER; however, many systems still suffer from high computational costs and redundant feature representations. In this paper, we propose a novel emotion recognition system that addresses these limitations via a multi-stage approach. The proposed system decreases computational cost and discards ineffective features by selecting emotion-specific geometric and appearance features, then fusing them into a compact representation. This not only enhances efficiency but also improves classification performance. The proposed pipeline is structured into four fundamental stages: pre-processing, feature extraction, feature fusion, and classification. To further enhance model transparency, we incorporate Explainable AI (XAI) technique that highlight the most influential features contributing to each decision. Experimental results indicate that our approach enhances recognition accuracy while ensuring interpretability and computational efficiency.
Download

Paper Nr:	179
Title:	Optimizing Sequential Models through Temporal Landmark Selection and Normalization for Sign Language Recognition
Authors:	Sergio Esteban-Romero, Iván Martín-Fernández, Cristina Luna-Jiménez, Manuel Gil-Martín, Fernando Fernández-Martínez and Elisabeth André
Abstract:	This paper presents a comprehensive study of lightweight and efficient models for isolated sign language recognition that achieve consistent performance in two benchmark datasets, AVASAG100 and WLASL100. Our pipeline leverages ViTPose to extract keypoints from every video frame, which are then processed by Transformer-based architectures. We investigate the impact of keypoint normalization and centering across frames, as well as the role of a learnable weighted pooling mechanism compared to uniform averaging of encoder representations. Our findings show that normalization, particularly centering, is crucial for stable and effective learning. Furthermore, while learnable weighted pooling enhances performance in LSTM-based models, it provides limited benefit for Transformers, indicating that Transformer representations are inherently more robust. Moreover, an ablation study related to a simple yet effective frame selection strategy based on mean hand-confidence thresholds led to improved efficiency efficiency. On WLASL100, the proposed compact Transformer with only 2.2M parameters achieves strong results, 65.99% M-F1, 66.24% W-F1, and 67.83% Accuracy, demonstrating the effectiveness of our approach in balancing efficiency and performance.
Download

Paper Nr:	194
Title:	A Data-Driven Approach for Fibres Recognition via Spectrophotometry
Authors:	Megan Robinson, Saikat Ghosh, Chenyu Du, Parikshit Goswami and Mauro Vallati
Abstract:	The increasing volume of textile waste presents significant environmental and economic challenges, necessitating the development of efficient automated sorting techniques to support a more effective textile waste recycling. Automated sorting is a notoriously complex task, due to deployment constraints and to the variability of textiles. To advance the work on automated textile sorting, this study investigates the use of data-driven approaches on spectrophotometer-based reflectance measurements for recognising fibres. Spectrophotometry offers significant advantages in terms of operational simplicity and reliability, making it a promising choice for use in textile sorting facilities where environmental conditions are difficult to control. Considering an extensive dataset of specifically acquired pure textile samples, in this work we leverage on AutoML solutions to determine the best architecture to discriminate between cotton and polyethylene terephthalate (PET) fibres.
Download

Paper Nr:	200
Title:	Enhancing Data Privacy in Alzheimer's Research: Leveraging Gaussian Noise for Reliable Defense in Transformer Models
Authors:	Sameh Ben Hamida, Boussad Ait Salem, Faten Chakchouk Chaieb, Hichem Mrabet and Abderrazak Jemai
Abstract:	Ensuring patient privacy is a major challenge in adopting machine learning systems in healthcare. The sensitive nature of medical records requires strict safeguards, yet recent studies have shown that advanced models are vulnerable to attacks targeting confidentiality. Among these, Membership Inference Attacks (MIAs) represent a significant threat, as they exploit model predictions to infer whether specific patient data were included in the training set, potentially exposing private information. This issue is particularly critical in areas such as neurodegenerative disease research, where datasets are often limited and imbalanced. In this work, we propose a privacy-preserving framework for the prediction of Alzheimer's disease based on transformer architectures. To counteract MIAs, we integrate carefully calibrated Gaussian noise into the model during training. This strategy allows us to reduce the ability of adversaries to infer membership information while retaining strong predictive performance. Our results demonstrate that the proposed method reduces the AUC of MIA attacks by more than 20% compared to baseline models, without causing a clinically significant decline in diagnostic accuracy. These findings indicate that it is feasible to design deep learning systems that achieve both robust privacy protection and practical utility, offering a promising direction for secure medical AI applications.
Download

Paper Nr:	203
Title:	Deep Maritime Awareness through Sequential Modeling of AIS Trajectories
Authors:	Najeh Abdeladhim, Bechir Alaya and Haifa Touati
Abstract:	Maritime surveillance is crucial for navigation safety, vessel monitoring, and regulatory compliance. Detecting unusual vessel trajectories from Automated Identification System (AIS) data helps identify irregular behaviour and reduce risks in busy waterways. Analysing large-scale AIS datasets is challenging due to their volume and complexity. This study proposes an unsupervised deep learning framework for anomaly detection in vessel movements. The approach integrates a Long Short-Term Memory (LSTM) autoencoder with a One-Class Support Vector Machine (SVM). The LSTM autoencoder learns typical vessel patterns and generates a compact representation of the data. This latent representation is then used by the One-Class SVM to identify deviations from normal behaviour. The hybrid model overcomes limitations of individual methods in handling high-dimensional data. Experiments show that the approach improves detection accuracy and reliability. The method provides an effective, scalable solution for real-time maritime surveillance.
Download

Paper Nr:	205
Title:	Meta Prompting for Clinical AI: Iterative Refinement of GPT-5 in Psychomotor Assessment
Authors:	Sarah Ayad and Pamela Kattar
Abstract:	The LAby 5–12 maze is a clinical test used to assess children’s motor coordination and thinking skills. Two main indicators are measured: Line Crossing (LC), which shows how well the child controls their hand movements, and Deviation Measure (MD), which reflects planning mistakes like going into wrong paths. We tested four computer vision models to detect LC, including MobileNetV2 and a hybrid OpenCV+CNN approach. All maze images were preprocessed using the Segment Anything Model (SAM) to focus on the child's trace. MobileNetV2 gave the best results for LC, reaching 84\% accuracy. However, all vision models performed poorly on MD. To improve MD scoring, we introduced a five-step prompting method with GPT-5. The final version (V5) included examples and error rules, reaching 86% accuracy with clear, explainable outputs. This study shows that combining MobileNetV2 for LC and GPT-5 for MD can create an effective and interpretable system for automatic psychomotor assessment, even with a small dataset.
Download

Paper Nr:	209
Title:	Cooperative Intelligent Agents Based Approach for Proactive Defense and Geolocation of BGP Hijacking Attacks in SD-WAN Systems
Authors:	Hichem Dachraoui, Wafa Mefteh and Ali Frihida
Abstract:	Software-Defined Wide Area Networks (SD-WAN) have revolutionized enterprise networking with their adaptive and cost-effective connectivity solutions. However, their reliance on Border Gateway Protocol (BGP) introduces unprecedented security risks, particularly BGP hijacking attacks, which compromise network integrity and data confidentiality. This paper proposes a novel cooperative and intelligent agents based proactive distributed defense solution using machine learning based detection and real-time geo-location analysis for protecting SD-WAN infrastructures against BGP hijacking attacks. Our approach employs a multi-agent architecture with Random Forest classifiers achieving 99.3% accuracy in detecting hijacking attempts, complemented by a web application providing real-time monitoring and visualization of network topology and potential threats. Experimental evaluation demonstrates the system's capability to detect and mitigate BGP hijacks within minutes while maintaining low false positive rates of approximately 17.5 suspicious cases per day for internet-wide monitoring.
Download

Paper Nr:	210
Title:	Weighted Spatio-Temporal Graph Neural Network: A Novel Approach for Video Anomaly Detection
Authors:	Linda Zitouni, Nahla Majdoub Bhiri and Anouar Ben Khalifa
Abstract:	The ability of human skeletons to identify unusual activity in video surveillance has attracted considerable interest recently. In particular, skeleton-based Video Anomaly Detection (VAD) has extensively used Graph Neural Networks (GNNs). However, conventional models often suffer from limited generalizability and inefficient feature representation. To address these challenges, we propose the Weighted Spatio-Temporal Graph Convolutional Network (WST-GCN), which improves feature aggregation by integrating weighted ST-GCN layers. We begin by extracting human skeletons from video frames, after which the WST-GCN module assigns adaptive weights to skeleton joints and employs a multi-layer perceptron to learn representations of normal and abnormal behaviors. Compared to existing GNN models, our suggested approach enhances the AUC-ROC (Area Under the Curve -Receiver Operating Characteristic) of classification. We validate the effectiveness of WST-GCN on the UBnormal dataset, where it achieves an improved AUROC of 70.4%, surpassing current state-of-the-art models.
Download

Paper Nr:	213
Title:	Human Oversight in the AI Act: From Legal Obligation to Operational Safeguards for Fundamental Rights
Authors:	Silvia Tessaro Trapani, Nicole Inverardi, Ilaria Penco and Manuela Zaidan
Abstract:	Human oversight occupies a central role in the European Artificial Intelligence Act (Regulation (EU) 2024/1689), serving as the primary safeguard for protecting individuals’ fundamental rights in high-risk AI applications. This article offers a normative and interpretative analysis, emphasizing that oversight is not merely procedural, but a structural guarantee embedded both in system design (Article 14) and in operational practice through qualified human overseers (Article 26). Oversight transforms the abstract commitments of the AI Act into tangible protections, enabling trained and empowered individuals to monitor, evaluate, and, if necessary, override AI outputs. Its integration with the Fundamental Rights Impact Assessment (FRIA, Article 27) ensures that risk identification leads to concrete mitigation, bridging the gap between regulatory principles and real-world outcomes. The paper further elaborates on the three dimensions of effective oversight-authority, competence, and technical capacity-and presents a structured analytical checklist to assess compliance in organizational contexts. By highlighting the dual responsibilities of providers and deployers, the study demonstrates how human oversight operationalizes the AI Act’s human-centric and rights-based objectives, while also addressing practical challenges arising from fragmented responsibilities across the AI value chain.
Download

Paper Nr:	216
Title:	Experiences in Exploiting Reinforcement Learning for Network Traffic Classification and Attack Detection
Authors:	Salvo Finistrella, Stefano Mariani and Franco Zambonelli
Abstract:	Cybersecurity demands adaptive solutions against sophisticated evolving threats. This work presents an experimental study exploiting Reinforcement Learning (RL) for real-time network traffic classification and attack detection within a software-defined networking environment. We extensively compare five RL algorithms, including both tabular methods and deep learning ones. Agents are trained in a simulated Mininet network, where an adversarial traffic generator introduces heterogeneous traffic patterns. Observing per-second flow statistics, agents learn by receiving rewards based on the correctness of their classifications. Our experiments reveal that tabular methods often demonstrate rapid learning and stable convergence for traffic detection, while deep learning methods, despite exhibiting greater training variability, show strong generalization capabilities in attack detection. In addition, the experiments offer practical insights into algorithm selection and the impact of hyperparameters on learning outcomes.
Download

Paper Nr:	221
Title:	On Predicting Electoral Outcomes Using TikTok Comments
Authors:	Ștefania-Alexandra Cristea and Bogdan Ichim
Abstract:	This paper investigates the potential of using public comments from the TikTok platform in order to estimate electorate intentions and predict the final results of the elections. In particular we focus on the Romanian presidential elections of 2024 – 2025. More than 653,000 comments were collected from the official accounts of the main candidates, after which they were automatically annotated and partially validated manually along three dimensions: sentiment, emotion, and political context. Relevant political entities were extracted through Named Entity Recognition (NER). Several classification models, including Bidirectional Encoder Representations from Transformers (BERT), Multi-Task BERT and Random Forest, were trained in order to analyse opinion polarization and to compute popularity scores for the elections candidates. Our findings indicate that the ratio of positive to negative comments strongly correlates with the elections results, although the presence of bots may introduce some bias. The study highlights the usefulness of TikTok comments and Natural Language Processing methods as a complementary source for estimating public opinion and underlines the need for hybrid approaches that combine social media data with traditional polling methods.
Download

Paper Nr:	223
Title:	Experimental Evaluation of Multi-Agent Consensus Protocols Using Aerial Swarms
Authors:	Mikhael Sunil, Akshat Singh, Agniva Banerjee, Arijit Sen and Sujit P. B.
Abstract:	This work presents the experimental evaluation of multiple consensus protocols using Crazyflie drone swarm. Unlike prior studies limited to simulations, we implement and compare classic consensus, fractional-order consensus, double-integrator schemes, and formation control in real-world conditions. We analyze convergence speed, stability under disturbances, oscillatory behavior, and formation accuracy. The results reveal distinct trade-offs: fractional-order dynamics provide improved oscillation damping, double-integrator protocols enhance formation stability, and advanced formation control achieves precise dynamic coordination. By addressing noise, communication delays, and real-time execution, this study provides novel insights into adaptive and scalable swarm consensus, bridging theoretical models with practical deployment. All of the codes, rosbags, and videos for this paper are provided on GitHub link Video link.
Download

Paper Nr:	232
Title:	Evolutionary Transfer Learning for Dragonchess
Authors:	Jim O'Connor, Annika Hoag, Sarah Goyette and Gary B. Parker
Abstract:	Dragonchess, a three-dimensional chess variant introduced by Gary Gygax, presents unique strategic and computational challenges that make it an ideal environment for studying the transfer of artificial intelligence (AI) heuristics across domains. In this work, we introduce Dragonchess as a novel testbed for AI research and provide an open-source, Python-based game engine for community use. Our research investigates evolutionary transfer learning by adapting heuristic evaluation functions directly from Stockfish, a leading chess engine, and subsequently optimizing them using Covariance Matrix Adaptation Evolution Strategy (CMA-ES). Initial trials showed that direct heuristic transfers were inadequate due to Dragonchess’s distinct multi-layer structure and movement rules. However, evolutionary optimization significantly improved AI agent performance, resulting in superior gameplay demonstrated through empirical evaluation in a 50-round Swiss-style tournament. This research establishes the effectiveness of evolutionary methods in adapting heuristic knowledge to structurally complex, previously unexplored game domains.
Download

Paper Nr:	237
Title:	Sentiment-Driven Stock Price Prediction: Analysing Green Finance News Using Large Language Models for Tesla
Authors:	Amir Lorvand, Halil Yetgin, Serengul Smith, Duaa Alkubaisy and Luca Piras
Abstract:	The rise of green finance has had a significant impact on the way investors evaluate a company’s performance, prompting them to consider environmental, social and governance (ESG) criteria alongside traditional financial metrics. Nevertheless, research investigating the potential of ESG-related news to improve stock price prediction models is still limited. This study examines the effectiveness of sentiment analysis derived from green finance and ESG-related news to improve stock price forecasts specifically for Tesla Inc. Historical stock price data was combined with sentiment scores extracted from ESG-related news articles using FinBERT. Two recurrent neural network models (RNN) and two long short-term memory (LSTM) models were developed and evaluated, both with and without the inclusion of sentiment features. The results show that sentiment features have only a modest impact on predictions during stable market conditions, while they provide significant improvements during periods of market volatility. Specifically, during a volatile period when Tesla’s share price fluctuated between $340 and $480, the inclusion of sentiment features in the LSTM model reduced the root mean square error (RMSE) from 9.87 to 4.06. This significant reduction underscores the improved accuracy achieved by incorporating textual sentiment data. Overall, these results emphasise the importance and potential benefits of combining traditional financial indicators with textual data analysis, especially in times of increased market uncertainty.
Download

Paper Nr:	238
Title:	Content-Based Arabic Documents Recommendation for Smart Learning in Campus
Authors:	Ons Meddeb, Mohsen Maraoui and Mounir Zrigui
Abstract:	The rapid growth of smart learning environments, driven by Internet of Things (IoT) technologies, has intensified the need for adaptive and personalized educational content delivery. However, existing recommender systems suffer from cold-start limitations and delayed adaptation, challenges that are further amplified in Arabic educational contexts due to linguistic complexity and data sparsity. This paper proposes RT-DeepARS, a Real-Time Deep Arabic Recommendation System designed to provide continuously adaptive content-based recommendations in smart learning environments. The proposed framework integrates semantic representation learning from Arabic document metadata with contextual modeling and incremental learning, enabling real-time adaptation to evolving user preferences and newly available learning resources. A data-driven architecture enables continuous updates and low-latency recommendations. Experiments conducted on an enriched version of the BRAD 1.0 dataset demonstrate that RT-DeepARS outperforms static baselines, while reducing training latency. These results confirm that combining deep semantic modeling with incremental learning significantly improves both accuracy and responsiveness. Overall, RT-DeepARS offers a scalable and effective solution for intelligent and adaptive Arabic recommendation in smart learning environments.

Paper Nr:	240
Title:	Generative Adversarial Networks for Malicious Social Bot Detection: Insights, Taxonomy, Datasets, and Future Directions
Authors:	Ghada Feki, Nèle Impouma, Nourhène Ben Rabah, Benedicte Le-Grand and Chokri Ben Amar
Abstract:	The escalating sophistication of malicious social bots (MSBs) poses a significant and evolving threat to social cybersecurity, challenging the integrity of online discourse and trust. While traditional machine learning (ML) approaches have been widely deployed for bot detection, they exhibit critical limitations, including a reliance on static features and an inability to generalize to novel, evolving bot strategies. This paper argues that Generative Adversarial Networks (GANs) present a powerful and versatile framework to overcome these challenges and advance the state of social cybersecurity. Our contributions are fourfold: First, we provide a systematic identification and analysis of the inherent limitations in traditional ML-based bot detection. Second, we develop a novel multidimensional taxonomy to categorize and elucidate the landscape of GAN architectures relevant to security tasks. Third, we present a comprehensive overview of the datasets employed in GAN-based bot detection research. Fourth, we conduct a critical evaluation of existing GAN-based approaches, assessing their efficacy in anticipating attacks and adapting to dynamic bot behavior. We outline promising future research directions and the next wave of challenges, emphasizing that the path to resilient, next-generation social cybersecurity systems lies in addressing issues of scalability, adaptability, and model interpretability for complex, multi-platform environments.
Download

Paper Nr:	247
Title:	KGiRAG: An Iterative GraphRAG Approach for Responding Sensemaking Queries
Authors:	Isabela Iacob, Melisa Marian and Gheorghe Cosmin Silaghi
Abstract:	Recent literature highlights the potential of graph-based approaches within large language model (LLM) retrieval-augmented generation (RAG) pipelines for answering queries of varying complexity, particularly those that fall outside the LLM’s prior knowledge. However, LLMs are prone to hallucination and often face technical limitations in handling contexts large enough to ground complex queries effectively. To address these challenges, we propose a novel iterative, feedback-driven GraphRAG architecture that leverages response quality assessment to iteratively refine outputs until a sound, well-grounded response is produced. Evaluating our approach with queries from the HotPotQA dataset, we demonstrate that this iterative RAG strategy yields responses with higher semantic quality and improved relevance compared to a single-shot baseline.
Download

Paper Nr:	248
Title:	Optimizing Vision Transformers for Brain Tumor Diagnosis from MRI Images
Authors:	Khaled M. Hassen Wane, Sawsen Guendouz, Nessrine Touahar, Khaled Bensid, Mouhamed Laid Abimouloud and Monji Kherallah
Abstract:	This study investigates the application of optimization algorithms in Vision Transformers (ViTs), a deep learning architecture based on self-attention mechanisms, for automated multi-class classification of brain tumors from Magnetic Resonance Imaging (MRI) scans. ViTs analyze MRI images as sequences of patches, enabling them to capture complex global and spatial relationships that are often critical for accurate tumor characterization. To address the high computational cost and extended training time typically associated with ViTs, we implemented hyperparameter tuning and compared optimization algorithms Adam, Adamax, and SGD, to enhance model efficiency and performance. The optimized ViT model, trained on 7,023 MRI images across four classes glioma, meningioma, pituitary tumor, and no tumor achieved an accuracy of 95.8%, with precision and sensitivity exceeding 95%, and an AUC of 99.15%. Compared to existing ViT-based studies, our optimized model achieves comparable diagnostic accuracy with substantially reduced training time and computational requirements, making it more practical for real-world clinical integration. These results indicate that ViT-based models can provide radiologists with a reliable, automated second opinion for brain tumor diagnosis, potentially reducing diagnostic variability, improving early detection, and expediting clinical decision-making. This work highlights the promise of ViT-driven approaches as clinically valuable tools in computer-aided neuro-oncology diagnosis.

Paper Nr:	255
Title:	Artificial Neural Networks Algorithm for Bioconvection Flow Considering Magnetic Potential
Authors:	Merve Gurbuz-Caldag, Bengisen Pekmen and Hakan F. Oztop
Abstract:	The present study develops an artificial neural network (ANN) model to predict bioconvective flow generated by magnetotactic bacteria in a square cavity with a rounded upper corner under an external magnetic field. The induced magnetic field is directly incorporated into the governing equations of bioconvection. A dataset is constructed using radial basis function (RBF) method by varying key physical parameters, including the Rayleigh, bioconvective Rayleigh, Peclet, Lewis, Hartmann and magnetic Reynolds numbers, and radius of the rounded corner. The ANN is trained and tested using multiple architectures, activation functions, and partition ratios to evaluate performance. Results indicate that a trilayer ANN with ReLU activation and an 80:20 training-to-testing split achieves the lowest mean squared error across all target outputs. Unlike conventional numerical solvers, the proposed ANN acts as a fast model capable of accurately predicting heat, mass, and bioconvective transport indicators across a high-dimensional parameter space.
Download

Paper Nr:	259
Title:	Real Time Exit Blockage Detection via Door-Centric Change Monitoring for Emergency Evacuation
Authors:	Hamed Talebi, Nahid Khoshk Angabini, Arezoo Sarkheyli-Hägele and Erdal Akin
Abstract:	The availability of unobstructed exits is crucial for ensuring safe and efficient evacuation during emergencies. However, real-time exit blockage detection, especially in buildings with multiple exits, poses a critical challenge, and dynamic unforeseen obstructions can cause confusion, delays, and even injuries as evacuees may be directed toward inaccessible routes. Human monitoring is often proven unreliable under such rapidly changing conditions. In addition, existing vision-based approaches rely on exhaustive obstacle catalogs and struggle to adapt to new and unforeseen blockage scenarios. In this work, we propose a novel, door-centric change detection model that bypasses the need for object-specific obstacle recognition. Instead of identifying all possible obstructions, our method focuses solely on monitoring the visual state of doors. Using a lightweight YOLOv11s model, we track the bounding box geometry of each door and detect blockages based on deviations from a known unobstructed reference. This approach simplifies the detection process while improving adaptability to unknown obstruction types. We evaluate our model on diverse and previously unseen video streams containing various types of door blockage. The results demonstrate high accuracy in estimating accessible door areas and show strong generalization capabilities beyond the limitations of traditional object detection methods. Notably, the model supports real-time inference on low-power edge devices, such as the Raspberry Pi, enabling practical deployment in resource-constrained environments.
Download

Paper Nr:	260
Title:	On the Prediction of the Pareto Distribution with a Nonstationary Shape Parameter Based on Bayes Decision Theory
Authors:	Daiki Koizumi
Abstract:	A prediction problem for the Pareto distribution with a nonstationary shape parameter is investigated within the framework of Bayes decision theory. The proposed nonstationary Pareto distribution includes a single hyperparameter used to express the nonstationarity of its shape parameter. Furthermore, the proposed predictive algorithm utilizing the observed time series data is based on both the posterior distribution of the shape parameter and the predictive distribution of data, all within a Bayesian context. Each predictive estimator derived from the proposed algorithm achieves Bayes optimality, which guarantees the minimum expected error under the proposed nonstationary Pareto distribution, a prior distribution of the shape parameter, and an absolute-error loss function, according to Bayes decision theory. Moreover, an approximate maximum likelihood estimation method for the hyperparameter based on numerical calculation is also presented. Finally, the predictive performance of the proposed algorithm has been evaluated in terms of both model selection theory and the predictive mean absolute error through comparisons with the stationary Pareto distribution using real financial time series data.
Download

Paper Nr:	261
Title:	Revisiting Representer Point Selection for Interpretable Predictions
Authors:	Zalán Bodó and Ábel Portik
Abstract:	Explaining the predictions of complex machine learning models, particularly deep neural networks, has emerged as a central challenge in artificial intelligence. Despite progress through both intrinsic and post-hoc interpretability methods, no general solution ensures model transparency across architectures. Among various approaches, using the representer theorem to trace model decisions back to training examples offers an appealing theoretical framework. However, as we demonstrate in this work, practical limitations – such as the need for strictly convex loss functions and sparsity in expansion coefficients – restrict its applicability. This paper investigates these constraints and proposes deep support vector machines for explaining the decisions of a neural network. We also extend the saliency map-based visualization using a decomposed Grad-CAM approach and find the optimal layer along existing dataset annotations applying a layer diversity score. The work builds on and critiques existing methodologies, presenting a structured evaluation and highlighting directions for future research. The code of the experiments is available at https://github.com/miafranc/rrepps.
Download

Paper Nr:	266
Title:	Evaluating Translation Models and LLMs in Poetic Texts
Authors:	Beatriz Ribeiro Borges, Paulo Henrique Ribeiro Gabriel and Elaine Ribeiro de Faria
Abstract:	This study investigates the performance of specialized machine translation (MT) models and large language models (LLMs) in the automatic translation of poetry across six language pairs (FR-EN, FR-PT, EN-FR, EN-PT, PT-FR, PT-EN). Automatic evaluations using BLEU, METEOR, and BERTScore were complemented by human assessments focusing on poetic structure, stylistics, fluency, meaning, and overall impression. Results indicate that LLMs, particularly ChatGPT overall and Maritaca AI for translations into Portuguese, outperform specialized MT models in semantic fidelity and fluency, although Google Translate also performed very well, surpassing other MT models such as MarianMT, mBART, and OpenNMT (RNN). Despite these successes, all models struggle with poetic form, rhyme, and figurative nuances.
Download

Paper Nr:	267
Title:	Fed-SHAP-IDS: Federated SHAP-Based Intrusion Detection System for IoMT
Authors:	Mohammed Yacoubi, Omar Moussaoui and Cyril Drocourt
Abstract:	The use of the Internet of Medical Things (IoMT) has significantly increased in recent years, introducing new vulnerabilities and security challenges. Numerous attacks now target the healthcare ecosystem, exploiting weaknesses in connected medical devices. Traditional intrusion detection methods are increasingly insufficient to cope with the growing number of attacks and the emergence of zero-day threats. Artificial Intelligence (AI) has demonstrated strong potential in this context; however, centralized AI-based detection solutions expose sensitive medical data, create a single point of failure, and struggle with non-IID (non–independent and identically distributed) data across geographically dispersed devices. To overcome these limitations, we propose Fed-SHAP-IDS (Federated SHAP-based Intrusion Detection System), a novel federated learning framework specifically designed for IoMT environments. This system preserves data confidentiality by training models locally without sharing raw data, while effectively handling non-IID data distribution across heterogeneous clients. Furthermore, Fed-SHAP-IDS integrates the SHAP (Shapley Additive Explanations) method to capture local feature importance, enabling interpretable, explainable, and interoperable insights for each client. By combining federated learning with SHAP-driven feature aggregation, our approach enhances both privacy and transparency, offering a robust and trustworthy solution for IoMT intrusion detection.
Download

Paper Nr:	268
Title:	GradSwap: Training-Free Diffusion for Efficient Face Swapping
Authors:	Emna BenSaid, Marwa Jabberi, Mohamed Neji and Adel M. Alimi
Abstract:	Face swapping remains a challenging problem in computer vision as it requires maintaining the source identity while adapting to the target's appearance, expressions, and pose. Existing methods often compromise either identity consistency or the fidelity of target attributes and typically rely on task-specific training or paired datasets. In this work, we introduce GradSwap, a gradient-guided diffusion framework built upon a pretrained unconditional Denoising Diffusion Implicit Model (DDIM). Rather than passively conditioning the diffusion process, GradSwap actively controls it through multi-gradient optimization, where each guidance term enforces a distinct visual constraint. This unified formulation enables a fine-grained balance between identity preservation, expression transfer, and photorealistic rendering within a single diffusion trajectory, achieving coherent and realistic swaps without any retraining. We evaluate our method on the FFHQ and FaceForensics++ datasets, showing that it surpasses state-of-the-art approaches in both visual realism and identity–attribute consistency across diverse conditions.
Download

Paper Nr:	269
Title:	Abductive Relative Majoritary Explanations for Random Forest Classifiers in the Context of Multi-Class Classification
Authors:	Louenas Bounia and Juba Agoun
Abstract:	This work extends the concept of majoritary reasons to random forests in multi-class classification. The ma-joritary reasons proposed by (Audemard et al., 2022d) are limited to binary classification by using a fixed cardinality constraint unsuited to relative majority voting. We introduce relative majoritary reasons, which dynamically adapt to the number of votes received by each class. Our contributions include: (1) theoretical results demonstrating that only trees voting for the predicted class are necessary to construct explanations, (2) an optimized encoding reducing complexity from O(n · \|F\|) to O(n · \|Fc\|) while preserving the same formal guarantees. Relative majoritary reasons can be computed in linear time via a greedy algorithm, and their minimal versions via PARTIAL MAXSAT encodings. We also propose preferred majoritary reasons via WEIGHTED PARTIAL MAXSAT. This extension improves the interpretability of random forests in practical contexts where multi-class classification is frequent.

Paper Nr:	271
Title:	DREAM: Dynamic, Reinforced, and Evasive Attack Model on Spatio-Temporal GNNs
Authors:	Mandru Suma Sri, Uddhav Narayan Gilda, Tarun Bisht and Virendra Singh
Abstract:	Graph Neural Networks (GNNs) have emerged as powerful tools for dynamic spatial and temporal predictions in recent years, due to their ability to understand and learn spatiotemporal data. However, their vulnerability to adversarial attacks still remains an open challenge, especially in dynamic and real-world settings. Existing attacks majorly focus on the effectiveness of the attack by applying static perturbations to all nodes, overlooking the attack's stealthiness. In this work, we present a dynamic adversarial attack framework, DREAM, targeting spatio-temporal GNNs, where vulnerable nodes are selected at each time step using Reinforcement Learning(RL), to which hybrid perturbations are generated via a context-aware formulation that incorporates global variability, mean deviation, and local temporal differences, and are clipped to preserve distributional realism. Extensive experiments on traffic datasets demonstrate that DREAM achieved the attack's stealthiness by retaining the Global RMSE while increasing Local RMSE to 899%, showing attack's effectiveness. Furthermore, the attack strategy is transferred to other graph models that are trained on non-traffic datasets, to show that DREAM is model-agnostic. Notably, DREAM is 14× faster than the state-of-the-art methods, highlighting its effectiveness, stealthiness, and efficiency of attack.
Download

Paper Nr:	282
Title:	Analyzing Degradation Mechanisms: An Explainable Multi-Task Learning Approach for Battery Forecasting
Authors:	Théo Heitzmann, Amel Hidouri, Ahmed Samet, Tedjani Mesbahi and Romuald Boné
Abstract:	This paper proposes an exogenous, explainable multi-task learning framework based on Long Short-Term Memory (LSTM) neural networks to model the behavior of lithium-ion batteries in electric vehicles and identify the key factors contributing to their degradation. Using a sliding window-based approach, the model enables to forecast not only the State of Health (SoH) but also the Internal Resistance (IR), two critical indicators of battery health. The objective is to predict the shape of battery degradation curves while simultaneously identifying the features that influence their evolution. Time series features on the statistical, temporal and spectral domains such as the Root-Mean-Square (RMS) value, the maximum frequency or the power bandwidth, are extracted from raw data consisting of voltage, current and temperature measurements. A Sequential Forward Floating Selection (SFFS) algorithm is then integrated to identify the most relevant features for prediction. To integrate an explainable layer into our model, we employ the SHapley Additive exPlanations (SHAP) framework to link feature importance with the evolution of battery degradation. This integration enhances transparency in battery health forecasting while providing deeper insight into the internal mechanisms of our data-driven approach. Finally, through extensive comparative experiments with state-of-the-art models, the proposed Explainable Multi-Task Long Short-Term Memory (EMT-LSTM) framework achieved demonstrably higher predictive accuracy.
Download

Paper Nr:	283
Title:	Topology-Guided Quantum GANs for Constrained Graph Generation
Authors:	Tobias Rohe, Markus Baumann, Michael Poppel, Gerhard Stenzel, Maximilian Zorn and Claudia Linnhoff-Popien
Abstract:	Quantum computing (QC) promises theoretical advantages, benefiting computational problems that would not be efficiently classically simulatable. However, much of this theoretical speedup depends on the quantum circuit design solving the problem. We argue that QC literature has yet to explore more domain specific ansatz-topologies, instead of relying on generic, one-size-fits-all architectures. In this work, we show that incorporating task-specific inductive biases – specifically geometric priors – into quantum circuit design can enhance the performance of hybrid Quantum Generative Adversarial Networks (QuGANs) on the task of generating geometrically constrained K4 graphs. We evaluate a portfolio of entanglement topologies and loss-function designs to assess their impact on both statistical fidelity and compliance with geometric constraints, including the Triangle and Ptolemaic inequalities. Our results show that aligning circuit topology with the underlying problem structure yields substantial benefits: the Triangle-topology QuGAN achieves the highest geometric validity among quantum models and matches the performance of classical Generative Adversarial Networks (GAN). Additionally, we showcase how specific architectural choices, such as entangling gate types, variance regularization and output-scaling govern the trade-off between geometric consistency and distributional accuracy, thus emphasizing the value of structured, task-aware quantum ansatz-topologies.
Download

Paper Nr:	286
Title:	A Lightweight Spatial-Temporal Graph Neural Network for Long-Term Time Series Forecasting
Authors:	Henok Tenaw Moges and Deshendran Moodley
Abstract:	We propose Lite-STGNN, a lightweight spatial-temporal graph neural network for long-term multivariate forecasting that integrates decomposition-based temporal modeling with learnable sparse graph structure. The temporal module applies trend-seasonal decomposition, while the spatial module performs message passing with low-rank Top-K adjacency learning and conservative horizon-wise gating, enabling spatial corrections that enhance a strong linear baseline. Lite-STGNN achieves state-of-the-art accuracy on four benchmark datasets for horizons up to 720 steps, while being parameter-efficient and substantially faster to train than transformer-based methods. Ablation studies show that the spatial module yields 4.6% improvement over the temporal baseline, Top-K enhances locality by 3.3%, and learned adjacency matrices reveal domain-specific interaction dynamics. Lite-STGNN thus offers a compact, interpretable, and efficient framework for long-term multivariate time series forecasting. The source code is available at https://github.com/HTMoges/Lite- STGNN.
Download

Paper Nr:	289
Title:	NextGen: Transforming Industrial Training through LLM-Driven 3D Environment Generation
Authors:	Mariam Loutfy, Nada Nasser, Nada Sharaf and Slim Abdennadher
Abstract:	Large Language Models (LLMs) are increasingly integrated into workflows ranging from text generation to code automation. This paper explores their potential in industrial training, focusing on automating the generation of 3D simulations of industrial machines. We propose a system that bridges CAD models with Unity environments via JSON structures generated with the aid of LLMs. The pipeline extracts geometric data from FreeCAD, reformats it through an LLM into Unity-compatible JSON, and automatically renders training-ready 3D scenes. Evaluation demonstrates that the approach significantly reduces development effort while preserving structural accuracy. Tests on simple shapes and multi-component assemblies confirm the feasibility of this pipeline, offering a promising step toward Industry 5.0 intelligent automation.
Download

Paper Nr:	292
Title:	A Case for Causal Reinforcement Learning in Longitudinal Vehicle Control
Authors:	Jule Schmidt, Xin Tao, Chelsea Sidrane, Swarup Kumar Mohalik, Akhil Prasad, Jana Tumova and Nils Jansen
Abstract:	Deep Reinforcement Learning (DRL) has succeeded in various applications, including autonomous driving. Still, it often struggles with robustness, generalizability, and training efficiency. Causal Reinforcement Learning (CRL) aims to tackle these problems, showing improved robustness and efficiency over standard Reinforcement Learning (RL) in many studies. However, these studies focus on simplified discrete benchmarks and tabular RL, while applications of deep CRL in complex settings are scarce. We apply deep CRL to longitudinal control, a continuous task in autonomous driving, using a physics-based environment with complex goals, and compare with standard DRL on various performance aspects. Our results show that the CRL controller’s training decreases the required training time by 40% compared to the DRL controller. with lower variance across different seeds. Whereas the CRL controller improves gradually during training, the DRL controller experiences setbacks. In robustness experiments, the CRL controller outperforms the DRL controller, maintaining a 24% higher mean reward with a 65% lower standard deviation under distribution shifts. These findings highlight the advantages of CRL in terms of training efficiency, stability, and robustness.
Download

Paper Nr:	293
Title:	Towards Intelligent Monitoring System Using Computer Vision
Authors:	Sandy Hoffmann, Arthur Rodrigues Fernandes, Vinicius Wolosky Muchulski, Stefan Sarkadi, Aldo von Wangenheim and Alison R. Panisson
Abstract:	This paper presents an intelligent monitoring system designed to enhance security in surveillance environments. The system leverages existing camera infrastructure, combining classical computer vision techniques with advanced deep learning models. To demonstrate the proposed approach, we explore two monitoring scenarios in a residential condominium: one involving the garage gate and another focused on the main pedestrian entrance. The garage gate scenario aims to determine the gate’s status and identify individuals attempting to enter, while the main entrance scenario focuses on detecting waiting individuals, such as delivery personnel. Both solutions employ multimodal Large Language Models (LLMs) to perform contextual analysis and classify detected individuals. The integrated system offers a practical and scalable solution for automated surveillance. In addition to an evaluation demonstrating the effectiveness of the proposed approach, we also discuss implementation costs and suggest future improvements, including integration with communication channels and hardware for automated gate closure.
Download

Paper Nr:	296
Title:	Are User Cognitive Load and Engagement Affected by the Improvement of AI-Generated Images Based on User-Input Prompts?
Authors:	Mari Saito and Seiji Yamada
Abstract:	While generative AI is reported to reduce academic and business workloads, the emergence of a new task-prompt engineering, which enables users to generate intended images-requires specific user expertise. This study investigates whether cognitive characteristics such as cognitive load and self-efficacy, previously examined in human task performance, also apply to collaborative tasks between humans and AI. To this end, we focus on the cognitive load of prompt input when AI generates images and examine the relationship with self-efficacy. Our findings indicate that prompt input imposes a higher cognitive load than automatic image generation, although self-efficacy can mitigate this load. Moreover, greater satisfaction with and perceived improvement in output images are associated with higher self-efficacy and lower cognitive load. This suggests that the human tendency to achieve goals through incremental steps also applies to collaborative tasks with AI. Specifically, in generative AI, the cognitive load of prompt input can be reduced by building interactions that make users feel the engagement is enhanced and the results are reflected, leading to a high sense of achievement in collaborative tasks with AI.
Download

Paper Nr:	298
Title:	Autonomous Control for Reusable Rocket Landing: Deriving a Robust Guidance through Deep Reinforcement Learning in a Custom 3D Simulation Environment
Authors:	Atin Chowdhury, Piyush Joshi, Vanya Awasthi and Nikhil Tripathi
Abstract:	Rocket landings in today’s aerospace industry are predominantly achieved using finely tuned control systems derived from precise analytical models. While effective, these traditional controllers demand extensive manual design and often struggle under highly nonlinear or uncertain dynamics. In this work, we propose the first reinforcement learning (RL)-based framework for autonomous rocket landing that eliminates the dependence on handcrafted control laws. We develop a custom 3D physics simulation environment that models key dynamics of rocket descent-including gravity, thrust, torque from off-center exhaust, rotational motions, and fuel consumption. The environment provides continuous state feedback encompassing position, velocity, orientation, angular velocity, and fuel level, while the agent performs thrust and attitude adjustments through discrete control inputs. Landing is formulated as a continuous-control RL task with a reward function that jointly optimizes landing precision, attitude stability, descent rate, and fuel efficiency. Through training, agents learn robust policies capable of achieving soft and stable landings on a designated pad without any prior knowledge of the system dynamics. Our experimentation demonstrates that reinforcement learning can autonomously discover control strategies that offer greater adaptability and scalability to complex coupled dynamics. The proposed framework provides an open foundation for future research in AI-driven aerospace guidance and control.
Download

Paper Nr:	299
Title:	SAMU-6GNet: DRL-Based Network Architecture for Dynamic Resource Allocation in Constrained Medical Environments
Authors:	Mahdia Slimi, Ibtissem Brahmi and Faouzi Zarai
Abstract:	Critical medical environments (emergency services, disaster zones, remote areas) require ultra-reliable communication networks capable of dynamically adapting to stringent operational constraints. This paper introduces SAMU-6GNet, a novel Deep Reinforcement Learning (DRL)-based network architecture for dynamic resource allocation in 6G-enabled medical systems. Our solution addresses key challenges including extreme demand variability, device heterogeneity (IoT(Internet of Things) sensors, medical equipment, drones), and strict QoS (Quality of Service) requirements for life-critical traffic. SAMU-6GNet’s core innovations include: 1. Medical relationship modeling through Graph Neural Networks (GNNs), reducing latency by 72% versus conventional approaches. 2. A medically-guided reward function enabling automatic priority scheduling, improving urgent data delivery by 23%. 3. Dynamic 6G network slicing that isolates critical flows, guaranteeing 98.7% QoS compliance for time-sensitive applications like robotic surgery (<10 ms latency). Experimental results demonstrate SAMU-6GNet’s superiority over existing methods (Round-Robin, static QoS) in: i)Energy efficiency (0.8 mJ/packet), ii) Responsiveness (5 ms reconfiguration), iii) Scalability (100+ node deployments). The architecture enables self-adaptive, resilient medical networks that align with clinical protocols while complying with emerging 6G standards.
Download

Paper Nr:	301
Title:	Quality Assessment of Datasets for Named Entity Recognition & Linking
Authors:	Hubert Nacmer and Szymon Olewniczak
Abstract:	The quality of datasets is an important issue among the natural language processing field. However, there is an apparent lack of methods for assessing quality of datasets for those problems. This study proposes potential dataset’s aspects that could be indicative of the dataset’s quality and presents experimental metrics for their measurement. The suggested aspects are: the grammar correctness of texts in the dataset, the occurrence of labels and entities across the dataset, the coherence of entities contained in the documents, the consistency of annotations throughout the dataset, and the completeness of annotations in the dataset. Furthermore, the experiments were performed on different datasets, which are available in multiple versions, i.e. in different stages of revision, and with different annotation approaches. Results show the relevance of the metrics in the assessment of the perceived dataset quality and demonstrate the impact of different revision and processing approaches on the metrics’ values across the different versions.
Download

Paper Nr:	302
Title:	A Lightweight Reinforcement Learning Framework for Othello Agents: Exploring Reward Shaping and Efficient Exploration
Authors:	Vanya Awasthi, Piyush Joshi, Atin Chowdhury and Nikhil Tripathi
Abstract:	Although Othello has been weakly solved through exhaustive search, developing lightweight reinforcement learning agents that perform competitively under strict computational constraints remains challenging. In this work, we study how dense reward shaping and exploration strategies influence the learning efficiency of CPU-only Othello agents trained using Deep Q-Networks (DQN). We evaluate eight exploration policies within a unified framework using a bounded, exponentially-scaled dense reward shaping signal derived from state-dependent positional features, combined with a shallow, search-limited dynamic programming (DP) lookahead used only during action selection. This design provides informative intermediate feedback while avoiding the cost of full tree search. Experimental results from self-play and random-opponent evaluation show that dense reward shaping substantially improves training stability and convergence. Among the evaluated strategies, Randomized Value Functions (RVF) achieve the highest performance, reaching a training win rate of approximately 53% and a test win rate of 70.86%. These findings provide empirical insights into the design of efficient reinforcement learning algorithms for board games under computational constraints.
Download

Paper Nr:	305
Title:	An Efficient ZDD-Based Method for Enumerating All Cost-Bounded Solutions of Graph Problems: Applications to Weighted Vertex Cover and Hamiltonian s-t Path Problems
Authors:	Teruyuki Miyamoto, Jun Kawahara and Shin-ichi Minato
Abstract:	This paper presents a unified framework for efficiently enumerating and compactly representing low-cost solutions to various combinatorial optimization problems on graphs using Zero-suppressed Binary Decision Diagrams (ZDDs). We propose a method for implicitly enumerating graph structures, such as vertex covers, independent sets, and paths, with a total weight below a specified threshold, and for efficiently constructing a compact ZDD that represents them. Furthermore, we introduce two speed-up techniques for the proposed method. Experimental results on multiple datasets demonstrate the efficiency of our approach, outperforming existing methods. For the vertex cover problem, in the best case, our method exhaustively enumerated more than 61,058 low-cost solutions 210 times faster than a commercial integer programming solver. These results suggest that our method is a valuable tool for real-world applications requiring diverse, high-quality solutions.
Download

Paper Nr:	308
Title:	Efficient Repair of Binarized Neural Networks Using Binary Activation Lookup Tables
Authors:	Sahar Alzahrani, Sven Schewe and Xiaowei Huang
Abstract:	Binarized Neural Networks (BNNs), with weights and activations constrained to {−1,+1}, provide high efficiency for edge and IoT devices but remain highly vulnerable to single-pixel perturbations that can trigger misclassifications in safety-critical applications. We propose a lookup-table (LUT)-based repair framework that corrects such misclassifications without retraining or modifying the BNN architecture. Repair tables are constructed from clean and perturbed training samples, mapping binary activation patterns to correct outputs. Three strategies are evaluated: VL1-only (using only the last hidden layer activations v L−1), Nested (hierarchical: v L−1 primary, with v L−2 used only to disambiguate low-confidence patterns), and Tuple (concatenation of both layers). Experiments are conducted on MNIST and FashionMNIST under two single-pixel perturbation scenarios: 1Pert and 10Pert. Both VL1-only and Nested are strictly correctness-preserving—they never degrade overall accuracy and always produce substantially more wrong-to-correct repairs than harmful correct-to-wrong changes. VL1-only achieves the highest overall accuracy improvements, while Nested provides superior repair quality by reducing harmful over-corrections and achieving higher repair success rates, with comparable coverage. Repairs are triggered infrequently (coverage below 23%) but are highly reliable when applied, with success rates above 58% and reaching 100% in some settings. The Tuple strategy proves ineffective due to extreme overspecificity. Overall, triggered repairs are predominantly beneficial. This generalizable, correctness-preserving approach enhances BNN reliability for lightweight deployment.
Download

Paper Nr:	312
Title:	CARJAN: Agent-Based Generation and Simulation of Traffic Scenarios with AJAN
Authors:	Leonard Frank Neis, André Antakli and Matthias Klusch
Abstract:	User-friendly modeling and virtual simulation of urban traffic scenarios with different types of interacting agents such as pedestrians, cyclists and autonomous vehicles remains a challenge. We present CARJAN, a novel tool for semi-automated generation and simulation of such scenarios based on the multi-agent engineering framework AJAN and the driving simulator CARLA. CARJAN provides a visual user interface for the modeling, storage and maintenance of traffic scenario layouts, and leverages SPARQL Behavior Tree-based decision-making and interactions for agents in dynamic scenario simulations in CARLA. CARJAN provides a first integrated approach for interactive, intelligent agent-based generation and simulation of virtual traffic scenarios in CARLA.
Download

Paper Nr:	316
Title:	A Human-in-the-Loop Framework for Integrating Human Perception in Image Clustering
Authors:	Jingbo Yan and Seiji Yamada
Abstract:	Although many studies have explored how human judgment can support model interpretability, effective evaluation methods are still lacking. There is also a need for interactive systems that can turn human feedback into concrete model improvements. To address these issues, we design a novel user interface that enables humans to perform patch-level clustering on image data. The clustering annotations are converted into pairwise constraints. These constraints guide constrained K-means to iteratively update the clustering results across multiple interaction rounds, allowing users to progressively refine the clusters. We validate the framework through a user study that includes pre- and post-interaction clustering quality assessments and questionnaires on quality improvements and system usability. Our user studies demonstrate that this interface successfully captures consistent perceptual structures, offering a systematic methodology for collecting human visual organization data that can inform and guide the development of more interpretable models.
Download

Paper Nr:	322
Title:	UAVs That Speak: Integrating VLMs and LLMs for UAV-Assisted Emergency Response
Authors:	David Lelis and Ayan Dutta
Abstract:	Timely and coordinated response is critical in emergency scenarios where first responders must rapidly locate and assist victims. We present a novel autonomous emergency response framework that integrates unmanned aerial vehicle (UAV) technologies, large language models (LLMs), and vision-language models (VLMs) to enhance situational awareness and decision-making. Our system leverages a server-UAV architecture in which the server ingests victim location data, computes optimal navigation paths, and coordinates communication with UAVs and first responders. Deployed UAVs autonomously capture aerial imagery, pre-process it, and apply deep learning pipelines, including the YOLO model for human detection and VLMs for semantic scene description. The resulting descriptions are then interpreted by an LLM, enabling high-level scene analysis and generating actionable responses. These insights are transmitted back to the server to advise first responders in real time. By combining autonomous path planning, robust perception, and natural language reasoning, our approach provides an autonomous, intelligent, and human-centric emergency response mechanism. We demonstrate the feasibility of the system and discuss its implications for future AI-driven disaster management, highlighting its ability to bridge the gap between autonomous sensing and effective human intervention. The results of the study showcase the framework’s ability to reliably recognize people and interpret the scene from an image and a message from the victim. The framework also shows potential in providing assistance instructions to first responders in providing effective aid. Experimental results show that the proposed framework achieves high classification accuracy (F1-scores ≈0.97) and strong semantic alignment across all evaluation metrics, demonstrating reliable scene understanding and response generation in both simulated and real-world emergency scenarios.
Download

Paper Nr:	324
Title:	A High-Performance and Explainable Framework for Movie Ticket Sales Forecasting
Authors:	Lorenzo Lefeve, Nourhène Ben Rabah, Lionel Prevost and Bénédicte Le Grand
Abstract:	The movie industry, long a cornerstone of popular entertainment, has recently encountered significant challenges. The COVID-19 pandemic drastically reduced theater attendance, while the rapid growth of streaming platforms continues to reshape audience habits and divert viewership. Consequently, investing in this highly competitive sector has become increasingly risky for industry professionals. This evolving landscape has spurred the development of new box office prediction designed to support data-driven decision-making through machine learning techniques. However, the development and deployment of such systems face a critical challenge: achieving the right balance between predictive performance and model explainability. A common limitation of existing approaches is the frequent omission of audience feedback data, as well as the difficulty of reconciling accuracy and transparency. High-performance models, often act as “black boxes,” making it difficult for professionals to interpret and trust their outputs. To overcome these shortcomings, we propose an end-to-end explainable and high performing ML framework for predicting cinema ticket sales. This framework aims to enhance model explainability without significantly compromising predictive performance. It covers the entire pipeline, including data collection, preprocessing, feature engineering, model training, and evaluation. We apply both model-specific and model-agnostic explainability methods to machine learning models. Our findings underscore the pivotal role of feature engineering and the importance of incorporating survey-based information. The results show that machine learning models perform best when integrating diverse information types, combining marketing, movie-related, and audience opinion features. This multi-faceted approach allows models to outperform industry experts in box office forecasting.
Download

Paper Nr:	325
Title:	Hybrid Learning Framework for Multi-Source Forest Monitoring
Authors:	Rim Douss and Imed Riadh Farah
Abstract:	High-dimensional spatio-temporal data present significant challenges for machine learning. Redundancy, non-linearity and class imbalance can hinder the efficient representation of data and make models difficult to interpret. This paper introduces a hybrid learning framework that combines representation learning and feature selection to overcome these limitations. It integrates linear and nonlinear dimensionality reduction techniques, such as Principal Component Analysis (PCA), Uniform Manifold Approximation and Projection (UMAP) and deep auto-encoders, with supervised feature selection based on Recursive Feature Elimination (RFE) and SHapley Additive exPlanations (SHAP). A multi-task learning paradigm is adopted to predict correlated forest indicators, such as species composition, basal area and structural diversity, jointly. When applied to real and synthetic multi-source datasets combining Sentinel-2 spectral–temporal data and LiDAR structural features, the framework achieved over 70% dimensionality reduction and 40% faster training, with predictive accuracy exceeding 85%. Further analysis of feature interpretability revealed strong relationships between key predictors and ecological processes. Overall, the proposed approach offers a scalable, interpretable and domain-independent solution for the intelligent modelling of complex spatio-temporal systems.
Download

Paper Nr:	326
Title:	A Health-Aware Multi-Objective Approach to Home Healthcare Routing and Scheduling
Authors:	Radhia Zaghdoud
Abstract:	Home Healthcare Routing and Scheduling (HHRS) requires balancing competing stakeholder objectives between operational efficiency, fair workload distribution, and patient-centered care. This paper presents a novel multi-objective optimization model that simultaneously minimizes total travel distance, balances caregiver workloads, and maximizes patient preference satisfaction. Our solution employs a hybrid algorithm that enhances the NSGA-II with a health-aware heuristic, dynamically prioritizing caregiver assignments for clinically vulnerable patients, such as the elderly and those with high blood pressure. Comprehensive evaluation across 21 adapted Solomon instances (C, R, and RC series) demonstrates the algorithm’s consistent performance. The method achieves an average hypervolume of 0.72, indicating strong overall solution quality across all test scenarios. Critically, the algorithm maintains substantial patient preference satisfaction, with average rates of 56–84% across different problem scales, successfully integrating quality-of-care considerations into routing decisions. The approach generates an average of 6–29 high-quality Pareto-optimal solutions per instance, providing planners with multiple viable routing options that capture different trade-offs between operational objectives and patient satisfaction. This work provides healthcare organizations with a practical decision-support tool that effectively strikes a balance between cost efficiency, workforce fairness, and patient-centered care delivery.
Download

Paper Nr:	328
Title:	Using Spatiotemporal Changes in Memory as a Form of Attention
Authors:	Fernando Fradique Duarte, Nuno Lau, Artur Pereira and Luís Paulo Reis
Abstract:	Attention-based agents have achieved much success in many areas of Artificial Intelligence. In the scope of Deep Reinforcement Learning, attention has proved to be an important technique, having contributed significantly to the improvement of both the performance of the agents and their interpretability. Over time, several different formulations of attention have been proposed in the literature. This work proposes a new formulation of attention based on the spatiotemporal changes that occur in memory at each timestep of agent-environment interaction. The formulation proposed is simple to implement and its computational complexity depends mainly on the distance metric used. This work uses the cosine similarity, but other distance metrics can be used. Several Deep Recurrent Reinforcement Learning architectures were used to test the proposed attention formulation in 6 Atari 2600 videogames. The Soft Top-Down Spatial Attention architecture was used as the baseline for all the performance comparisons carried out. The results obtained seem to indicate that the attention formulation proposed can achieve similar, and in some cases, better performance results when compared to the baseline architecture.
Download

Paper Nr:	330
Title:	A Multiplex Hypergraph-Based Approach for Community Detection in Scientific Collaboration Networks
Authors:	Imen Hamed, Wala Rebhi and Narjes Bellamine Ben Saoud
Abstract:	The increasing complexity of modern social networks, particularly scientific collaboration networks, has made community detection a central challenge in network analysis. While various approaches, ranging from simple graph-based approaches to multiplex-based approaches, have been proposed to model these networks and detect communities, each tends to overlook certain key aspects of real-world interactions. In this paper, we propose the use of multiplex hypergraphs, which effectively represent both multilayer relationships and higher-order interactions. Then, we introduce a novel algorithm based on Non-Negative Matrix Factorization (NMF) to detect cross-layer communities within a multiplex hypergraph of researchers. Our approach performs joint clustering of nodes and edges while incorporating regularization terms that reflect the hypergraph’s topology and node attributes. Finally, to assess the effectiveness of our approach, we compare it with existing methods based on hypergraph and multiplex graph models, using real-world scientific collaboration networks.
Download

Paper Nr:	331
Title:	A Neural Approach to ADHD Detection in Children: Enhanced EEG Analysis with Wavelet-Transformer Synergy
Authors:	Raja Dhiabi, Rim Walha and Fadoua Drira
Abstract:	The prediction of Attention Deficit Hyperactivity Disorder (ADHD) using Electroencephalogram (EEG) data poses a considerable challenge due to the intricate nature of brain signals and variations among individuals. Early diagnosis is vital for effective management. This study introduces an innovative ADHD prediction system that utilizes advanced EEG data processing techniques along with a transformer-based predictive model to accurately identify ADHD symptoms in children. The research employs an efficient pipeline for EEG data preprocessing and analysis, integrating the wavelet transform to generate informative 2D time–frequency representations, providing a richer description of neural activity compared to raw time-domain data. A transformer encoder is then applied directly to these 2D representations, leveraging the attention mechanism’s strength in modeling global relationships across time and frequency domains. Furthermore, a comparative analysis of various deep learning architectures is conducted to assess their performance in ADHD detection. The findings demonstrate the effectiveness of the proposed ADHD prediction system, showcasing enhanced performance compared to existing approaches by achieving an accuracy and an F1-score exceeding 99%.
Download

Paper Nr:	333
Title:	A Core Ontology for the Abductive Theory of Method
Authors:	Tezira Wanyana and Deshendran Moodley
Abstract:	We introduce the ATOM core ontology (ACO), a mid-level, reusable ontology that operationalizes the abduc-tive theory of method (ATOM) for scientific discovery. ACO formalizes the scientific discovery process from data to phenomena and phenomena to explanatory theories, and distinguishes theory maturation from plausible theories to potential theories and potential theories to strong theories. It is dual-anchored i.e., ontologically to BFO and to PROV-O for provenance, leveraging a published BFO-PROV-O mapping. We developed ACO following UPON and authored it in Protégé making use of standard OWL constructs. Consistency was verified with standard reasoners, and fitness for use was assessed by instantiating ACO in two domains i.e., domestic electricity consumption and arrhythmia detection where competency questions were executed as SPARQL queries over the domain ontologies. Using established design criteria for core ontologies, ACO fully meets axiomatization and formal precision, modularity extensibility, reusability, and separation of concerns. ACO offers a method-grounded, provenance-aware scaffold that accelerates the construction of domain application ontologies for scientific knowledge discovery.
Download

Paper Nr:	335
Title:	Agentic 3D-CNN: Cognitive-Aware Deep Model for Early Dyslexia Detection in Children
Authors:	Mohammad Abuhamdah, Hamdi Kchaou and Boudour Ammar
Abstract:	Early and accurate detection of dyslexia is critical for timely intervention and improved learning outcomes. This study proposes a 3D Convolutional Neural Network (3D-CNN) framework with agentic optimization for dyslexia classification in children aged 8–10 years using structural Magnetic Resonance Imaging (MRI). We first develop a baseline 3D-CNN architecture that learns hierarchical spatial patterns from volumetric brain imaging data, achieving 92% accuracy on a limited dataset of 125 subjects (20 dyslexic, 105 control). To further enhance performance, we introduce an agentic optimisation framework incorporating autonomous hyperparameter tuning, adaptive learning rate scheduling, and intelligent checkpoint selection. The enhanced Agentic 3D-CNN achieved 96.15% test accuracy, an F1-score of 0.96, and a Cohen’s Kappa of 0.9210, representing a 4.15 percentage-point improvement over the baseline. These preliminary results on a limited dataset suggest potential for early dyslexia detection, warranting larger-scale validation in clinical settings. This work presents initial evidence for a computationally efficient AI-assisted approach to learning disability detection.

Paper Nr:	336
Title:	Adaptive Multimodal Temporal Feature Attribution (AMTFA): A Novel Explainable AI Framework for IoT Intrusion Detection
Authors:	Moez Zhioua and Asma Amdouni
Abstract:	The proliferation of Internet of Things (IoT) devices has created unprecedented security challenges, necessitating robust intrusion detection systems (IDS) supported by explainable artificial intelligence (XAI) to ensure transparency and trust. This paper introduces a novel XAI framework, Adaptive Multimodal Temporal Feature Attribution (AMTFA), designed for comprehensive feature attribution in IoT security using the N-BaIoT dataset. AMTFA integrates multimodal data fusion, temporal pattern recognition with attention mechanisms, adaptive meta-learning, and hierarchical cross-modal validation to provide superior explanations of feature impbortance. By combining network traffic patterns, device behavioral profiles, and temporal sequences, AMTFA addresses critical limitations in existing XAI methods that treat IoT data as static or uniform. Experimental results demonstrate that AMTFA achieves superior performance across multiple evaluation metrics, including fidelity (0.847), robustness (0.923), consistency (0.891), and explanation quality (0.756), compared to baseline methods such as SHAP, LIME, and Permutation Importance. This work represents a significant advancement in explainable IoT security, providing actionable insights for security practitioners while maintaining high interpretability and reliability.
Download

Paper Nr:	340
Title:	ChronoSHAP Framework: Time-Aware SHAP for Time Series with ARIMA-Based Surrogates
Authors:	Yosr Baccari, Nesrine Khabou and Ismael Bouassida Rodriguez
Abstract:	Recent progress in explainable artificial intelligence (XAI) has yielded numerous techniques for interpreting deep neural networks. However, the vast majority of these methods are tailored for static, tabular, or image data, and often fail to capture the temporal dependencies intrinsic to time-series models. This paper introduces ChronoSHAP, a novel time-aware Shapley-based explanation framework that extends classical SHAP theory to sequential data by defining temporally coherent coalitions of features. The approach computes attributions over contiguous time windows rather than isolated points, preserving contextual information through autoregressive imputation strategies. To illustrate the feasibility and interpretability of the method, ChronoSHAP is applied to a heart-rate monitoring scenario using a simple recurrent prediction model. The resulting explanations demonstrate the ability of ChronoSHAP to localize temporal segments responsible for abnormal predictions while maintaining mathematical consistency with Shapley axioms. This study lays the initial validation for a future extended version incorporating large-scale experiments on predictive maintenance datasets and industrial sensor streams.
Download

Paper Nr:	345
Title:	Unveiling Concise Counterfactual Explanations and Feature Relevance in Machine Learning through Decision Trees
Authors:	Jurandir Silva, Eduardo Aguilar and Rogerio Salvini
Abstract:	Interpretability is defined as the ability of a human to understand why an AI model makes certain decisions. Interpretability can be achieved through the use of interpretable models, such as linear regression and decision trees, and through model-agnostic interpretation methods, which treat any predictive model as a “black box”. Another concept related to interpretability is that of Counterfactual Explanations, which show the minimal changes in inputs that would lead to different results, providing a deeper understanding of the model’s decisions. The approach proposed in this work exploits the explanatory power of Decision Trees to create a method that offers more concise explanations and counterfactual explanations. A key advantage of this approach is the automatic selection of relevant features, without the need for predefined parameters, enabled by the hierarchical structure of Decision Trees, which streamlines interpretation. Furthermore, their structure naturally supports intuitive counterfactual explanations by highlighting minimal changes in branching decisions. The results of the study indicate that Decision Trees not only explain the “why” of model decisions, but also show how different attribute values could result in alternative outputs.
Download

Paper Nr:	349
Title:	On the Non-Stationarity of Rewards in MCTS
Authors:	Peter Guba
Abstract:	The most popular tree policy used in MCTS, UCB1, is taken from the standard Multi-armed Bandit setting. One of its underlying assumptions is that the reward distributions of the arms remain stationary. In this paper, we argue that, in MCTS, that is often not the case and taking this into account should result in improved performance of the algorithm. We demonstrate this on random trees generated using our custom toolkit which allows us to compare the algorithm’s estimates to the trees’ true minimax values.
Download

Paper Nr:	352
Title:	Data-Driven Decision Support System for Managing Discrepancies in Textile Ordering Operations
Authors:	Miguel Martins, Daniela Quintas, Ivan Gomes, Cláudia Mendes Araújo and Rui Fonseca
Abstract:	Subcontracted dyeing in the textile industry frequently causes discrepancies between the quantities of fabric sent and received, resulting in stock imbalances and scheduling inefficiencies. This study presents a machine learning (ML) based decision support system, developed using real industrial data, to predict fabric losses and width, thereby reducing uncertainty. The framework integrates two regression models (for fabric loss percentage and final width), a classification model for categorical loss prediction, and an unsupervised clustering model to identify production patterns. Implemented with XGBoost and validated through 10-fold cross-validation, the models achieved high accuracy and interpretability. The proposed approach enables planners to anticipate losses, optimize subcontracting quantities, and improve inventory reliability. Results show a significant potential to reduce stock shortages and enhance subcontracting decisions, demonstrating the prac-tical impact of AI-based decision support in textile manufacturing.
Download

Paper Nr:	355
Title:	Improved Convolutional Neural Network for Imbalanced Multi-Class Mammogram Images Classification Based on BI-RADS
Authors:	Hana Mechria and Lamjed Ben Said
Abstract:	Convolutional neural networks (CNNs) have shown impressive results in classifying mammogram images, however, there are always ongoing efforts to improve their effectiveness through different strategies. In this paper, we focused on improving CNN, especially EfficientNetV2, by optimizing weight adjustment for more accurate classification. Our approach is designed to yield robust performance and tackle two key issues typically faced in mammogram analysis, including multi-class and imbalanced datasets. In order to accurately assess its effectiveness, the proposed approach was evaluated on an imbalanced dataset of multi-class mammogram images comprising six classes based on BI-RADS. The experimental results demonstrated consistently significant improvements across all main effectiveness measures, namely accuracy, recall, precision, F1-score, Area Under the Curve (AUC), and Precision-Recall Curve (PRC), highlighting the robustness and adaptability of the proposed approach. For instance, the improved EfficientNetV2 achieved an accuracy of 95.62% compared to 90.76% for the baseline EfficientNetV2. This improvement provides robust classification of imbalanced multi-class mammogram images based on CNN, enabling a more accurate Computer-Aided Diagnosis (CAD) system in breast cancer screening.
Download

Paper Nr:	358
Title:	Intelligent Mutation Enhanced PSO for Efficient and Resilient Microservice Scheduling
Authors:	Maroua Douiri, Imen Ben Mansour and Moncef Tagina
Abstract:	The rapid adoption of microservice architectures has reshaped the design of modern cloud and edge applications. This evolution is strongly supported by containerization technologies, which allow multiple services to be efficiently deployed on shared physical infrastructure. However, despite numerous studies on container-based microservice placement, many existing solutions remain limited by unstable service performance, significant communication overhead, and ineffective resource distribution. To overcome these challenges, this study introduces IM-PSO, an adaptive three-objective particle swarm optimization approach incorporating an intelligent mutation mechanism. The proposed algorithm exploits the ε-indicator to steer the search process toward the Pareto-optimal region while preserving diversity and ensuring balanced exploration and exploitation. Extensive experimental evaluations confirm that IM-PSO achieves superior placement decisions by effectively lowering network communication costs, reducing workload imbalance, and enhancing service robustness. These results demonstrate that the proposed approach provides a scalable and reliable solution for microservice deployment in dynamic and heterogeneous cloud environments, outperforming several established optimization techniques.
Download

Paper Nr:	363
Title:	Financial Time Series Augmentation Using Transformer Based GAN Architecture
Authors:	Andrzej Podobiński and Jarosław A. Chudziak
Abstract:	Time-series forecasting is a critical task across many domains, from engineering to economics, where accurate predictions drive strategic decisions. However, applying advanced deep learning models in challenging, volatile domains like finance is difficult due to the inherent limitation and dynamic nature of financial time series data. This scarcity often results in sub-optimal model training and poor generalization. The fundamental challenge lies in determining how to reliably augment scarce financial time series data to enhance the predictive accuracy of deep learning forecasting models. Our main contribution is a demonstration of how Generative Adversarial Networks (GANs) can effectively serve as a data augmentation tool to overcome data scarcity in the financial domain. Specifically, we show that training a Long Short-Term Memory (LSTM) forecasting model on a dataset augmented with synthetic data generated by a transformer-based GAN (TTS-GAN) significantly improves the forecasting accuracy compared to using real data alone. We confirm these results across different financial time series (Bitcoin and S&P500 price data) and various forecasting horizons. Furthermore, we propose a novel, time series specific quality metric that combines Dynamic Time Warping (DTW) and a modified Deep Dataset Dissimilarity Measure (DeD-iMs) to reliably monitor the training progress and evaluate the quality of the generated data. These findings provide compelling evidence for the benefits of GAN-based data augmentation in enhancing financial predictive capabilities.
Download

Paper Nr:	364
Title:	AGMO: Attention-Guided Metaheuristic Optimization for High-Dimensional Hyperparameter Tuning in Tiny-MLP Based Intrusion Detection
Authors:	Nasreddine Hamdi, Akram Belazi, Safya Belghith and Héctor Migallón
Abstract:	Hyperparameter optimization (HPO) remains computationally expensive, especially for deep learning models. This paper introduces Attention-Guided Metaheuristic Optimization (AGMO), a framework that enhances population-based optimizers using a Query–Key–Value (QKV) attention mechanism from Transformer architectures. The population serves as queries, elite solutions as keys, and historical search directions as values, enabling adaptive aggregation of informative directions during optimization.We demonstrate AGMO with Particle Swarm Optimization (AGMOPSO), where attention-weighted updates improve search efficiency. Experiments on the CEC2017 benchmark show AGMOPSO achieves faster convergence and superior accuracy versus state-of-the-art methods. In real-world HPO for intrusion detection (CIC-IDS2017 dataset), AGMOPSO reaches higher Average Precision Scores with fewer evaluations.AGMO provides a plug-and-play attention layer for metaheuristics, advancing attention-driven optimization for complex AutoML tasks.
Download

Paper Nr:	376
Title:	Filtering Goals of Necessity-Optimal Agents in Qualitative Possibilistic Recognition via Planning
Authors:	Brandon Rozek and Selmer Bringsjord
Abstract:	Rationally inferring the goal of an agent from observations of their actions is challenging. In the goal-recognition-as-planning literature, it is often assumed that the initial state of the environment is known. However, actors and observers do not always operate with complete knowledge. Instead, agents may be working with little information regarding the uncertainty of the environment. In this light, we revisit goal recognition in the context of qualitative possibilistic planning (QPP). Agents in this setting do not know the exact probability of an event occurring but are able to determine whether one event is more likely than another. More specifically, agents describe the uncertainty regarding the initial state and the outcome of actions qualitatively. We show that for rational actors, the observer should not filter goals solely based on necessity thresholds and instead propose a technique that takes into account whether the actor followed a necessity-optimal plan. Using our novel compilation CQPR, we find those necessity-optimal plans that additionally satisfy the observed action sequence by casting the overall problem as a QPP problem. Our formal results and experiments show that this approach is sound and may narrow down the potential goals that a necessity-optimal agent is pursuing.
Download

Paper Nr:	377
Title:	CAPRE: Counterfactual-Augmented Path-Based Reasoning for Extrapolative Link Prediction on Temporal Knowledge Graphs
Authors:	Dang Huynh, Minh Khau, Thanh Le and Bac Le
Abstract:	Temporal knowledge graphs have tremendous potential for applications, yet the approaches taken to perform link prediction on them continue to suffer from either degrading performance at distant timesteps for history-based interaction forecasting methods or ineffectiveness against over-squashed neighbors for reinforcement learning-based path reasoning algorithms, let alone the temporal embeddings, which are incapable of extrapolation beyond training timestamps. As an effort to tackle the extrapolative link prediction with generalization capability and efficient neighborhood handling, we present CAPRE, or Counterfactual-Augmented Path-based Reasoning for Extrapolation, a methodological framework built upon path-based reasoning to ensure extrapolative prediction while also sampling negative counterfactuals to help the scoring network distinguish paths of true positives against highly similar negatives. The proposed method also demonstrates generalizability with increasing performance at future timestamps, beyond the training’s temporal cutoff, in stark contrast with most of the advanced extrapolation models. Extensive benchmarks on GDELT and ICEWS reveal CAPRE surpasses state-of-the-art methods by over 10% in MRR and Hits@1 and 8% in Hits@3. Temporal evaluations further confirm high extrapolation capability, aligning with design goals and enabling future improvements.
Download

Paper Nr:	378
Title:	Adapter-Based Multi-Document Summarisation: Opinion Summarisation Use Case
Authors:	Kushan Hewapathirana, Nisansa de Silva and C. D. Athuraliya
Abstract:	This study explores adapter-based fine-tuning to enhance the PRIMERA model for opinion summarisation. PRIMERA, a state-of-the-art multi-document summarisation (MDS) model, exhibits strong transfer potential owing to its pre-training on large-scale MDS corpora. Leveraging adapter architectures, this work demonstrates substantial improvements when extending PRIMERA to opinion summarisation through parameter-efficient fine-tuning. In addition, an LLM-based evaluation paradigm is introduced using the DeepEval framework, enabling semantic and sentiment-aware assessment beyond lexical-overlap metrics such as ROUGE. To improve training efficiency, an agentic optimisation framework is proposed, where evaluation reasoning guides iterative adapter configuration, reducing fine-tuning cycles while maintaining performance. Results show that adapter-augmented PRIMERA surpasses the opinion summarisation baseline ADASUM, establishing a reproducible, interpretable, and computationally efficient path for low-resource MDS. Overall, this work highlights how adapter-based fine-tuning and reasoning-guided optimisation together advance both performance and applicability in opinion summarisation.
Download

Paper Nr:	379
Title:	Not All Countermodels Are Equal
Authors:	Daniel Crowley, Daniel Le Berre and Yakoub Salhi
Abstract:	We introduce a semantic framework for analysing the robustness of counterexamples in propositional reasoning. The central concept is that of a (T,k)-resistant countermodel: given a finite set T of extensive (transformation) operators and a nonnegative integer k, an interpretation is (T,k)-resistant if it continues to falsify the formula after up to k successive applications of operators from T. Building on this notion, we define a quantitative fragility measure for proof obligations, capturing how many successive relaxations are required before a given countermodel ceases to exist. We then study two key families of weakening operators: (1) Hamming-based dilations, which expand the model set via metric neighborhoods, and (2) Forgetting-based operators, which remove literal occurrences, entire variables, or arbitrary subformulas. For both families, we analyse the resistance properties of countermodels, with a focus on computational complexity. We also provide SAT-based encodings to effectively compute (T,k)-resistant countermodels.
Download

Paper Nr:	381
Title:	AMSAS: A BERT-ViT Fusion Framework for Real-Time Multimodal Sentiment Analysis
Authors:	Molka Kharoubi, Ibtissem Khedher and Neila Rjaibi
Abstract:	In the Big Data era, social media platforms like Twitter and Instagram have become rich sources of multimodal content where users express emotions through both text and visual elements. While traditional sentiment analysis methods focus primarily on textual data, they prove inadequate for multimodal content where images often modify or intensify the expressed sentiment. This critical gap has spurred significant research interest in multimodal sentiment analysis. Early multimodal approaches relied on simplistic feature concatenation or averaging across modalities, failing to capture the intricate relationships between text and images. Current state-of-the-art methods, while improved, still demonstrate limited accuracy, resulting in frequent sentiment misclassifications that hinder practical applications. To overcome these challenges, we present AMSAS (Advanced Multimodal Sentiment Analysis System), an innovative framework that achieves breakthrough performance in emotion classification. Our system combines BERT’s advanced natural language processing capabilities with ViT’s powerful visual feature extraction, enabling more sophisticated cross-modal analysis. Through an intelligent fusion mechanism, AMSAS attains unprecedented classification accuracy of 95%, while maintaining real-time processing capabilities essential for social media applications.
Download

Paper Nr:	383
Title:	Source Speech Reconstruction for Many-to-Many and One-to-One Voice Conversion
Authors:	Zbyněk Lička, Anton Firc and Kamil Malinka
Abstract:	As voice conversion (VC) systems grow in realism and accessibility, concerns over their misuse – particularly in deepfake scams – have intensified. In this paper, we demonstrate an approach to reconstructing attacker speech, which is applicable to one-to-one (O2O) and many-to-many (M2M) voice conversion. Unlike previous solutions based on extracting high-dimensional embeddings, we demonstrate that the attacker voice is deeply embedded into the VC model and thus it cannot be removed. Furthermore, we demonstrate that we can subsequently reconstruct their voice using the original VC model in order to link the attacker identity back to the deepfake audio. We show that current O2O VC models enable reconstruction without additional training. For M2M models, we introduce a source speaker classifier to facilitate reconstruction. Empirical evaluation across state-of-the-art models (MaskCycleGAN-VC, StarGANv2-VC) demonstrates that reconstructed speech achieves as low as EER 0.14 % for M2M VC reconstruction and 5.15 % EER for O2O VC reconstruction. As VC models improve, this method offers a scalable path for source speaker attribution in forensic and security applications.
Download

Paper Nr:	386
Title:	RAG-Enhanced Prompt Compression with Need-Oriented Knowledge for Dialog Based Embodied Navigation
Authors:	Hiroaki Shimoma, Sudesna Chakraborty, Takeshi Morita, Aoi Ohta, Masaki Asada, Shusaku Egami, Takanori Ugai and Masahiro Hamasaki
Abstract:	In the field of Embodied AI, where agents learn through interaction with virtual environments, Large Language Models (LLMs) and commonsense reasoning based on environmental knowledge are used to interpret user requests. Existing dialog-based navigation systems in VirtualHome embed all environmental knowledge into navigation prompts. However, as the number of objects increases, the prompt length increases, reducing the accuracy and increasing the computational cost. To address this, we propose a prompt compression technique that uses Retrieval-Augmented Generation (RAG) with a need-oriented environmental knowledge base grounded in Murray’s theory of needs for dialog-based embodied navigation. Our system links each need to the corresponding actions, descriptions, and objects, and infers user needs using an RAG-based approach. This approach aims to achieve prompt compression by storing and referencing environmental knowledge externally and retrieving only the relevant objects that satisfy the user’s needs using an LLM. We evaluated using an OpenEQA-based VirtualHome dataset and compared it with a purely LLM-based approach. Our results confirmed that the RAG-based approach effectively reduced the prompt length.
Download

Paper Nr:	387
Title:	Comparison of Dimensionality Reduction Methods for MCI Classification Using Machine Learning and Static Functional Connectivity Based on Resting-State Functional Magnetic Resonance Imaging
Authors:	Ryosuke Minami, Ryo Hatano and Hiroyuki Nishiyama
Abstract:	Early detection and treatment of dementia are important, and mild cognitive impairment (MCI) has attracted considerable attention. However, diagnosing MCI accurately using any single medical technology is challenging. In this study, we examined resting-state functional magnetic resonance imaging (rs-fMRI) data to classify patients as exhibiting MCI or as cognitively normal (CN). To classify MCI, the analysis of functional connectivity (FC), specifically static FC (sFC), from rs-fMRI data was performed. Although FC analysis is often used in conjunction with rs-fMRI and AI technology; it typically generates a large number of features, which causes overfitting. Therefore, we focused on dimensionality reduction (DR), specifically, the three DR methods: recursive feature elimination (RFE), recursive feature selection (RFS), and recursive feature addition (RFA). A comparison of their classification performances and computational costs revealed that although RFS exhibited the lowest computational cost; RFE showed the highest performance, with an F1 score of 1.000. Thus, the proposed approach may contribute to the early and efficient detection of MCI patients.
Download

Paper Nr:	392
Title:	Tool-Augmented Disambiguation LLM Agent Utilizing UMLS for Biomedical Entity Linking
Authors:	Luka Blašković, Nikola Tanković and Ivo Ipšić
Abstract:	Biomedical entity linking (BEL) aims to normalize textual mentions of medical concepts to standardized identifiers within the Unified Medical Language System (UMLS), which integrates over 190 vocabularies and 3.45 million Concept Unique Identifiers (CUIs). The scale and lexical variability of UMLS make BEL a persistent challenge. Existing lexical and neural methods achieve limited disambiguation accuracy and underuse the ontology’s structure. This paper presents a tool-augmented large language model (LLM) agent that performs self-directed disambiguation by orchestrating lexical, semantic, and relational UMLS utilities. The agent integrates SapBERT-based candidate retrieval with semantic-type validation, ontology-driven neighbor expansion, and iterative reasoning to refine concept selection. This hybrid symbolic–neural framework enables dynamic grounding and interpretable decision-making. On a 1,000-mention subset of the MedMentions ST21pv benchmark, the proposed system achieves 68.3% top-1 exact CUI match, outperforming QuickUMLS (20.5%), BioBERT (33.2%), GPT-4o single-stage few-shot prompting approach (35.1%), SciSpacy (39%), and SapBERT (51.8%). The study demonstrates that ontology-grounded tool use and agentic reasoning substantially improve BEL precision compared to static embedding or direct LLM inference. Contributions include (i) a unified agentic framework for UMLS-guided linking, (ii) integration of structured knowledge with LLM reasoning via function-calling tools, and (iii) an empirical benchmark establishing the benefits and trade-offs of tool-augmented BEL in accuracy, interpretability, and cost.
Download

Paper Nr:	394
Title:	Benchmarking and Transfer Learning for Hyperparameter Optimization of Graph Neural Networks
Authors:	Marek Dědič and Michal Bělohlávek
Abstract:	Graph Neural Networks (GNNs) rely critically on effective hyperparameter optimization (HPO), but comprehensive HPO benchmarks for graph learning tasks remain limited. This paper addresses this gap by presenting an extensive comparison of HPO strategies for GNNs, ranging from simple random search to sophisticated Sequential Model-Based Optimization (SMBO) techniques like Bayesian Optimization (BO) and Tree-structured Parzen Estimators (TPE). Furthermore, we explore meta-learning for transfer learning, investigating its potential to accelerate HPO on new tasks by leveraging knowledge from past runs. Our results provide a practical comparative guide and demonstrate the viability of using meta-learned knowledge to significantly accelerate GNN HPO.
Download

Paper Nr:	399
Title:	Evaluating the Sample Efficiency of Solomonoff-Inspired Predictive Models
Authors:	Cleyton Slaviero, Edward Hermann Haeusler and Luiz Carlos Castro Guedes
Abstract:	Solomonoff induction offers a universal Bayesian view of learning but it is incomputable in its full form. In this work, we propose a computable approximation framework based on time- and length-bounded truncations of the universal distribution. We analyze how these truncated forms relate to PAC guarantees, showing that they induce generalization bounds driven by the Kolmogorov complexity of the target hypothesis. We then introduce a compression-based proxy that preserves these theoretical insights while remaining practical. We propose an experiments on synthetic data to illustrate that both the predicted error rates and the sample complexity reflect the complexity of the underlying hypothesis class, even when the learner is not time-bounded. These results support the claim that Solomonoff-style bounds can serve as an alternative perspective on PAC analysis, connecting structural properties of data and hypotheses with effective generalization behavior.
Download

Paper Nr:	401
Title:	Modeling Fantasy Premier League Player Performance with a Two-Component Poisson Mixture Transformer
Authors:	Weronika Wiechno, Piotr Duch and Bartosz Bartosik
Abstract:	Accurate prediction of Fantasy Premier League (FPL) player performance is a challenging task due to multiple sources of variability affecting match outcomes. In this work, we propose the Two-Component Poisson Mixture Transformer (2PM-Transformer), a novel deep learning architecture that extends the standard Transformer by modeling player points as a two-component Poisson mixture. This probabilistic formulation explicitly captures both the excess of low-scoring outcomes and the high-scoring performance regimes observed in FPL data. Extensive experiments compare the 2PM-Transformer with baseline models, including a vanilla Transformer, Long Short-Term Memory (LSTM) network, and Moving Average, using historical FPL data from the 2021/22 and 2023/24 seasons. Results show that the 2PM-Transformer consistently outperforms all baselines, achieving a Mean Squared Error (MSE) of 9.31 compared to 9.89 for the Transformer, 9.80 for the LSTM, and 10.51 for the Moving Average. Incorporating player data from additional football leagues, translated into FPL scoring rules, further improves performance, reducing the MSE to 9.19.
Download

Paper Nr:	406
Title:	Empirical Analysis of Adversarial Robustness and Explainability Drift in Cybersecurity Classifiers
Authors:	Mona Rajhans and Vishal Khawarey
Abstract:	Machine learning (ML) models are increasingly deployed in cybersecurity applications such as phishing detection and network intrusion prevention. However, these models remain vulnerable to adversarial perturbations—small, deliberate input modifications that can degrade detection accuracy and compromise interpretability. This paper presents an empirical study of adversarial robustness and explainability drift across two cybersecurity domains: phishing URL classification and network intrusion detection. We evaluate the impact of L∞-bounded Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) perturbations on model accuracy and introduce a quantitative metric, the Robustness Index (RI), defined as the area under the accuracy–perturbation curve. Gradient-based feature sensitivity and SHAP-based attribution drift analyses reveal which input features are most susceptible to adversarial manipulation. Experiments on the Phishing Websites and UNSW-NB15 datasets show consistent robustness trends, with adversarial training improving RI by up to 9% while maintaining clean-data accuracy. These findings highlight the coupling between robustness and interpretability degradation and underscore the importance of quantitative evaluation in the design of trustworthy, AI-driven cybersecurity systems.
Download

Paper Nr:	409
Title:	Multi-Modal Forecasting: An Advanced Framework for Emerging Technology Prediction in Blockchain Patents
Authors:	Emna Rassâa and Asma Amdouni
Abstract:	Forecasting emerging technologies is pivotal for strategic decisions in research, industry, and policy. This paper proposes Multi-Modal Forecasting (MMF), a unified framework that integrates semantic, network, temporal, and citation analyses with multi-metric emergence quantification to improve accuracy and interpretability beyond single-modality methods. Applied to 3,596 blockchain patents (1990–2020), MMF identifies major technology clusters, maps their evolution, and forecasts future trajectories. MMF attains 87% predictive accuracy (R2), outperforming the strongest baseline (LSTM) by 7.4%. Emergence scoring highlights digital assets, smart contracts, and blockchain infrastructure as leading candidates, with projected three-year growth of 85.7%, 44.4%, and 37.2%, respectively. MMF offers a strong balance of accuracy, interpretability, and computational efficiency for technology forecasting and strategic planning.
Download

Paper Nr:	418
Title:	Hallucinate or Memorize? The Two Sides of Probabilistic Learning in Large Language Models
Authors:	Junichiro Niimi
Abstract:	Large language models (LLMs) have been increasingly applied to a wide range of tasks, from natural language understanding to code generation. While they have also been used to assist in citation recommendation, the hallucination of non-existent papers remains a major issue. Building on prior studies, this study hypothesizes that an LLM’s ability to correctly produce bibliographic records depends on whether the underlying knowledge is generated or memorized, with highly cited papers (i.e., more frequently appear in the pretraining corpus) showing lower hallucination rates. We therefore assume citation count as a proxy for training data redundancy (i.e., the frequency with which a given bibliographic record appears in the pretraining corpus) and investigate how citation frequency affects hallucinated references in LLM outputs. Using GPT-4.1, we generated and manually verified 100 citations across twenty computer-science domains, and measured factual consistency via cosine similarity between generated and authentic metadata. The results revealed that (i) citation count is strongly correlated with factual accuracy, (ii) bibliographic information becomes almost verbatim memorized beyond roughly 1,000 citations, (iii) memory interference occurs when multiple highly cited papers share similar content, and (iv) memorization is hierarchical that paper titles and first authors are prioritized over journal names and numeric metadata. These findings indicate a threshold where generalization shifts into memorization, with highly cited papers being nearly verbatim retained in the model.
Download

Paper Nr:	419
Title:	Reinforcement Learning for Parameterized Quantum State Preparation: A Comparative Study
Authors:	Gerhard Stenzel, Michael Kölle, Tobias Rohe, Leo Sünkel, Julian Hager and Claudia Linnhoff-Popien
Abstract:	We extend directed quantum circuit synthesis (DQCS) with reinforcement learning from purely discrete gate selection to parameterized quantum state preparation with continuous single-qubit rotations Rx, Ry, and Rz. We compare two training regimes: a one-stage agent that jointly selects the gate type, the affected qubit(s), and the rotation angle; and a two-stage variant that first proposes a discrete circuit and subsequently optimizes the rotation angles with Adam using parameter-shift gradients. Using Gymnasium and PennyLane, we evaluate Proximal Policy Optimization (PPO) and Advantage Actor–Critic (A2C) on systems comprising two to ten qubits and on targets of increasing complexity with λ ranging from one to five. Whereas A2C does not learn effective policies in this setting, PPO succeeds under stable hyperparameters (one-stage: learning rate approximately 5 × 10−4 with a self-fidelity-error threshold of 0.01; two-stage: learning rate approximately 10−4). Both approaches reliably reconstruct computational basis states (between 83% and 99% success) and Bell states (between 61% and 77% success). However, scalability saturates for λ of approximately three to four and does not extend to ten-qubit targets even at λ = 2. The two-stage method offers only marginal accuracy gains while requiring around three times the runtime. For practicality under a fixed compute budget, we therefore recommend the one-stage PPO policy and outline avenues to improve scalability.
Download

Paper Nr:	421
Title:	AIgmented HUMANity: A Pragmatic Framework for AI-Guided Human Enhancement
Authors:	Kacper Boś, Nikolas Jerzy Feduniewicz and Anna Bryniarska
Abstract:	Human enhancement is entering a new phase where artificial intelligence (AI) links biological and mechanical dimensions of capability. This paper introduces AIgmented HUMANity – a pragmatic, 1-2-year concept for integrating sensing, reasoning, and actuation across medical and assistive systems. Rather than proposing speculative transhumanism, it unifies existing, evidence-based technologies under a transparent and privacy-preserving framework. We synthesize clinical evidence from AI-driven prevention, rehabilitation, and augmentation, including randomized trials in cancer screening, cardiology, robotics, and neuroprosthetics. These results demonstrate that AI can operate safely in real-world contexts and achieve parity with expert performance. Building on this evidence, we outline a modular blueprint for low-data, interpretable AI systems capable of adaptive human-machine interaction. The proposed framework establishes a pathway from assistive automation toward synergistic human-AI integration. It emphasizes clinical transparency, ethical alignment, and technical feasibility within current regulatory constraints, offering a near-term vision of augmentation as an extension of care rather than replacement of human agency, and a foundation for future adaptive human-AI ecosystems.
Download

Paper Nr:	426
Title:	Empowering Financial Controlling with Artificial Intelligence: An Empirical Evaluation
Authors:	Korbinian Markl, Michael Diener, Michaela Polz and Felix Rößle
Abstract:	This study empirically evaluates the performance of Artificial Intelligence in executing authentic financial controlling tasks, addressing a gap in research on AI in financial controlling and management accounting. We employ a triangulated evaluation framework integrating self-assessment, third-party review by advanced language models, and benchmarking against expert solutions by the authors. Drawing on authentic and complex case studies from Horváth et al. (2012), encompassing Reading Comprehension, Data Visualization Interpretation, and Multi-Format tasks, we evaluate models including Grok and Perplexity. Outputs were rated on a scale from 0 (worst) to 10 (best) across five criteria (correctness, completeness, transparency, soundness, and relevance) - with each task repeated 10 times to account for non-determinism. Results reveal inflated self-perceptions tempered by external scrutiny, yielding aggregate downgrades for both Grok and Perplexity, with these deviations being statistically significant. Multi-Format tasks outperformed others, with relevance being consistently strong, but soundness lagging. While Large Language Models demonstrate promise in automating routine tasks and enhancing decision support, persistent issues in methodological rigor and output variability underscore the need for human oversight to ensure reliability in volatile business environments.
Download

Paper Nr:	428
Title:	Towards Trustworthy LLM Decision Support through MCDA Integration
Authors:	Andrii Shekhovtsov, Amirkia Rafiei Oskooei and Wojciech Sałabun
Abstract:	Large Language Models (LLMs) have become popular for decision support, yet their inherent limitations, such as non-deterministic outputs, opaque reasoning, and susceptibility to biases, raise serious concerns about reliability. On the other hand, Multi-Criteria Decision Analysis (MCDA) offers mathematically rigorous, deterministic, and transparent frameworks for structured decision-making, but remains largely inaccessible to non-experts due to complexity and technical requirements. In this position paper, we argue that integrating LLMs with MCDA through an agent-based architecture creates a powerful hybrid system addressing the complementary weaknesses of both approaches. We propose a framework where LLM-based agents serve as an orchestrators, providing natural language interfaces for problem formulation, preference elicitation, and data gathering, while delegating core decision logic to a deterministic MCDA tool. This integration democratizes access to robust decision-making tools while grounding LLM outputs in transparent, reproducible mathematical frameworks, positioning MCDA-enhanced LLM agents as a promising solution for reliable, accessible, and trustworthy decision support.
Download

Paper Nr:	432
Title:	Sample Complexity of Identifying the Nonredundancy of Nontransitive Games in Dueling Bandits
Authors:	Shang Lu and Shuji Kijima
Abstract:	Dueling bandit is a variant of the Multi-armed bandit to learn the binary relation by comparisons. Most work on the dueling bandit has targeted transitive relations, that is, totally/partially ordered sets, or assumed at least the existence of a champion such as Condorcet winner and Copeland winner. This work develops an analysis of dueling bandits for non-transitive relations. Jan-ken (a.k.a. rock-paper-scissors) is a typical example of a non-transitive relation. It is known that a rational player chooses one of three items uniformly at random, which is known to be Nash equilibrium in game theory. Interestingly, any variant of Jan-ken with four items (e.g., rock, paper, scissors, and well) contains at least one useless item, which is never selected by a rational player. This work investigates a dueling bandit problem to identify whether all n items are indispensable in a given win-lose relation. Then, we provide upper and lower bounds of the sample complexity of the identification problem in terms of the determinant of A and a solution of x ⊤A = 0 ⊤ where A is an n×n pay-off matrix that every duel follows
Download

Paper Nr:	433
Title:	A-RailYOLOM: Unified Framework for Railway Scenes Understandin
Authors:	Dorsaf Sebai and Manel Zouaoui
Abstract:	This paper introduces the very first multitask lightweight unified model designed for comprehensive perception in railway environments. Building on transfer learning, we adapt the A-YOLOM architecture, originally developed for automotive applications, to jointly perform object detection and semantic segmentation of railway-specific elements. The proposed approach constitutes the first attempt to integrate both tasks within a single network specifically tailored for railway scene understanding. The model is fine-tuned on the RailSem19 dataset to detect vehicles, pedestrians, signs, and signals, while segmenting structural components such as rails, tracks, and poles. To evaluate generalization, we conduct both in-domain and out-of-domain experiments. As a first ever railway-specific multi-task model, in-domain evaluation on RailSem19 achieves strong performance. Out-of-domain testing on a new dataset collected from the Tunisian National Railway Company (SNCFT) under varying environmental and acquisition conditions shows slight performance drops, highlighting robustness to domain shift. Results also demonstrate the feasibility of deploying the proposed model on edge devices for real-world railway scene understanding.

Paper Nr:	434
Title:	Towards AI-Enabled Training Needs Analysis Using Dual AI-Agent Collaboration
Authors:	Nikilkumar Patel, Peter J. Barclay, Janice McMillan and David McGuire
Abstract:	A comprehensive Training Needs Analysis is essential for effective HR development and organisational growth. However, traditional approaches often fall short due to limitations in scale, labour intensiveness, or resource and expertise constraints. To address these challenges, we propose an AI-driven automated platform for conducting TNA at scale with unstructured data. Our prototype features a dual-agent system powered by Large Language Models. The Disseminator Agent performs knowledge extraction and deep data analysis, then the Formulator Agent produces novel intellectual ideas, actionable insights, and formatted reports, facilitating final human verification and decision-making. We also outline a pragmatic plan for AI monitoring and platform evaluation – critical components for successful AI adoption in industry. Our proposed design is currently being implemented for evaluation and deployment in a business setting.
Download

Paper Nr:	436
Title:	Dynamic K-Prototypes for Mixed Data Stream Clustering
Authors:	Siwar Gorrab, Fahmi Ben Rejab and Kaouther Nouira
Abstract:	Many of the real-life data mining applications work with dynamic datasets. Herewith, humongous volumes of categorical and numeric data are continuously generated at a very giant speed and volume, making it complicated for traditional clustering algorithms to create and maintain the desired clusters. Incremental and decremental learning is the key of this challenge. In this paper, we present a novel dynamic k-prototypes clustering method that is capable of handling a bulk of updates, owing to the training mixed samples which become available one after another over time. It deals with both incremental and decremental attribute, object and class learning spaces within time and memory restrictions. This was based on the split and merge techniques. Experiments performed on real mixed data sets emphasize our proposal’s efficiency and that it outperforms the conventional k-prototypes and other similar methods based on different evaluation measures.
Download

Paper Nr:	437
Title:	Toward Explainable Diagnosis: A Neurosymbolic Approach
Authors:	Ciro Listone, Vadim Malvone and Aniello Murano
Abstract:	The growing adoption of Artificial Intelligence (AI) in medicine highlights the limitations of neural models, which often lack transparency and verifiability. This work presents a theoretical formalization of a neurosym-bolic game model that integrates an autoencoder with Obstruction Logic (OL) to model the diagnostic process as structured reasoning guided by knowledge and formal constraints. The autoencoder learns relationships among symptoms from data, producing continuous representations that enrich the clinical picture and capture latent structure. These representations serve as input to a logical module based on OL, a dynamic and tractable formalism that progressively eliminates diagnostic hypotheses inconsistent with established medical knowledge. The combined system merges the inferential power of neural networks with the rigor and reliability of symbolic reasoning, offering interpretable, consistent outcomes. This approach aims to enable decision-support tools that maintain clinical trustworthiness while leveraging data-driven insights, providing a pathway toward more transparent and verifiable AI-assisted medical diagnostics.
Download

Paper Nr:	438
Title:	A Novel Hybrid Deep Learning Model for Context-Aware Subjectivity Classification of Textual Data
Authors:	Rim Chiha, Radhia Toujani and Hela Ltifi
Abstract:	Subjectivity detection is a key task in opinion mining, aiming to distinguish factual from subjective content. This paper introduces a hybrid BERT–CNN–Fuzzy Logic framework to capture both explicit and implicit subjective cues. CNNs are employed to extract local patterns from BERT embeddings, while fuzzy logic addresses uncertainty and semantic ambiguity during feature extraction. A POS-tag–guided optimization strategy is further applied to reduce redundancy and preserve linguistically relevant features. Experimental results on two benchmark datasets show that the proposed approach achieves competitive performance and improved robustness compared with state-of-the-art methods.
Download

Paper Nr:	444
Title:	Detection of Dangerous Driving Event Using Qualitative Spatial Reasoning from the Video
Authors:	Kazuko Takahashi and Yurika Yamaguchi
Abstract:	We describe a method for extracting dangerous driving behavior using data recorded by dashboard cameras. While a machine learning approach is used in most studies, it typically requires a vast amount of data and extensive time for training each event. Furthermore, the semantics of each scene or event are not explicitly provided. To address these drawbacks, we propose a logical approach using qualitative spatial reasoning (QSR). We define a novel QSR tailored to our objectives, represent each scene through relative positional relationships between objects without numerical values, and extract events in the video by logical reasoning. We apply our method to both simulator-generated and real-world video data. As a result, we have found that it is possible to extract events correctly with only a small amount of data.
Download

Paper Nr:	445
Title:	Optimizing Road Flow with Genetic Algorithms and Deep Learning
Authors:	Alex Lewis, Rina Azoulay and Esther David
Abstract:	In this study, we propose a hybrid framework that combines deep learning and genetic algorithms to improve road-capacity utilization by jointly managing traffic speed and flow. The framework follows a two-stage pipeline. First, we train predictive models, based on Long Short-Term Memory (LSTM) networks and Graph Neural Networks (GNNs) to forecast traffic speed from both current observations and historical patterns. Second, we employ a genetic algorithm that leverages these predictions to search for near-optimal flow configurations that enhance network performance. Unlike approaches that treat speed or flow in isolation, our optimization simultaneously considers multiple interacting factors, enabling a more balanced trade-off between efficiency and feasibility across the road network. By complementing guideline-based planning practices (e.g., Highway Capacity Manual–style analysis) with a data-driven optimization layer, the proposed methodology offers transportation planners additional flexibility for evaluating capacity-oriented interventions. Experiments on real traffic data suggest that the approach can improve network-level efficiency and provides a practical tool for smart transportation decision support.
Download

Paper Nr:	448
Title:	Retrieval-Augmented Generation for Alzheimer’s Disease Detection: A Multi-Dataset Validation Study
Authors:	Sabrine Brahmi, Marwa Afnouch, Wided Ben Daoud and Faouzi Zarai
Abstract:	Alzheimer’s disease (AD) poses a major challenge for clinical diagnosis, highlighting the need for accurate and reliable automated decision-support tools. In this study, we introduce a retrieval-augmented generation (RAG) framework for detecting Alzheimer’s disease from structural MRI that combines deep learning with explicit medical knowledge retrieval. The proposed method integrates a three-dimensional convolutional neural network with an attention-based retrieval mechanism, allowing the model to leverage similar historical cases while maintaining strong predictive performance and interpretability. The framework was evaluated on three large-scale public neuroimaging datasets-ADNI, OASIS, and AB-MRI-and achieved state-of-the-art classification accuracies of 96.8%, 95.3%, and 94.7%, consistently outperforming existing approaches. Results across datasets demonstrate robust generalization and highlight the positive impact of incorporating knowledge retrieval into the diagnostic process. By enabling transparent, case-based reasoning, the proposed framework helps overcome key limitations of conventional black-box deep learning models and offers clinically meaningful interpretability for Alzheimer’s disease diagnosis.
Download

Paper Nr:	449
Title:	Adaptive Structured Pruning of Convolutional Neural Networks for Time Series Classification
Authors:	Javidan Abdullayev, Maxime Devanne, Cyril Meyer, Ali Ismail-Fawaz, Jonathan Weber and Germain Forestier
Abstract:	Deep learning models for Time Series Classification (TSC) have achieved strong predictive performance but their high computational and memory requirements often limit deployment on resource-constrained devices. While structured pruning can address these issues by removing redundant filters, existing methods typically rely on manually tuned hyperparameters such as pruning ratios which limit scalability and generalization across datasets. In this work, we propose Dynamic Structured Pruning (DSP), a fully automatic, structured pruning framework for convolution-based TSC models. DSP introduces an instance-wise sparsity loss during training to induce channel-level sparsity, followed by a global activation analysis to identify and prune redundant filters without needing any predefined pruning ratio. This work tackles computational bottlenecks of deep TSC models for deployment on resource-constrained devices. We validate DSP on 128 UCR datasets using two different deep state-of-the-art architectures: LITETime and InceptionTime. Our approach achieves an average compression of 58% for LITETime and 75% for InceptionTime architectures while maintaining classification accuracy. Redundancy analyses confirm that DSP produces compact and informative representations, offering a practical path for scalable and efficient deep TSC deployment.
Download

Paper Nr:	455
Title:	A Survey and Case Study on Radar Perception for Object Detection
Authors:	Rajarshi Biswas, Kshitij Kumar, Mohamed Salman and Laura Bies
Abstract:	In this paper, we look at the recent advances in the field of radar processing for perception applications. We review the underlying principles of Frequency Modulated Continuous Wave (FMCW) Radar and discuss the classical approaches for converting raw ADC data into usable representations to allow detection of targets in the scene, in both static and dynamic settings. We also examine deep neural network (DNN) approaches that operate directly on radar measurements to achieve end-to-end learning. Subsequently, we present a case study on applying modern object-detection networks on 2D image representation of radar data for detecting and classifying items inside containers. We demonstrate the feasibility of using deep learning models for radar perception on these image representations. Finally, we briefly discuss the future applications of our work and possible extensions to our current methodology.
Download

Paper Nr:	456
Title:	Domain-Adaptive Named Entity Recognition for Specialized Arabic Texts: Medical, Legal, and Financial Entity Extraction with Minimal Supervision
Authors:	Hassen Mahdhaoui, Abdelkarim Mars, Mazin S. Mohammed, Alaa Abid Muslam Abid Ali, Salah Zrigui and Mounir Zrigui
Abstract:	Named Entity Recognition (NER) in specialized Arabic domains-medical, legal, and financial-presents formidable challenges due to domain-specific terminology, scarce annotated resources, and catastrophic performance degradation of general-purpose models. While general Arabic NER systems achieve F1-scores exceeding 85%, these models experience dramatic performance drops to 40–50% F1 when applied to specialized texts containing complex domain entities such as disease names, legal references, or financial instruments. Manual annotation of domain-specific corpora remains prohibitively expensive, requiring expert annotators with specialized knowledge. This work presents a comprehensive domain-adaptive NER framework enabling efficient transfer from general Arabic to specialized domains with minimal supervision. Our contributions comprise: (1) novel domain-adaptive architecture integrating terminology-aware encoding, domain-specific adapters, and meta-learning mechanisms, (2) terminology-enhanced training leveraging domain lexicons and ontologies, (3) LLM-based synthetic data generation producing high-quality domain-specific training instances, (4) meta-learning approach enabling rapid adaptation with fewer than 500 domain examples, and (5) comprehensive multi-domain benchmark spanning medical (8,200 sentences), legal (6,500 sentences), and financial (7,100 sentences) domains. Our model achieves 73.4% macro F1 with only 500 domain examples (compared to 48.2% zero-shot baseline), reaching 85.8% with 2,000 examples and approaching supervised performance ceiling. This work establishes foundations for practical deployment of Arabic NER in critical specialized applications.
Download

Paper Nr:	459
Title:	Explainable Quantum Convolutional Neural Network for Attack Detection in Healthcare IoMT Systems Using SHAP and Grid5000 Computing
Authors:	Marwen Amara, Sami Mnasri, Marwa Amara and Thierry Val
Abstract:	The growth of the Internet of Medical Things (IoMT) improves patient care and data analysis, but it also makes healthcare systems newly vulnerable to cyberattacks.Current intrusion detection systems (IDS) struggle to work with the many different IoMT protocols. More importantly, they are ”black boxes” that cannot explain their decisions, which is a major problem in healthcare where safety is critical. This paper proposes a hybrid explainable quantum convolutional neural network (QCNN–SHAP) framework for accurate and transparent attack detection in IoMT environments. The QCNN exploits quantum feature encoding and entanglement to capture complex correlations in high-dimensional IoMT traffic data, while SHAP (SHapley Additive exPlanations) provides feature-level interpretability for each detection decision. We tested our model on the CICIoMT2024 dataset, a benchmark for multi-protocol IoMT security assessment, using the Grid5000 distributed computing platform. The proposed QCNN–SHAP model achieved a detection accuracy of 98.05%, outperforming current models like CNN, LSTM–Autoencoder, and Transformer–XAI models while maintaining explainability and moderate computational cost. SHAP analysis showed that key network features such as packet size variance, flow duration, and protocol type were the most influential in identifying attacks. The results validate the model’s robustness, interpretability, and suitability for real-time IoMT intrusion detection in healthcare contexts.
Download

Paper Nr:	460
Title:	SAC: A Framework for Measuring and Inducing Personality Traits in LLMs with Dynamic Intensity Control
Authors:	Adithya Chittem, Aishna Shrivastava, Sai Tarun Pendela, Jagat Sesh Challa and Dhruv Kumar
Abstract:	Large language models (LLMs) have rapidly gained traction across fields, with growing expectations for them to exhibit human-like personalities. To meet this expectation, various studies have explored personality mod-elling through psychometric frameworks. However, most existing approaches rely on the Big Five (OCEAN) model, which offers only broad personality dimensions and lacks mechanisms for fine-grained control of trait intensity. We address this gap by extending the Machine Personality Inventory (MPI), which originally used the Big Five model, to incorporate the 16 Personality Factor (16PF) model, allowing expressive control over sixteen distinct traits. We also developed a structured framework known as Specific Attribute Control (SAC) for evaluating and dynamically inducing trait intensity in LLMs. Our method introduces adjective-based semantic anchoring to guide trait intensity expression and leverages behavioural questions across five intensity factors: Frequency, Depth, Threshold, Effort, and Willingness. We tested this methodology on four state-of-the-art LLMs and benchmarked it against human responses, finding that modelling intensity as a continuous spectrum yields more consistent and controllable personality expression than binary trait toggling. We also find that altering a target trait’s intensity systematically shifts related traits in psychologically coherent directions, indicating that LLMs internalize multi-dimensional personality structures rather than treating traits independently. These findings open new directions for developing LLMs with adaptive, transparent, and context-sensitive personality control.
Download

Paper Nr:	464
Title:	Multi-Objective Text Detoxification with Policy Gradient Optimization and Curriculum Learning
Authors:	Amira Dhokar, Olfa Mabrouk and Lobna Hlaoua
Abstract:	Toxic language in online spaces creates significant challenges for healthy digital communication. While automated detoxification systems can rewrite offensive text, they often fail to balance three critical goals: removing toxicity, preserving meaning, and maintaining natural-sounding language. This paper introduces TRIAD, a three-stage detoxification framework combining intelligent filtering, progressive training with curriculum learning, and quality validation. Using reinforcement learning guided by a curriculum strategy, the system achieves balanced performance across all objectives (0.79 for toxicity reduction, semantic preservation, and fluency), representing a 46.3% improvement over traditional methods on the Jigsaw dataset (38,198 texts). Evaluation on ParaDetox shows competitive performance (Joint Score of 0.81), validating that progressive training with dynamic objective balancing provides an effective solution for automated text detoxification.

Paper Nr:	465
Title:	Energy-Efficient IoT Routing Procedure Using Reinforcement Learning
Authors:	Awatef BenFradj Guiloufi, Nouha Alyaoui, Nawres Bouabid, Hajer Ben Fradj and Karim Chabir
Abstract:	The Internet of Things (IoT) has transformed how physical objects interact and share data, enabling automation and intelligent decision-making many sectors such as industry, smart homes, and healthcare. As IoT systems become integral to mission-critical applications ranging from home surveillance to seismic detection and military tracking ensuring their reliability and energy efficiency has become a pressing challenge. In particular, extending the operational lifetime of devices, especially in scenarios where battery replacement is impractical, is a key research field. Reinforcement Learning (RL), a machine learning paradigm based on trial-and-error learning, offers a promising solution for dynamic, resource-constrained environments like IoT. In this paper, we propose a novel RL-based routing algorithm designed to minimize energy consumption while considering essential metrics such as device lifespan, hop count, and Q-values. Through this approach, we aim to enhance the sustainability, adaptability, and efficiency of IoT networks, addressing core challenges related to energy management and long-term system performance. Simulation results show that the proposed procedure tends to increase network lifetime by minimizing the energy consumption for each node.
Download

Paper Nr:	472
Title:	Belief Revision with Two Kinds of Trust
Authors:	Aaron Hunter and Deanna Lepke
Abstract:	In practical applications, belief revision is impacted by trust in information sources. But trust is a complex concept, which can influence belief change in different ways. In this paper we focus on two distinct forms of trust: trust based on expertise, and trust based on reliability. Each of these has been formalized independently in a belief change setting, but the interaction has not been addressed. In this paper, we introduce a simple framework that includes both forms of trust. Moreover, we demonstrate a concrete form of interaction between the two, based on counting how frequently a source provides information beyond its expertise. In addition to a formal framework, we also introduce an implemented software tool that allows us to automatically calculate the new belief state following a sequence of reports from partially trusted sources. Future directions and applications are discussed.
Download

Paper Nr:	473
Title:	I-SteganoGAN++: Fusion-Encoder and Dual-Attention Decoder for High-Capacity Generative Steganography
Authors:	Mounir Telli, Mohamed Othmani and Hela Ltifi
Abstract:	I-SteganoGAN++ is a high-capacity generative steganography framework that overcomes the fundamental trade-off between payload capacity and visual imperceptibility. Unlike traditional methods that modify an existing cover image, our approach generates the stego image end-to-end, allowing for optimal integration of the secret message during the synthesis phase. The proposed architecture features an adaptive fusion encoder and a dual-attention (spatial and channel) U-Net decoder, designed to exploit both spatial and feature dimensions. It is optimized through multi-objective adversarial training that simultaneously targets visual fidelity, message recovery, and detection resistance. A Reed–Solomon error-correction mechanism ensures message robustness. Our experiments demonstrate an effective embedding capacity of up to 4.66 bits per pixel while maintaining exceptional visual quality (PSNR > 45 dB, SSIM > 0.98). A rigorous security evaluation against modern steganalyzers (SRM, Yedroudj-Net, SRNet) shows detection error rates approaching 0.5 (random guessing). Comprehensive ablation studies validate our architectural choices, and computational analysis confirms practical feasibility.
Download

Paper Nr:	474
Title:	Detecting New Attacks under Concept Drift in Intrusion Detection Systems Using Autoencoder Networks and Gaussian Mixture Models
Authors:	Lincoln de Queiroz Vieira, Ricardo Choren and Ricardo Sant'ana
Abstract:	Intrusion Detection Systems (IDS) face the challenge of evolving cyber threats that cause concept drift, degrading the performance of static models. This paper proposes a novel IDS model that integrates contrastive autoencoders and Gaussian Mixture Models (GMMs) to detect new and variant attacks under drift conditions. The approach learns a structured latent space that enhances class separability, while GMMs model the distribution of known attacks to identify samples with low likelihood as new threats. Experiments on the CICIDS2018 dataset show that the proposed method achieves an F1-score of 0.69 for new attack detection and 0.87 for known attack classification, outperforming the CADE baseline. These results highlight the effectiveness of probabilistic modeling combined with contrastive representation learning for adaptive intrusion detection.
Download

Paper Nr:	476
Title:	RAG-KGRec++: Enhancing the Robustness and Explainability of Knowledge-Graph-Based Recommender Systems through Semantic Traceback Filtering
Authors:	Mohamed Anouer Sdiri, Hamza Gharsellaoui and Sadok Bouamama
Abstract:	Explainable recommendation systems have recently leveraged information aggregation from knowledge graphs and large language models (LLMs) to produce personalized suggestions with natural justifications. However, these approaches face several critical limitations: they are sensitive to prompt formulation, dependent on the quality of the knowledge graph (KG), and struggle to balance accuracy, robustness, and explainability. In this paper, we propose a new method called RAG-KGRec++, which introduces a novel retroactive semantic filtering mechanism. It consists of analyzing the explanations generated by the LLM to identify the graph facts actually used in reasoning, and then dynamically refining the extracted subgraphs during future queries. This feedback loop reduces noise in prompts, stabilizes recommendations, and improves their semantic consistency. Our experiments on two datasets—MovieLens 1M and Amazon Beauty—show that our approach improves the robustness of explanations while reducing the size of prompts, without compromising the quality of recommendations. We also demonstrate better resilience to errors or gaps in the initial graph.
Download

Paper Nr:	481
Title:	A Retrieval-Augmented Generation Approach for Bilingual Radiological Recommendation System
Authors:	Abir Baâzaoui and Lamis Kaddoussi
Abstract:	The surge in expensive imaging and healthcare costs has intensified examination of radiologists’ practices, specifically focusing on the frequency and necessity of their recommendations for relevant exam choice. This study presents a bilingual medical recommendation system designed to assist clinicians in selecting appropriate radiological examinations based on clinical case descriptions in French or English. The system employs a Retrieval-Augmented Generation (RAG) approach, integrating semantic search with multilingual sentence embeddings and the FAISS library for efficient retrieval of relevant clinical cases. Clinical data, extracted from structured Excel files, including case descriptions and corresponding radiological recommendations, are preprocessed and indexed into a unified vector database. Upon receiving a user query, the system retrieves the top-3 semantically similar cases using cosine similarity, providing tailored exam recommendations. The system was evaluated using precision, recall, and F1-score metrics, with the Multilingual E5 Large model achieving a precision of 85.2%, recall of 94.4%, and F1-score of 88.9% on a curated dataset. This approach enhances the accessibility of multilingual clinical knowledge, reduces diagnostic uncertainty, and supports evidence-based decision-making in radiology.
Download

Paper Nr:	482
Title:	When Senses Collide: Reasoning with Contradictory Multimodal Inputs
Authors:	Marcus Valtonen Örnhag, Katja Szybek, Tobias Widmark and Anastasia Grebenyuk
Abstract:	This paper presents a real-time, multimodal AI assistant for Augmented Reality (AR) navigation that operates efficiently with incomplete or outdated environment maps. Our proposed system embodies a conversational AI avatar, receiving inputs from the AR headset to interpret user intent, set navigation goals, and provide step-by-step guidance via a spatial path planner. We primarily focus on the discrepancy between a pre-mapped environment and live perceptual input, such as newly encountered obstacles. Our method enables the Large Language Model (LLM) to prioritize between a route computed on-device with a local path planner and real-time visual cues through prompt control. We demonstrate the system’s performance in an indoor setup, evaluating its ability to handle dynamic goal changes and navigate around unmapped obstacles. The results show a robust pipeline that combines the reasoning strengths of an LLM with the reliability of a classical planner, offering responsive and adaptive navigation even when the prior map is incomplete. Finally, we demonstrate the limitations of this approach and suggest possible mitigation strategies.
Download

Paper Nr:	488
Title:	A Hybrid Ontology Alignment Approach Using Syntactic and Linguistic Similarity with Supervised Machine Learning
Authors:	Faten Abbassi and Yousra Bendaly Hlaoui
Abstract:	As knowledge continues to grow, the same concepts are often represented differently in multiple ontologies, leading to syntactic, terminological, conceptual and semantic heterogeneity. To address this issue, we propose a novel ontology alignment approach that combines supervised machine learning models with schema-matching techniques. Our approach parses ontologies and their alignments to extract and normalize various ontological elements, thereby achieving a coherent representation that enhances the accuracy of alignments. These normalized components are processed through syntactic and external similarity techniques to construct a final similarity matrix that represents entity correspondences between two ontologies. Eight machine learning models (Logistic Regression, Random Forest Classifier, Neural Network, LinearSVC, Gradient Boosting Classifier, AdaBoost Classifier, Bagging Classifier and Recurrent Neural Network) are then used to estimate similarity degrees and generate RDF alignment file containing the corresponding entities, presented in the same format as the reference alignment files. Experimental evaluations conducted on the benchmark and conference tracks, based on reference alignments provided by the Ontology Alignment Evaluation Initiative (OAEI) competition, demonstrate that our approach achieves higher f-measure than previous work and confirms its robustness as well as its ability to generalize.
Download

Paper Nr:	490
Title:	Evaluating the Trustworthiness of LEMNA versus LIME and SHAP in Explainable Malware Detection
Authors:	Kefas Rimamnuskeb Galadima, Peter Anthony, Roderik Ploszek, Štefan Balogh, Pavol Zajac and Martin Homola
Abstract:	Deep learning algorithms have demonstrated state-of-the-art accuracy in malware detection, yet their black-box nature and their inherent lack of transparency created critical roadblocks in their usability for security operations where a transparent and trustworthy decision is essential. This paper presents a systematic evaluation of three widely used explainable artificial intelligence(XAI) techniques, LEMNA, SHAP, and LIME, for interpreting malware classifiers. Experiments are conducted on the EMBER Windows malware dataset, with cross-domain validation performed on the DREBIN Android dataset to assess explanations across platforms. We evaluate explanation quality based on fidelity, robustness, and agreement in inter-explainer feature attribution. Results show that a deep learning classifier achieves the highest detection performance on both datasets, with the lowest false positive rates and high accuracy. Among the explanation methods, LEMNA provides the highest fidelity, LIME exhibits lower fidelity but comparatively high robustness, and SHAP offers a balanced trade-off between fidelity and robustness. Analysis using Jaccard similarity reveals low agreement between explainers, indicating that each method highlights different aspects of the model’s decision process. These findings emphasise the limitations of relying on a single explainer and provide practical guidance for the selection of explainable AI techniques in high-stakes malware detection scenarios.
Download

Paper Nr:	494
Title:	A Comparative Study of Feature Identification Methods for Sparse Autoencoders in Board Game Representations
Authors:	Jonathan A. Zea, Lorena Isabel Barona and Marco E. Benalcázar
Abstract:	This paper investigates whether gradient-based methods can accelerate feature identification in Sparse Autoencoders (SAEs) trained on board game representations. We compare a traditional dataset sweeping method with Projected Gradient Descent (PGD) for identifying board configurations that maximally activate or deactivate SAE features in a 5×5 Tic-tac-toe variant. Our results demonstrate that while PGD can discover activating configurations for all features, including those considered dead in the training distribution, these configurations often lack human-interpretable meaning. Despite achieving similar activation levels, PGD requires significantly more iterations than dataset sweeping, suggesting limited practical benefits for computational efficiency. We evaluate SAE performance across multiple feature dimensions (25, 75, 100, and 512) using established interpretability metrics, finding that higher-dimensional SAEs generally capture more interpretable concepts while exhibiting diminishing returns. This work contributes to the developing field of mechanistic interpretability by providing insights into the strengths and limitations of gradient-based approaches for feature identification.
Download

Paper Nr:	499
Title:	Automatic Query-Intent Annotation: A Log-Free Agentic LLM Framework
Authors:	Zahra Taheri and Ziad Kobti
Abstract:	Short and ambiguous queries remain a major challenge for search engines, often reducing the effectiveness of information retrieval (IR) systems. Prior work shows that identifying a query’s intent, such as navigational, informational, or transactional, can significantly improve retrieval accuracy and interpretability. However, most intent-labeling methods depend on user search-log histories, which not only require extensive collection and preprocessing but also introduce substantial privacy and data provenance concerns, limiting their availability to researchers and their deployability in privacy-sensitive domains. This paper introduces a log-free, agentic annotation framework that uses a pipeline of LLM-driven agents to automatically assign intent labels to queries. The system comprises four components: a loader, an LLM-based annotator (zero-shot or few-shot), a validator enforcing structural and confidence constraints, and an evaluator that compares predictions against gold labels when available or uses an LLM-as-judge otherwise. We adopt a five-class taxonomy (navigational, factual, transactional, instrumental, and abstain) and evaluate the approach on labeled data (orcas-i-2m) as well as unlabeled benchmarks including antique, clueweb09-b, robust04, gov2, and dbpedia. Empirical results show that the few-shot variant consistently outperforms zero-shot prompting, and that the end-to-end pipeline maintains stable cross-domain performance under GPT-5-based auditing. Because the framework operates without relying on user interaction logs, it significantly reduces privacy exposure while lowering annotation cost and improving scalability. Automating query-intent annotation avoids dependence on user search logs, thereby mitigating privacy risks while reducing the time and cost of labeling large-scale datasets. The code-base to support reproducibility is available at https://github.com/ZahraTaherikhonakdar/Agentic-Annotation.
Download

Paper Nr:	502
Title:	Optimized Deep Learning Multimodal Medical Images Registration: Validation on MRI/PET
Authors:	Aymen Chaouch, Nada Haj Messaoud, Asma Ben Abdallah and Mohamed Hedi Bedoui
Abstract:	Image registration represents a challenging problem in the field of computer vision, despite its numerous applications in areas such as image fusion, stereoscopic vision, and motion analysis. Conventional methods generally perform this registration by estimating a global homography matrix between the source images, either by matching characteristic points or by direct inference. However, the inherent complexity of multimodal or morpho-functional medical images, characterized by significant variations in depth, renders the use of a single homography matrix inadequate for comprehensively characterizing geometric relationships at the pixel level. We propose a robust and innovative approach to address this limitation by estimating homography while explicitly integrating depth variations; moreover, we reformulate the mutual information estimator as a GAN-inspired discriminative objective. The proposed methodology is based on the development of an optimized convolutional neural network (CNN) architecture inspired by HomographyNet, designed to estimate geometric transformations for accurate image alignment. Clinically, this framework is particularly relevant for PET–MRI registration in patients with drug-resistant epilepsy and negative structural MRI findings, where precise multimodal alignment is critical for identifying subtle hypometabolic regions associated with epileptogenic foci. By improving morpho-functional fusion, the proposed approach has the potential to enhance pre-surgical evaluation, support more reliable localization of epileptogenic zones, and ultimately contribute to improved clinical decision-making in epilepsy management.
Download

Paper Nr:	505
Title:	Real-Time Indoor Obstacle-to-Exit Distance Measurement Using a Monocular Camera
Authors:	Adam Pirkhdrie, Hamed Talebi, Azad Shokrollahi, Erdal Akin and Arezoo Sarkheyli-Hägele
Abstract:	Effective monitoring of indoor environments is critical for safety and emergency preparedness, particularly in evacuation scenarios where blocked exits can have severe consequences. This paper introduces a practical method for real-time distance measurement between obstacles and evacuation routes using a single conventional camera. By integrating object detection with geometric triangulation and Euclidean distance estimation, the system reliably identifies and measures static objects such as chairs and doors. A reference-based calibration converts pixel coordinates into accurate real-world distances, removing the need for costly 3D sensors. Experiments show a mean deviation of approximately 7cm, demonstrating sufficient accuracy for deployment in smart building applications. The approach is cost-effective, integrable with existing IoT infrastructures, and enables rapid risk assessment, supporting enhanced safety and more effective emergency response in dynamic indoor environments.
Download

Paper Nr:	508
Title:	Adaptive Penalty-Guided VND with CP-SAT for Curriculum-Based Course Timetabling
Authors:	Tam Minh Nguyen
Abstract:	This paper presents an Adaptive Penalty-Guided Variable Neighborhood Descent (VND) framework with CP-SAT initialization for solving the ITC-2007 Track 3 problem. The approach begins with a feasible timetable generated by a constraint programming model, which guarantees satisfaction of all hard constraints. The initial solution is then refined through a VND procedure that alternates between two complementary neighborhoods: single-event swap moves and Kempe-chain based moves enhanced with a min-cost max-flow model for efficient room assignment. To prevent premature convergence, an adaptive penalty-guided perturbation mechanism introduces diversification by perturbing high-penalty events, where the perturbation strength is adjusted dynamically based on search stagnation. This adaptive design allows the algorithm to balance intensification and diversification throughout the search process. Experiments on ITC-2007 benchmark instances demonstrate that the proposed approach achieves competitive solution quality within a reasonable runtime, indicating its effectiveness for large-scale timetabling problems.
Download

Paper Nr:	509
Title:	An Attention-Residual 3D Framework for Robust Hippocampal Segmentation with K-Fold Cross-Validation
Authors:	Nesrine Jazzar, Raghda Jouirou, Basma Mabrouk and Ali Douik
Abstract:	Accurate segmentation of the hippocampus from three-dimensional magnetic resonance imaging (3D MRI) is crucial for the quantitative assessment of neurodegenerative disorders, especially Alzheimer’s disease. This study introduces a robust and reproducible framework for volumetric hippocampal segmentation. The method employs an Attention-Residual 3D network enhanced with Squeeze-and-Excitation (SE) modules to improve contextual feature representation and adaptive channel weighting. To support precise boundary delineation, a hybrid loss function integrating Dice, Binary Cross-Entropy, and boundary-based terms is utilized. In addition, K-Fold cross-validation and anatomically guided data augmentation are incorporated to reinforce model stability and generalization. Experiments on the Medical Segmentation Decathlon Hippocampus dataset show that the proposed framework attains high Dice scores and consistent performance across folds, highlighting its potential for clinical use and neuroscience research.

Paper Nr:	510
Title:	Recent Progress in Compilation-Based Approaches for Multi-Agent Path Finding
Authors:	Pavel Surynek
Abstract:	Multi-Agent pathfinding (MAPF) is a fundamental enabler of application successes of warehouse logistics. The task in the standard MAPF is to find discrete paths through which agents can navigate themselves from their starting positions to individual goal positions without collisions. The requirement to generate optimal paths with respect to various objectives such as makespan or sum of costs makes the problem computationally challenging. Two major approaches to optimal MAPF solving include dedicated search-based methods, and compilation-based methods that reduce a MAPF instance to an instance in a different formalism, for which an efficient solver exists. In this position paper, we summarize main ideas behind compilation-based solvers for MAPF using SAT and MILP formalisms. Although currently the dedicated solvers dominate in MAPF, we believe that ideas from compilation approaches are still relevant, and we will explain why we believe so.
Download

Paper Nr:	522
Title:	Fuzzy TransE: A Fuzzy Type Semantic-Based Learning for Translating Embedding Models
Authors:	Italo F. Zoppis, Sahar Shah, Sara L. Manzoni and Davide E. Ciucci
Abstract:	Translating Embeddings (TransE) is a widely adopted model for knowledge graph completion which represents relationships as vector translations in an embedding space. While it is known for its efficiency and simplicity, TransE struggles with capturing complex relational patterns, particularly in one-to-many (1-to-N), many-to-one (N-to-1), and many-to-many (N-to-N) relationships. This challenge arises from its rigid distance-based formulation. In this paper, we propose an enhancement to TransE that integrates fuzzy type constraints, which provide a soft regularization of entity embeddings based on their degree of membership in semantic categories (e.g., city, person). This extension aims to improve the model’s ability to represent intricate relationships and enhance the overall performance in knowledge graph tasks.
Download

Paper Nr:	525
Title:	Selecting Network Flow Attributes: A Proposal to Identify Internet Video Streaming Traffic Exploring Fuzzy-Based Classifiers
Authors:	Eduardo M. Monks, Giancarlo Lucca, Gabriel Rosa, Bruno M. P. Moura, Helida Santos, Adenauer C. Yamin and Renata Reiser
Abstract:	This paper addresses the challenges of Internet traffic classification arising from large data volumes, widespread encryption, and the increasing prevalence of adaptive services such as video streaming. We propose a hybrid approach for selecting network flow attributes that integrates Machine Learning techniques with Interval-Valued Fuzzy Logic (IvFL), enabling uncertainty modeling and improved interpretability. The method combines expert knowledge with data-driven inference to support the classification of video streaming traffic. Experimental evaluations using IvFL-based and fuzzy rule-based classifiers demonstrate that the proposed strategy provides reliable attribute selection and competitive classification performance under varying network conditions.
Download

Paper Nr:	526
Title:	DeepXFusion: A Hybrid Representation Learning and Explainability Framework for Trustworthy Machine Learning
Authors:	Wajih Abdallah and Majed Alruwaili
Abstract:	Deep learning methods attain high predictive performance but are generally resistant to being understood or interpreted, a hindrance to their usefulness when the situation requires reliable and interpretable reasoning. Traditional, post-hoc explainability methods, such as gradient-based attribution or perturbation method, could potentially provide insights but are always limited and unstable because they are decoupled from the model’s internal representation, and also inconsistent across perturbations. This paper presents DeepXFusion, an intrinsically explainable learning framework that can provide representation learning and interpretability in a single, differentiable architecture. The proposed approach embeds a deep encoder with an intrinsic explainer, optimized together with a Consistency-Augmented Explanation Loss (CAEL), that enforces consistency between the explanations and model behaviour while guarding against small input perturbations. The method is evaluated using extensive experiments on MNIST, Fashion-MNIST, and CIFAR-10, achieving almost 99% accuracy, high fidelity explanation (0.94 at epoch 50), and stable attribution maps with significantly less drift than previously provided by the state-of-the-art post-hoc methods. Additionally, latent-space projections display clustered representations that are compact and well separated, which demonstrated that explanation-driven training develops structured, semantic representations. We demonstrate the incorporation of interpretability via training directly into the optimization process results in stability and robustness, providing confidence with a principled approach to reliable deep learning systems.
Download

Paper Nr:	527
Title:	Failing on Bias Mitigation: A Case Study on the Challenges of Fairness in Government Data
Authors:	Hongbo Bo, Jingyu Hu, Debbie Watson and Weiru Liu
Abstract:	The potential for bias and unfairness in AI-supporting government services raises ethical and legal concerns. Using crime rate prediction with the Bristol City Council data as a case study, we examine how these issues persist. Rather than auditing real-world deployed systems or producing a generalizable benchmark, our position is that a case-based investigation to understand why bias mitigations applied to government data are not always effective. Experimental analysis reveals that the failure occurs not because of flaws in model architecture or metric selection, but due to the inherent properties of the data itself, further reinforcing that the origin of bias lies in the structure and history of government datasets. We then explore the reasons for the mitigation failures in predictive models on government data and highlight the potential sources of unfairness posed by data distribution shifts, accumulated historical bias, and delays in data release. This study provides a crucial existence proof and serves as a critical ‘early warning’ that biases in government data may persist even with standard mitigation methods.
Download

Paper Nr:	528
Title:	Simulation-Based Optimization of Demand Flexibility and Storage Capacity in Distributed Solar Energy Systems
Authors:	Marco A. Hudelist, Claudia Maussner, Erich Teppan and Elena Wiegelmann
Abstract:	The accelerating adoption of residential photovoltaic (PV) systems introduces complex optimization challenges for distributed energy management. Determining economically and ecologically justified battery capacity at the household level requires capturing the dynamic interplay between generation, consumption, and flexibility. This paper presents a simulation-based optimization framework that integrates demand-side scheduling and battery sizing for residential prosumers. Using high-resolution (15-minute) agent-oriented simulations, each household is modeled as an adaptive agent that shifts appliance operation within inferred flexibility windows derived through inverse load-profile disaggregation. Agents interact with a simulated PV–battery system incorporating round-trip efficiency, power limits, and tariff dynamics. For each storage configuration, the simulation computes import/export profiles and evaluates performance indicators including net present value, payback period, levelized cost of stored electricity, and net life-cycle CO2 impacts accounting for embodied emissions. Results show that schedule-aware agents shift the optimal battery size to a moderate range, maxi-mizing marginal gains in self-consumption and emissions reduction without over-investment. The framework provides a reproducible environment to explore the coupled decision space of load management and storage sizing, offering a data-driven tool for researchers, policymakers, and energy developers to advance both affordability and sustainability in distributed energy systems.
Download

Paper Nr:	529
Title:	Time-Aware Auto-SCS: An Enhanced Deep Learning Based Approach for POI Recommendation
Authors:	Chaima Laroussi and Raouia Ayachi
Abstract:	Point-of-Interest (POI) recommendation is a fundamental service for guiding users to relevant venues among a large number of suggestions. In light of this, academia and industry have extensively studied influential factors to improve recommendation efficiency in Location-Based Social Networks (LBSN). While existing approaches leverage spatial distance, social relationships, and POI categories, they often insufficiently consider the dynamic nature of user behavior. This paper introduces Time-Aware Auto-SCS, an enhanced deep learning-based approach for POI recommendation. It serves as a significant extension to the original Auto-SCS framework by focusing on overcoming its static limitations. The overall architecture comprises three phases, namely, feature extraction, multi-aspect fusion, and recommendation generation. The first phase includes the spatial proximity and social trust steps from the Auto-SCS model, along with a new step called time-aware categorical behavior. This step is a novel Convolutional Neural Network (CNN)-based method that processes check-in sequences. It generates for each user a dense embedding that captures their time-sensitive preferences. The multi-aspect fusion phase utilizes a deep autoencoder. It combines the three embeddings from the previous outputs to generate the recommendation task. Extensive experiments on real-world LBSN datasets demonstrate that Time-Aware Auto-SCS outperforms the state-of-the-art baselines.
Download

Paper Nr:	537
Title:	Enhanced Data-Driven Product Development via Gradient Based Optimization and Conformalized Monte Carlo Dropout Uncertainty Estimation
Authors:	Andrea Thomas Nava, Lijo Johny, Fabio Azzalini, Johannes Schneider and Arianna Casanova
Abstract:	Data-Driven Product Development (DDPD) leverages data to learn the relationship between product design specifications and resulting properties. To discover improved designs, we train a neural network on past experiments and apply Projected Gradient Descent to identify optimal input features that maximize performance. Since many products require simultaneous optimization of multiple correlated properties, our framework em-ploys joint neural networks to capture interdependencies among targets. Furthermore, we integrate uncertainty estimation via Conformalised Monte Carlo Dropout (ConfMC), a novel method combining Nested Conformal Prediction with Monte Carlo dropout to provide model-agnostic, finite-sample coverage guarantees under data exchangeability. Extensive experiments on five real-world datasets show that our method matches state-of-the-art performance while offering adaptive, non-uniform prediction intervals and eliminating the need for retraining when adjusting coverage levels.
Download

Paper Nr:	542
Title:	An Empirical Study of Architectural Trade-Offs in Vision-Based Traffic Sign Interpretation Systems
Authors:	Su Myat Noe, Ha Thanh Nguyen, May Myo Zin and Ken Satoh
Abstract:	Multi-Agent architectures are frequently proposed for complex AI tasks under the assumption that role specialization improves performance through distributed expertise. However, empirical evidence supporting these benefits remains limited, particularly in perception-to-law mapping domains such as traffic sign interpretation, where systems must accurately map visual inputs onto codified legal rules. This paper explores how different architectural design strategies, including holistic single-agent processing, structured prompting, and modular multi-agent decomposition, influence reasoning quality within this legally grounded task setting. Across 360 test instances comparing four architectures—Single Agent, Single Agent with Chain-of-Thought prompting, Single Agent with Chain-of-Instruction, and Multi-Agent with specialized perception-reasoning roles—we find that a well-designed single-agent system significantly outperforms the multi-agent approach (mean score 8.21 vs 7.74, p=0.012, Cohen’s d=0.378). The single agent achieves superior rule-mapping relevance (9.41 vs 8.41 on 10-point scale) and substantially lower failure rates (12.2% vs 23.3%), despite the multi-agent system’s marginal advantage in description completeness. Our analysis reveals that coordination overhead, error propagation through agent pipelines, and loss of contextual integration can offset theoretical benefits from specialization. These findings provide evidence-based guidance for architectural decisions in legal AI: simplicity and integration may be preferable to complexity and decomposition for tasks requiring tight perceptionreasoning coupling. We discuss implications for AI system design in compliance-critical domains and identify conditions under which multi-agent approaches might succeed.
Download

Paper Nr:	545
Title:	Residual U-Net with Spatial Feature Refinement and Channel Sensitivity Enhancement for Accurate Segmentation of Cochineal Insect Infestation
Authors:	Adel Benali, Rahma Fourati and Imen Jdey
Abstract:	The Opuntia cactus, a key crop in arid regions such as Tunisia, is increasingly threatened by the mealybug (Dactylopius opuntiae), leading to considerable economic losses. Rapid and accurate detection is therefore crucial to implement effective pest control strategies. This article presents a new segmentation architecture, ResAtt-SE-UNet, specifically designed for the automated detection of cochineal infestations on cactus cladodes. This innovative architecture synergistically integrates residual connections, spatial attention mechanisms and Squeeze-and-Excitation (SE) blocks within the U-Net framework, thus aiming to improve the representation of features, optimize the gradient flow and allow the model to dynamically focus on the most relevant spatial and channel characteristics for an accurate delineation of pest boundaries. Experiments conducted on a custom annotated dataset demonstrate that the proposed ResAtt-SE-UNet achieves peak performance, with a Dice coefficient of 82.23%, an IoU of 69.82% and an accuracy of 94.32% over the test set. These results highlight the ability of our architecture to effectively capture complex textures and varied scales of cochineal infestations, while reducing false positives and false negatives. Finally, a thorough comparison with several reference architectures, including standard U-Net, U-Net with ResNet backbones, VGG and InceptionV3, as well as U-Net++, confirms the superiority of our approach. This comparative study provides valuable insights into the effectiveness of different architectural choices for segmentation tasks in precision agriculture, and validates the contribution of attention and recalibration mechanisms in improving performance. Our work thus proposes a robust tool for the early detection of pests, opening important perspectives for crop protection in arid regions.
Download

Paper Nr:	547
Title:	A Novel Depth-First Scheduling for Spatially Dynamic Neural Networks
Authors:	Steven Colleman, Andrea Nardi-Dei, Marc C.W. Geilen, Sander Stuijk and Toon Goedemé
Abstract:	Efficient hardware execution of convolutional neural networks (CNNs) increasingly relies on methods that reduce both data movement and memory usage. Depth-first processing is a proven approach to minimize on-chip memory and off-chip bandwidth requirements. In parallel, Spatially Dynamic Neural Networks (SDyNNs) exploit runtime spatial sparsity to skip redundant pixel computations, achieving adaptive efficiency. However, existing hardware implementations for SDyNNs neglect the benefits of depth-first scheduling. This paper introduces a flexible multi-core hardware architecture that, for the first time, integrates depth-first execution with spatially dynamic pruning. We propose a novel scheduling strategy tailored to the concurrent execution of convolutional and decision layers in SDyNNs. Our approach modifies the depth-first paradigm to support dynamic pruning at the pixel level while maintaining full computational parallelism. Analytical results demonstrate that this architecture significantly reduces latency and memory demands compared to prior two-array implementations, particularly under realistic sparsity conditions.
Download

Paper Nr:	548
Title:	AUDITRON: Secure AI-Orchestrated Auditing with Privilege-Gated Data Actions
Authors:	Raja Hanafi, Amir Smati and Henda ben Ghezela
Abstract:	Auditing continues to serve as a fundamental pillar of organizational compliance and risk management; however, its implementation remains constrained by labor-intensive procedures, alongside the necessity to securely manage sensitive data. The recent advancement in incorporating artificial intelligence (AI), particularly large language models (LLMs), into audit processes offers promising enhancements in automation and adaptability; yet, it concurrently presents significant security challenges, including risks of unauthorized data access and privilege escalation. This paper presents AUDITRON, a secure AI-orchestrated auditing architecture implementing privilege-gated data actions and function-scoped control. AUDITRON layers a Model Context Protocol (MCP) and a Model Context Operator (MCO) over an LLM to enable natural auditor–AI interaction while enforcing strict privilege verification and routing all read and write operations through secure API functions. The design emphasizes intent classification, conflict checking, and evidence extraction from diverse data artifacts (e.g., CSV exports, PDF logs) without granting the LLM direct database visibility. By isolating model reasoning from execution and binding every possible action to explicit, auditable capabilities, AUDITRON accelerates audit tasks without compromising data integrity or governance guarantees. This work positions AUDITRON as a pathway toward trustworthy, scalable, and policy-compliant automation of audit supervision in sensitive environments.
Download

Paper Nr:	550
Title:	Cluster Purge Loss: Structuring Transformer Embeddings for Equivalent Mutants Detection
Authors:	Adelaide Danilov, Aria Nourbakhsh and Christoph Schommer
Abstract:	Recent pre-trained transformer models achieve superior performance in various code processing objectives. However, although effective at optimizing decision boundaries, common approaches for fine-tuning them for downstream classification tasks - distance-based methods or training an additional classification head often fail to thoroughly structure the embedding space to reflect nuanced intra-class semantic relationships. Equivalent code mutant detection is one of these tasks, where the quality of the embedding space is crucial to the performance of the models. We introduce a novel framework that integrates cross-entropy loss with a deep metric learning objective, termed Cluster Purge Loss. This objective, unlike conventional approaches, concentrates on adjusting fine-grained differences within each class, encouraging the separation of instances based on semantical equivalency to the class center using dynamically adjusted borders. Employing UniXCoder as the base model, our approach demonstrates superior performance in the domain of equivalent mutant detection and produces a more interpretable embedding space.
Download

Paper Nr:	553
Title:	Explaining Mental Health Predictions through Natural Language
Authors:	Dan Dodun-Des-perrieres and Madalina Raschip
Abstract:	Understanding and explaining mental health indicators in social media text is a critical challenge for computational psychiatry and affective computing. In this work, we present a multitask neural framework that performs both depression classification and explanation generation for social media posts, incorporating not only the textual content but also temporal features such as posting time. Our model leverages a pretrained BERT encoder for semantic representation, augmented with a Time2Vec module to encode periodic time-based patterns. A Transformer decoder, conditioned on both the predicted class label and the temporal embedding, generates natural language explanations justifying the classification decision. The model is trained and evaluated on a curated dataset of timestamped social media posts annotated with depression labels and written explanations. Results demonstrate that our approach achieves strong classification performance while producing coherent and temporally-aware rationales. Our findings highlight the value of temporally grounded explanation in mental health NLP and open new directions for transparent and context-sensitive mental health analysis.
Download

Paper Nr:	556
Title:	Physics-Constrained Machine Learning Framework for Parametric Optimization of Industrial Cooling Systems
Authors:	Md Mohsin Kabir, Shaibal Barua, Mobyen Uddin Ahmed, Behrouz Nourozi, Shahina Begum and Rebei Bel Fdhila
Abstract:	High-voltage direct current (HVDC) systems rely on extensive and complex cooling networks that incorporate both dry coolers and chillers. Designing these systems is very challenging because many parameters interact in nonlinear ways, and simulation-based optimization requires significant computational resources. This paper presents a machine learning framework incorporating physics-based constraints for the efficient optimization of parametric cooling systems. The framework includes three main stages: data exploration, modeling, and optimization. In the first stage, simulation data are analyzed to extract key parameters such as Cooling Efficiency, Power Consumption, Cooling Power, Chiller Capacity, Water Inlet Temperature, etc. The modeling stage uses machine learning models to predict a composite performance score that combines cooling effectiveness, energy efficiency, and CO2 savings. Finally, a physics-constrained optimization algorithm (L-BFGS-B: Limited-memory Broyden–Fletcher–Goldfarb–Shanno with Bounds) finds the best parameter combinations while ensuring physical consistency in energy balance, temperature limits, and COP values. The results show that the proposed AI model can accurately predict system performance and identify optimal configurations. For 20 system configurations, the model achieved an R2 of 0.93, showing strong predictive capability. Key influential factors include Fan Power, Cooling Efficiency, COP (Chiller) and Power Consumption. This approach has the potential to provide a accurate and reliable alternative to full-scale simulations.
Download

Paper Nr:	560
Title:	A Trust-Based Model for Document Credibility in Collaborative Annotation Systems
Authors:	Fatiha Naouar and Lobna Hlaoua
Abstract:	Credibility constitutes a fundamental criterion for evaluating the reliability of information and social interactions within digital social networks. In an era of massive, often unverified content, modeling trust between users is essential to distinguish reliable sources. This paper proposes a novel credibility model based on a graph representation of user relationships and the analysis of their annotations on shared documents. These interactions serve as key indicators for measuring the coherence and relevance of user contributions. The model identifies "trustworthy users", individuals whose annotations are consistently accurate and aligned with the network, who play a central role in propagating credibility. A subsequent re-ranking phase reorders documents by integrating this user-derived trust with traditional semantic relevance (TF-IDF). Our experimental evaluation demonstrates that this integrated approach significantly enhances recommendation quality. The proposed model achieved a precision of 0.55 and a recall of 0.9 for top-5 recommendations, outperforming a TF-IDF baseline (precision=0.35, recall=0.28) and confirming that balancing textual relevance with social credibility leads to a more accurate, robust, and explainable information retrieval system.
Download

Paper Nr:	561
Title:	Collaborative Spatio-Temporal Graph Reinforcement Learning for Multivariate Drought Forecasting
Authors:	Ahlem Ferchichi, Mejda Chihaoui and Haifa Harrouch
Abstract:	Drought forecasting in Africa is a complex challenge due to the multivariate nature of climate drivers and the strong spatio-temporal dependencies across regions. Traditional deep learning models, while effective in regression and classification tasks, often assume equally spaced data and struggle to capture dynamic interregional relationships in time-series prediction. Graph Neural Networks (GNNs) provide a natural way to incorporate graph structure into forecasting models, enabling the explicit modeling of spatial and temporal dependencies. In this study, we propose a Collaborative Spatio-Temporal Graph Reinforcement Learning (STGNN–RL) framework for multivariate drought forecasting in Africa. Within this framework, reinforcement learning governs adaptive policy updates, while intelligent agents embedded in a dynamic spatio-temporal graph collaborate to reason collectively and adjust their strategies in response to evolving climate regimes. This collaborative design enhances interpretability, adaptive intelligence, and forecasting robustness by explicitly modeling inter-agent relationships across African regions. Experimental results demonstrate that the proposed framework outperforms baseline models-including STGNN-only and non-collaborative RL-by achieving lower error rates, higher skill scores, and improved calibration reliability. Overall, this study introduces a novel collaborative graph reinforcement learning paradigm for multivariate drought forecasting, highlighting its potential to support climate resilience and agricultural planning across Africa.
Download

Paper Nr:	562
Title:	Variable Semantic Representation: A Replication Study and Hyper-Parameter Optimization Using Taguchi Methods
Authors:	Stefan-Octavian Custura, Radu Gaceanu and Andreea Vescan
Abstract:	As software systems are rapidly growing all around the world, identifier names are part of the fundamentals of understanding software programs, making themselves the core ingredients in program comprehension and code navigation. Moreover, current artificial intelligence-based systems use such identifiers for code completion, bug detection, and the suggestion of new variable names. However, the current literature has yet to develop a solution for analyzing the similarity of two identifiers, as two different identifier names can refer to the same component. We propose different ways to enhance the current capabilities of identifier name analysis, using different models and tuning them. This way, we notice that specialized embedding models for software engineering offer higher-quality results, as shown by existing benchmarks, and we used the Taguchi method to optimize possible results in that direction.
Download

Paper Nr:	565
Title:	DL-Based Lesion Segmentation and Novel Biomarkers for Early EDSS Prediction in Multiple Sclerosis
Authors:	Nada Haj Messaoud, Aymen Chaouch, Asma Ben Abdallah and Mohamed Hedi Bedoui
Abstract:	This study introduces an integrated pipeline for the early prediction of disability progression in multiple sclerosis (MS), targeting the low EDSS range (0–2.5). It combines multimodal data (MRI, age, clinical features) with a novel deep learning model called MSU-AttNet, which employs multi-scale dilated attention mechanisms to achieve accurate segmentation of MS lesions of varying sizes and locations. The model delivers Dice scores of 0.75 on a private dataset, 0.71 on the ISBI 2015 challenge, and 0.83 when tested on acute ischemic stroke lesions, demonstrating strong generalization. The pipeline also includes automatic brain lobe segmentation, standard MRI feature extraction, and two innovative EDSS-driven biomarkers: the EDSS-Topographic Lesion Proximity Score (ETPS), which quantifies strategic periventricular lesion burden, and the EDSS-Weighted Regional Lesion Index (ERLI), which integrates anatomically weighted lesion load adjusted for disease duration. Among several machine learning models evaluated, LightGBM, using these novel biomarkers together with conventional MRI features and clinical/demographic variables, provides the best performance with an RMSE of 0.51 and an MAE of 0.41, highlighting their key contribution to improving early disability risk assessment in MS.
Download

Paper Nr:	575
Title:	A New Tunisian Arabic Corpus and Benchmark for Automatic Speech Recognition
Authors:	Mohamed Ali Sghaier, Mohamed Lazhar Bellagha and Mounir Zrigui
Abstract:	Recent advances in Automatic Speech Recognition (ASR) for Arabic have been significant; however, Tunisian Arabic (Derja) remains both underrepresented and linguistically challenging. This paper introduces TuniSpeech-21h, a 21-hour speech corpus of Tunisian Arabic compiled from social media and broadcast materials. The corpus encompasses multiple domains, including politics, culture, and everyday interactions, and features diverse speakers, recording environments, and detailed metadata. It was developed through a transparent and reproducible pipeline for segmentation, transcription, and text normalization. To establish reliable baselines, several state-of-the-art ASR architectures are evaluated, including Wav2Vec 2.0 (XLSR/XLS-R, with and without the LinTO Tunisian n-gram language model) and Whisper (large-v2, large-v3, and Arabicadapted variants), using both TuniSpeech-21h and other publicly available Tunisian Arabic corpora. Experimental results indicate that Whisper large-v2 achieves the best overall performance (WER = 24.74 %, CER = 8.32 %, MER = 16.53 %), while Wav2Vec 2.0 + LM remains competitive for domain-specific Spoken Language Understanding (SLU) tasks. Error analysis identifies major challenges related to code-switching, morphophonemic variation, and orthographic inconsistency. We release detailed corpus specifications and evaluation protocols to support open, comparable research and the improvement of ASR systems for underrepresented languages.
Download

Paper Nr:	580
Title:	Comparative Study of Random Forest and XGBoost for Cardiovascular Risk Prediction Using SHAP Interpretability
Authors:	Raghda Jouirou, Nesrine Jazzar and Olfa Boussaid
Abstract:	Accurate prediction of cardiovascular risk is essential for preventive medicine and early intervention. This study presents a comparative analysis of Random Forest and XGBoost models using a real clinical dataset with demographic, clinical, and lifestyle features. The experiments demonstrate that XGBoost slightly outperforms Random Forest while providing robust and reliable predictions. To ensure transparency, SHapley Additive exPlanations (SHAP) are applied, revealing the most influential factors in cardiovascular risk assessment. The proposed framework combines predictive performance with interpretability, offering a practical and clinically relevant tool for supporting decision-making in healthcare.
Download

Paper Nr:	582
Title:	Learning Proportional Analogies: Lightweight Neural Network vs Large Language Models
Authors:	Stergos Afantenos, Miguel Couceiro, Emiliano Lorini and Van-Duy Ngo
Abstract:	Analogical reasoning often involves statements of the form α : β :: γ : δ, known as proportional analogies, which can be interpreted as “α differs from β as γ differs from δ” and “β differs from α as δ differs from γ”. In this paper, we study the learnability of proportional analogies from both theoretical and experimental perspectives. We show that, in the Boolean setting—where each element of a proportional analogy is represented as a Boolean vector—proportional analogies are efficiently PAC learnable. To validate this in practice, we instantiate proportional analogies in a perceptual scenario with 4-cell images, each cell containing a shape and a colour. We automatically generate a dataset of valid and invalid proportional analogies and train lightweight artificial neural networks (ANNs) as evaluators. We compare our ANN-based models against state-of-theart Large Language Models (LLMs) in proportional analogy verification (checking correctness), proportional analogy generation (producing missing elements), and proportional analogy generalisation (applying knowledge acquired during learning to unseen features). Our results show that lightweight ANNs i) match LLMs in verification and generalisation, and ii) outperform LLMs in generation, demonstrating that simple, efficient models can effectively learn and generalise proportional analogies while using far fewer resources.
Download

Paper Nr:	583
Title:	Emerging Patterns and Research Directions on Machine Learning in Software Defect Prediction: A Systematic Literature Review
Authors:	Camelia-Petrina Nadejde
Abstract:	Software defects remain a significant challenge in the development of reliable and maintainable software systems. Machine Learning (ML) techniques have emerged as effective tools for predicting software defects, enabling early identification of faulty modules and improving software quality. This paper presents a Systematic Literature Review (SLR) of ML-based Software Defect Prediction (SDP) approaches published between 2015 and 2025. The review analyzes widely used ML techniques such as Decision Trees, Support Vector Machines, and Neural Networks, as well as recent advances in ensemble and deep learning models. Datasets are evaluated, along with feature selection strategies and performance metrics. The key challenges identified include class imbalance, feature redundancy, and limited dataset availability. The study highlights trends such as the integration of explainable AI, transfer learning, and AutoML into defect prediction workflows. The findings provide insights into the current state of ML-based SDP and suggest directions for future research and industrial applications.
Download

Paper Nr:	584
Title:	An Agentic Hybrid LLM–RAG Framework for Explainable Clinical Decision Support
Authors:	Mohammed Anas Kapadia, Mohammed Maaz Memon, Pankaj Mishra and Shun Okuhara
Abstract:	The fast evolution of Large Language Models (LLMs) has provided new opportunities to intelligent Clinical Decision Support Systems (CDSS), but such issues as hallucination, absent interpretability, and poor factual foundation still exist. This paper proposes a Hybrid LLM-Retrieval Augmented Generation (RAG) model of evidence-based clinical reasoning, which would combine transformer-based contextual understanding and retrieval-based factual verification. This system is an agentic design and consists of four collaborative components, namely Retriever, Transformer Encoder, Generator, and Evaluator Agents that together guarantee accuracy, interpretability, and transparency. The suggested framework transforms the diagnostic reasoning into a probabilistic optimisation problem, and the recommendations are conditionalized by multimodal patient data and top-k evidence obtained in the biomedical literature. A composite loss is a loss that optimises diagnostic accuracy, semantic consistency and factual faithfulness. Experimental validation on benchmark datasets, such as MIMIC-III, PubMedQA and ADReSSo 2021, performs better than current models, including BioBERT, GPT-3.5 and Med-PaLM 2, with 93.7% accuracy, 0.926 AUROC, and 68% reduction in rate of hallucination. The findings prove that the Hybrid LLM-RAG model is feasible in the context of aligning linguistic fluency and clinical reliability, developing a reliable AI-based decision support in healthcare-related applications. Although the findings are promising, they are achieved in controlled experimental conditions and do not demonstrate competitive performance assertion.
Download

Paper Nr:	586
Title:	A Structural Equation Modeling Analysis of How Responsibility Attribution toward AI Influences Blame in AI-Assisted Medical Decision-Making
Authors:	Keito Miyake and Seiji Yamada
Abstract:	The rapid advancement of generative artificial intelligence (AI) has increased opportunities for collaborative decision-making between humans and AI systems, particularly in high-stakes domains such as medical diagnosis. In such contexts, understanding how responsibility attribution shapes public reactions is critical for designing socially acceptable AI governance models. This study investigated how perceived responsibility influences blame and trust in a collaborative diagnostic scenario involving a medical AI system (AI-1) and a physician (Dr. A). Participants evaluated responsibility, confidence in the physician’s decision-making, and blame directed toward the physician, AI-1, and the AI manufacturer in a misdiagnosis case. Structural equation modeling (SEM) revealed that perceived physician responsibility significantly increased blame toward the physician and decreased confidence in the physician’s decision-making. Perceived AI responsibility increased blame toward both AI-1 and its manufacturer. These findings suggest that even in AI-assisted settings, society continues to expect physicians to exercise strong supervisory judgment, while AI responsibility is extended beyond the system itself to its creators. Implications for responsibility allocation and trust formation in AI-enabled healthcare are discussed.
Download

Paper Nr:	590
Title:	Sample Policy Gradient: A Competitive Policy Optimisation Method for Off-Policy Reinforcement Learning
Authors:	Athanasios Trantas
Abstract:	Off-policy Reinforcement Learning spans a wide range of applications and offers many learning possibilities. Unlike on-policy methods, off-policy methods can learn an optimal policy while executing an exploratory policy, learn from human or agent demonstration, and learn multiple tasks in parallel. In this paper, we propose a competitive policy optimisation method for off-policy reinforcement learning, in which the action selection is derived from a deterministic policy. In particular, we extend a model-free, off-policy, actor-critic algorithm known as Sample Policy Gradient (SPG) by adding a regularising term to the objective function and by considering a prioritised buffer for storing experiences, naming it Sample Policy Gradient Regularised (SPGR). We perform extensive sets of experiments to assess the performance of the proposed method and compare it with state of the art methods. Experimental results show that SPGR outperformed the selected state-of-the-art approaches regarding the mean reward return. Code is available at: https://github.com/DjAzDeck/ SPG.
Download

Paper Nr:	591
Title:	A Multi-Architecture Evaluation for Meltdown Crisis Detection Based on Behavioral Analysis
Authors:	Marwa Masmoudi, Salma Kammoun Jarraya and Mohamed Hammami
Abstract:	This research describes and examines four complementary techniques for identifying abnormal and complex physical behaviors linked with meltdowns in autistic children. The first technique uses handcrafted spatiotem-poral skeleton features, whereas the second employs pretrained convolutional neural networks (CNNs) to extract deep spatial and temporal data. The third uses a Temporal Vision Transformer (TVT) to represent dynamic patterns, while the fourth employs an Attention Temporal Graph Convolutional Network (AT-GCN) and pretrained CNNs to capture relational relationships. The suggested architecture aims to build an intelligent Autistic Meltdown Detector (AMD) for real-time crisis detection and safety monitoring. Experiments on the Kinect-based ”MeltdownCrisis” dataset show promising results, with InceptionResNetV2 attaining up to 99.8% accuracy and AT-GCN obtaining 97.5%. These results demonstrate the efficacy of combining handcrafted and deep learning algorithms for accurate meltdown crisis detection.
Download

Paper Nr:	592
Title:	End-to-End Development of a Multimodal Fusion Model for Diabetic Retinopathy and Cardiovascular Risk Prediction
Authors:	Marwa Masmoudi, Salma Kammoun Jarraya and Mohamed Hammami
Abstract:	Diabetic Retinopathy (DR) and Cardiovascular Disease (CVD) are serious consequences of diabetes, yet current diagnostic processes address them independently. This research describes a compact end-to-end multi-modal approach for predicting DR severity and a continuous cardiovascular risk (CVR) score by combining retinal fundus pictures with structured clinical data. The system consists of an EfficientNet-B3 retinal encoder, an autoencoder-based clinical encoder, and a Transformer fusion module that monitors relationships between ocular biomarkers and systemic health indicators. To overcome class imbalance and the scarcity of high-risk data, Conditional GANs and Tabular VAEs create realistic synthetic pictures and clinical records. Experiments on the BRSET dataset demonstrate that the suggested technique outperforms unimodal and classic fusion baselines, with an AUC of 0.95 and F1 of 0.90 for DR classification and an MSE of 0.021 for CVR regression. These findings highlight the potential of multimodal learning for unified, early detection of diabetes-related ocular and cardiovascular problems.

Paper Nr:	593
Title:	Automating the NIST Phish Scale with Language Models: A Reduced Metric for Threat Analysis and Red Team Operations
Authors:	Antonio Skorin, Pedro Pinacho-Davidson and Fernando Gutierrez
Abstract:	This paper presents an automated system for applying the NIST Phish Scale reliably and at scale by leveraging large language models (LLMs) and complementary machine learning components. The original Phish Scale is a systematic and well-validated metric, but its manual application is time-consuming, requires expert judgment, and exhibits variability across evaluators. To address these limitations, we reformulate the cue-based questions of the scale into a reduced ordinal classification task that is more suitable for LLMs. Using a manually annotated dataset, we compare the direct and reduced application methods and show that the reduced metric significantly improves automated accuracy. We then evaluate the system on a diverse corpus including ham, hard ham, generic phishing, and LLM-generated spear phishing emails. The results reveal consistent patterns in cue categories and detection difficulty, highlighting the challenges posed by personalized attacks. The proposed method offers a practical tool for threat assessment workflows and supports Red Team activities aimed at improving user resilience against social engineering.
Download

Paper Nr:	594
Title:	A Multi-LLM Agent System for Modular Ontology Population: A Case Study on ADHD
Authors:	Ibrahim Traore, Yassine Belmabrouk, Imen Megdiche, Abdel-Rahman H. Tawil, Jérôme Marquet-Doléac and Lotfi Chaari
Abstract:	Modular ontologies play a crucial role in structuring and interpreting complex knowledge, where they enable the representation of contextual information from diverse sources. To fully exploit these ontological models, a population phase is essential, involving the enrichment of the ontology with concrete instances. However, the automatic population of such ontologies from heterogeneous sources remains a significant challenge. This study proposes a multi-agent approach for the automatic population of modular ontologies. Each agent is an LLM guided by precise instructions and enhanced with appropriate tools, dedicated to one or more ontological modules. Our methodology combines advanced prompt engineering techniques with strategies for hallucination mitigation to ensure extractions that are both accurate and compliant with the ontology’s specifications. Our experimental setup involves extracting RDF triples from diagnostic PDF reports and activity schedules stored in a MongoDB database, aimed at the automatic population of an ADHD ontology. The results show that our multi-agent approach consistently outperforms mono-agent methods relying on a single LLM, regardless of the configuration used. We observe significant gains in precision, recall, and F1-score as well as a notable reduction in errors and hallucinations.
Download

Paper Nr:	595
Title:	Improving Scalability of the Multicriteria Vehicle Routing Problem
Authors:	Temirlan Kurbanov and Jiří Vokřínek
Abstract:	Multicriteria Vehicle Routing Problems (VRPs) remain notoriously difficult due to the exponential growth of non-dominated route alternatives on large road networks. In the standard approach, the multicriteria VRP is modeled on a complete graph over the depot and customers, where each edge represents a full Pareto set of trade-off paths. As network size and customer count increase, the number of parallel edges expands dramatically, leading to a severe scalability bottleneck. In this work, we introduce the Steiner VRP, which departs from the complete-graph paradigm and formulates the multicriteria VRP on a compressed road network constructed using Hierarchical k-Path Covers (kPCs). This representation preserves relevant multicriteria trade-offs while substantially reducing the size of the planning graph. To study the impact of this representation, we develop an integer linear programming model for the multicriteria Steiner Distance-Constrained VRP and compare it to a classic ILP formulated on the complete graph. Experiments on synthetically generated, city-like directed networks with 100 to 250 vertices show that the Steiner VRP significantly mitigates graph growth, reducing the number of planning edges by up to a factor of 15 and achieving runtime speedups of up to a factor of 8 on harder instances, while maintaining comparable solution quality in terms of Pareto-front coverage and diversity as measured by hypervolume. These results suggest that the Steiner VRP provides a promising foundation for scalable multicriteria routing and can be combined with ILP-based or metaheuristic solution methods in future work.
Download

Paper Nr:	601
Title:	Improving Semantic Segmentation Accuracy with Stable Diffusion-Based Augmentation
Authors:	Dániel Mezei and Csanád Sándor
Abstract:	This paper explores enhancing semantic segmentation performance through diffusion-based synthetic data augmentation. A DeepLabv3 model is first trained on the KITTI-360 dataset to establish a baseline accuracy. Next, we employ ControlNet-a diffusion model conditioned on semantic masks-to generate synthetic images that preserve spatial structure while introducing controlled variations in color, texture, and environmental conditions. The augmented dataset is then used to retrain the segmentation model, and performance is evaluated under different synthetic-to-real data ratios. Experimental results show that incorporating a moderate proportion of synthetic images improves mean Intersection-over-Union compared to training solely on real data. Moreover, targeted generation of environmental conditions such as autumn, night, and winter substantially reduces the domain gap while maintaining baseline accuracy. These findings demonstrate that diffusion-based, structure-preserving augmentation can effectively enhance segmentation robustness and generalization without additional manual annotation. Code is available at https://github.com/dani-mezei/icaart-controlnet.
Download

Paper Nr:	604
Title:	Human-LLM Collaboration for Reliable Prompt Monitoring and Benchmarking of Domain-Specific Language Assistants
Authors:	Daniel Pawlowicz, Mirjam Wittich and Claudia Dukino
Abstract:	This paper proposes a hybrid evaluation framework for domain-specific language assistants, combining prompt-triggered automated benchmarking with targeted human-in-the-loop validation. Classic metrics such as BLEU fail to capture the complexity of structured entity extraction and semantic variability required in operational workflows. By integrating prompt monitoring, the system ensures that benchmarking is triggered only by meaningful changes, maintaining alignment with evolving prompt structures. Semantic evaluation leverages Large Language Models (LLMs) to match entity fields across domains, with comprehensive experiments confirming robust performance for translations and typographical variants. Weaknesses remain in distinguishing semantically close synonyms and certain typo types, highlighting the need for human oversight. The results demonstrate that Human-LLM collaboration enables accurate, context-sensitive monitoring and evaluation, providing a reproducible and scalable quality assurance strategy for real-world language assistant deployments.
Download

Paper Nr:	607
Title:	A Rank Aggregation-Based Framework for Finding Optimal k-Plexes in Large Graphs
Authors:	Hajer Raddaoui, Haythem Mahfoudh, Mourad Kmimech and Mohamed Othmani
Abstract:	Community detection is a core problem in network science, aimed at uncovering cohesive groups of nodes that capture the structural and functional organization of complex systems. Traditional clique-based definitions of community, while conceptually simple, are often too rigid to accommodate noise, missing edges, or partial connectivity observed in real-world networks. Ideal communities have often been modeled as maximal cliques; however, since complete pairwise connectivity is rare in empirical data, such models are overly restrictive. This limitation has motivated the adoption of more flexible cohesive subgraphs, notably the k-plex, defined as a subgraph of size n in which each vertex is adjacent to at least (n −k) others. The k-plex model allows limited relaxation of connectivity constraints, but selecting an appropriate k remains challenging, as certain k values may yield maximal k-plexes that fail to represent meaningful communities. In this spirit, we propose an enhanced computational framework for discovering maximal k-plexes and a novel rank aggregation approach to automatically determine the most representative k value. Our approach incorporates an enhanced CPLex enumeration algorithm alongside social choice–based rank fusion techniques (Borda, Plurality, and Veto) to integrate multiple community quality metrics, thereby achieving an effective balance between structural robustness and interpretability. Extensive experiments on synthetic and real-world networks demonstrate that this hybrid strategy produces communities that are both cohesive and semantically relevant, while achieving higher modularity, improved stability, and reduced runtime compared to traditional community detection algorithms. The proposed work bridges graph mining and social choice theory, offering a robust, interpretable, and efficient mechanism for analyzing complex network structures.
Download

Paper Nr:	610
Title:	SpiTranNet-LIF: A Spiking Neural Network–Transformer Framework for Efficient Motor Imagery Decoding
Authors:	Maryam Khoshkhooy Titkanlou, Alireza Hashemi and Roman Mouček
Abstract:	Classifying motor imagery (MI) is a crucial task in biomedical signal analysis and brain-computer interface (BCI). Although traditional deep learning models have demonstrated potential in these areas, they frequently have issues with energy efficiency and temporal sparsity, particularly in embedded or real-time applications. In this work,we introduce SpiTranNet-LIF, a novel architecture that integrates Spiking Neural Networks (SNNs) with Transformers using Spiking Multi-Head Attention (SMHA) and adaptive LIF neurons, enabling energy-efficient temporal processing while retaining global contextual modeling.This integration preserves global contextual modeling capabilities while allowing energy-efficient computations and biologically realistic temporal processing. The model is evaluated on the BCI Competition IV 2a benchmark EEG dataset for binary MI classification (right-hand/left-hand). Experimental results on this dataset show that SpiTranNet-LIF achieves comparable accuracy and robust temporal-spatial decoding compared to conventional machine learning and deep learning approaches, and its time and memory-efficient inference makes it well suited for deployment in real-world BCI systems.
Download

Paper Nr:	611
Title:	Smart Cuckoo Search Algorithm for Community Detection in Social Networks
Authors:	Randa Boukabene, Fatima Benbouzid-Si Tayeb and Narimène Dakiche
Abstract:	Detecting communities in social networks remains a challenging task, primarily due to the complex interplay of structural and semantic factors governing interactions between individuals. To address this challenge, this paper introduces a three-stage community detection approach based on the Smart Cuckoo Search Algorithm (SCSA). In the first stage, node embeddings are leveraged to learn informative representations that capture essential structural dependencies and neighborhood patterns within the graph. These representations are subsequently refined in the second stage through a Graph Convolutional Network (GCN) trained under a contrastive learning framework, producing discriminative node embeddings that improve community separability. In the final stage, the Cuckoo Search Algorithm (CSA) exploits these high-quality embeddings to efficiently uncover the underlying community structures. Extensive experiments conducted on both synthetic and real-world datasets demonstrate that the proposed framework consistently achieves high accuracy and outperforms several state-of-the-art community detection methods.
Download

Paper Nr:	612
Title:	Statistical Analysis and Predictive Modelling for Early Detection of Diabetic Retinopathy Using Retinal Image Features
Authors:	Adel Ali Alkhaibari and Ammar Mohammed Ammar
Abstract:	Diabetic Retinopathy (DR) remains one of the leading causes of preventable blindness worldwide, affecting approximately 35% of diabetic patients. Early detection through statistical analysis of retinal images can significantly reduce vision loss. This research applies statistical methods and probability-based models to analyse retinal fundus images for automated DR detection. Using the APTOS 2019 Blindness Detection dataset containing 3,662 retinal images, we performed comprehensive statistical preprocessing, correlation analysis, and regression modelling to identify significant features associated with DR severity. Our methodology employs outlier detection using z-score analysis, feature extraction through histogram-based intensity distribution, and logistic regression for classification. Statistical hypothesis testing revealed that microaneurysm density (p < 0.001), exudate area ratio (p < 0.003), and haemorrhage count (p < 0.002) are strongly correlated with DR progression. The proposed model achieved 87.3% accuracy with sensitivity of 85.6% and specificity of 89.1% in detecting referable DR cases. Results demonstrate that statistical preprocessing improved classification accuracy by 12.4% compared to raw feature analysis. This study contributes to accessible and cost-effective screening solutions for diabetic patients in resource-limited settings, potentially preventing vision loss through timely intervention.
Download

Paper Nr:	614
Title:	Context-Aware AI-Based Framework for Sports Analytics
Authors:	Konstandinos Aiwansedo, Brahim Daoud and Hamza Ariouich
Abstract:	Artificial Intelligence (AI) is transforming industries worldwide, offering unprecedented insights and efficiency. In sports analytics, it has emerged as a crucial driver of innovation, empowering coaches, athletes, and researchers to uncover key performance determinants. Despite its growing influence, current AI tools in sports remain fragmented, highly domain-specific, and challenging to adapt across diverse disciplines or contexts. Addressing this limitation, this study introduces AISport, a modular AI-based framework designed to process heterogeneous, athlete-related data across various fields of sport. AISport proposes three core modules, context-aware data processing, adaptive AI modeling, and a reusable knowledge-driven module, forming an interoperable framework that generalizes across diverse sports contexts. By providing a adaptive and generalizable approach, AISport has the potential to address existing limitations in sports analytics, offering conceptual guidance for integrating AI with sports performance analysis across diverse contexts.
Download

Paper Nr:	622
Title:	A Variational Text-to-Motion Model for Emotion-Aware 3D Human Movement
Authors:	Ciprian Paduraru, Monica Girbea and Alin Stefanescu
Abstract:	Realistic human motion is central to immersive digital experiences. Traditional animation and motion-capture pipelines remain costly when nuanced emotions must be portrayed. This work investigates emotion-conditioned human motion generation driven by short text descriptions and trained on three-dimensional pose sequences. A curated dataset of monocular video clips is introduced, recorded with volunteer actors and annotated with free-form prompts covering six basic emotions. Pose-estimation methods convert raw footage into temporally consistent skeleton sequences, which serve as supervision for a text-conditioned variational autoencoder architecture adapted to emotion-aware motion. The model is evaluated with established motion–language metrics and a lightweight diffusion-style baseline, and further assessed in a user study that measures perceived realism and emotional alignment. Results indicate that expressive emotion-aware motion priors can be learned from datasets of modest size, and that text conditioning improves controllability over purely motion-based generative models. An analysis of failure cases highlights the limitations of sparse skeleton representations and current metric proxies for affect, and suggests design principles for future emotion-centric motion generation systems.
Download

Paper Nr:	623
Title:	Prefix-Aware Deep Reinforcement Learning for Resource Allocation in Business Processes
Authors:	Jeroen Middelhuis, Zaharah Bukhsh, Ivo Adan and Remco Dijkman
Abstract:	Optimizing business processes is crucial for improving operational efficiency. In a business process, cases arrive that require several activities to be executed in order to complete them. In process mining, methods have shown that the prefix of a case can be leveraged to predict future process states. A prefix is the historical execution trajectory of an ongoing case. However, these prefixes are not utilized by existing business process optimization approaches, limiting the available state information and thereby reducing the effectiveness of the learned resource allocation policies. This paper addresses this gap by combining the predictive power of prefixes with the optimization capabilities of Deep Reinforcement Learning (DRL) to learn resource allocation policies that minimize the mean cycle time of cases. We evaluate our approach on four synthetic processes based on real-world settings where prefix information is relevant for decision-making. The results demonstrate that including prefix information in the state space enables the DRL agent to learn policies that outperform a DRL agent without prefix information by an average of 19.5%. These findings underscore the importance of incorporating case-specific information into resource allocation methods, setting the stage for future research and real-world applications of prefix-aware process optimization.
Download

Paper Nr:	624
Title:	Near-Lossless Motor Imagery EEG Compression Using Hybrid Discrete Cosine Transform and Convolutional Autoencoder
Authors:	Duc Thien Pham, Josef Kohout, Eyyüb Tayşi and Roman Mouček
Abstract:	Efficient compression of motor imagery (MI) Electroencephalography (EEG) signals is critical for enabling scalable brain computer interface (BCI) systems, where data are often collected at high sampling rates across multiple channels. This challenge is particularly evident in binary paradigms (resting and hand movement) and multiclass paradigms (left-hand movement, right-hand movement, and resting), which generate large volumes of data requiring both compact representation and high-fidelity reconstruction. In this study, we introduce a hybrid near-lossless compression method that integrates the Discrete Cosine Transform (DCT) and a Con-volutional Autoencoder (CAE) with soft thresholding is applied on the latent space to improve compression efficiency for MI datasets. The framework applies an user-defined threshold to select significant residuals, which are further compressed using the Lempel–Ziv–Markov chain Algorithm (LZMA). Experimental evaluation shows that the proposed method achieves a compression ratio (CR) of 6.08 with 0.9986 correlation in the binary case, and a CR of 6.18 with 0.9989 correlation in the multiclass case. Overall, the proposed method offers an effective balance between compression efficiency and reconstruction accuracy, ensuring preservation of MI-relevant neural patterns for both classification tasks and BCI deployment.
Download

Paper Nr:	627
Title:	CopperNet: Hybrid Deep Learning and LLM-Based Optimization for Copper Refining Processes
Authors:	Gourav Sarkar, Hindesh Akash, Dinesh Kumar Yadav, Rohit Padakanti and Aman De Sarker
Abstract:	Copper electrorefining is a vital metallurgical process demanding precise electrolyte composition control for high-quality cathode production. The traditional process involves feeding anodes with various reagents and adjusting parameters to yield an LME-grade cathode. However, a significant drawback is the multi-day delay in obtaining data on the produced cathode quality. This delay increases the risk of low-quality production and limits timely information on factors influencing cathode quality. To overcome this, we introduce CopperNet: Hybrid Deep Learning and LLM-Based Optimization for Copper Refining Processes. CopperNet predicts cathode impurities based on anode and process parameter data, enabling swift decision-making. It utilizes LLM-based notifications to prompt timely adjustments to process parameters, aiming to secure the best quality cathode. Our results demonstrate that CopperNet significantly outperforms traditional machine learning baselines, achieving an R2 of 0.91, and provides transparent, human-readable decision support for refinery operators.
Download

Paper Nr:	629
Title:	Enterprise-Ready Web Automation: A Framework for Democratizing the AI Agent
Authors:	Gaurav Adke and Ameya Divekar
Abstract:	Autonomous Artificial Intelligence (AI) agents are finding enormous interest from both industrial and academic developers. One such vital area of application of AI agents is in automated web and computer use. These automations were traditionally done by the Robotic Process Automation (RPA), which are rule based and static in nature. AI agents offer advantages over RPA by performing Web and Computer use automations using natural language instructions as well as they can dynamically adapt to changing interfaces using strong reasoning intelligence embedded in underlying AI models. In this paper, we present a study which can democratize the Autonomous Web agent within the corporate environment. This will help various departments to improve efficiency by independently fulfilling repetitive, monotonous and complex web navigation operations using basic Python proficiency. We have initially developed an autonomous web agent and then built user interaction layer (RPA Agent) on top of it for its general use. Together, the agent receives a high-level instruction and autonomously carries out the task, effectively imitating human-like actions. Various building blocks of this RPA agent are discussed in detail. To demonstrate its usability, RPA agent is used to solve the problem statement by the marketing department. Compared to conventional RPA development time, the agent has provided required market insights in significantly less timespan.
Download

Paper Nr:	632
Title:	COVID-19 Propagation Model Based on a Network of Cellular Automata
Authors:	Beatriz da Silva Vieira and Luiz Gustavo Almeida Martins
Abstract:	This study presents a hybrid framework to model the spatio-temporal spread of COVID-19 in municipalities of Minas Gerais during 2020, combining daily case prediction using CatBoost with spatial simulation via cellular automata (CAs). The predictive model captured short-term fluctuations, weekly patterns, and medium-term epidemic trends, achieving R 2 = 0.61 on the test set and R 2 = 0.8149 overall. When integrated into the CA simulation, the model reproduced inter-municipal diffusion and spatial heterogeneity, with 7-day moving average R 2 = 0.857, and municipal-level performance ranging from R 2 = 0.25 to 0.90. The approach incorporates epidemiological, demographic, mobility, and seasonal factors, applying feature selection and hyperparameter tuning to estimate new infections driving the CA dynamics. Limitations include static mobility parameters, absence of vaccination and reinfection effects, and challenges in capturing extreme peaks. Future work may extend the temporal window, include adaptive epidemiological parameters, vaccination scenarios, and local interventions to enhance predictive accuracy. The framework provides a flexible tool to support regional public health planning and decision-making during epidemic events.
Download

Paper Nr:	633
Title:	Generative Transformer-Based Cross-Modal EEG-HRV Fusion for MCI Detection
Authors:	Amal Boudaya, Siwar Chaabene, Bassem Bouaziz and Lotfi Chaari
Abstract:	Given the increasing trend of multimodal medical data, more sophisticated fusion models are demanded for optimal anomaly detection in complex biological processes such as Alzheimer’s disease (AD). Early detection of MCI, a stage prior to AD, is important for timely intervention. This research uses the electroencephalog-raphy (EEG) and heart rate variability (HRV) data to diagnose MCI during the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) protocol. Compared with the conventional early fusion methods, we propose a cross-modal attention fusion mechanism, which can well integrate EEG and HRV features, considering key intermodal interactions. In order to enhance the model’s generalization, we fuse a generative model with a variational autoencoder (VAE) for data augmentation. We use the cross-modal transformer (CMT) model with its self-attention, which splits subjects into MCI and healthy control (HC). The effectiveness of the proposed framework is evaluated in terms of accuracy, precision, recall, and robustness under subject-wise cross-validation, achieving 92.18% accuracy, 92.42% precision, and 91.90% recall for early MCI detection.
Download

Paper Nr:	636
Title:	A Hybrid Approach for Opinion Leader Detection Integrating Semantic, Behavioral, and Topological Features in Online Social Networks
Authors:	Amira Foughali, Lobna Hlaoua and Mohamed Nazih Omri
Abstract:	The rapid growth of social media has created influential online ecosystems in which a small number of opinion leaders significantly shape public opinion and consumer behaviour. Identifying these actors within large scale and noisy networks remains challenging. This paper proposes a hybrid opinion leader detection framework that combines content analysis with structural graph modelling. Opinion leadership is characterised using four dimensions, influence, expertise, activity, and novelty, which are aggregated into a composite score to guide the HITS algorithm. Experiments on two RepLab domains, banking and automotive, show that the proposed approach consistently outperforms structural and content based baselines, achieving relative improvements of 68% and 24% over HITS, respectively, with higher precision and robustness across domains.
Download

Paper Nr:	637
Title:	Multi-Expert Topic Modeling for Code-Switched Dialectal Arabic: A Dynamic Routing Approach to Tunisian Social Media Analysis
Authors:	Samia Ben Ismail and Abdelkarim Mars
Abstract:	Topic modeling over dialectal Arabic social media remains brittle due to pervasive code-switching, non-standard spelling, and dialect-specific morphology. We introduce MDTIS, a mixture-of-experts topic modeling framework that routes each document to a small set of specialized processing modules according to its linguistic profile. MDTIS combines (i) multilingual representations for Arabic–French mixing, (ii) phonetic-aware normalization for spelling variants, (iii) morphology-sensitive features, (iv) sentiment-aware filtering, and (v) cross-dialect adaptation through a lightweight fusion module. On TunSocial-2024 (487K posts), MDTIS substantially improves topic quality over BERTopic-based baselines (e.g., +0.15 NPMI coherence) while reducing inference cost via selective expert activation. Gains are strongest on linguistically difficult subsets such as code-switched and highly variable texts. Ablations confirm that routing and phonetic/code-switch experts contribute most to coherence. We discuss remaining limitations on extreme mixing and emerging slang and release the dataset and implementation to support reproducible dialectal topic modeling.
Download

Paper Nr:	641
Title:	DEPC:Derivative Based EEG Preprocessing Framework for Epilepsy Monitoring Classification Using Deep Learning Models
Authors:	Ines Bouzouita, Zayneb Brari and Safya Belghith
Abstract:	Electroencephalogram (EEG) signals are crucial for epilepsy monitoring; however, their analysis is complicated by inherent noise and variability, making reliable seizure detection complex. The key contribution of this work lies in the introduction of a derivative-based signal preprocessing approach applied to EEG signals in combination with deep learning methods. By computing temporal derivatives, this method emphasizes fluctuations associated with epileptic activity that are often diluted in raw signals. We also conducted a comparative analysis of the derivative approach and a wavelet-based method under deep learning paradigms. Importantly, the proposed method is computationally efficient with linear time and memory complexity, enabling real-time or large-scale applications. Using the Bonn and Bern Barcelona datasets, the integrated methods achieved accuracy rates up to 99%, confirming our hypothesis. In particular, CNNs with derivative preprocessing outperform other models because derivatives enhance edge features that CNNs are designed to detect.
Download

Paper Nr:	647
Title:	An Adaptive Differential Evolution with Boundary-Guided Mutation for Multi-Threshold Breast Cancer Image Segmentation
Authors:	Mohamed Slim Kassis, Olfa Fakhfakh and Ghaith Manita
Abstract:	This paper proposes an Adaptive Differential Evolution with Boundary-Guided Mutation (A-BGMDE) algorithm for multilevel breast image segmentation driven by Kapur’s entropy. The method extends standard Differential Evolution through two coupled mechanisms: (i) success–history-based adaptation of the scaling factor and crossover rate to balance exploration and exploitation without manual tuning, and (ii) a boundary-guided mutation operator that leverages gradient-derived edge information to bias candidate thresholds toward boundary-rich intensity levels. Experiments on the DMR-IR breast thermography dataset across multiple threshold levels compare A-BGMDE with canonical DE and ten recent metaheuristic optimizers. Results over fitness, PSNR, and SSIM indicate that A-BGMDE achieves the best or near-best performance across images and threshold settings, while Friedman-test ranking further supports its consistent superiority.
Download

Paper Nr:	648
Title:	BEFT vs. LoRA: Parameter-Efficient Fine-Tuning for Auto-Optimized Vision Transformers in Image Classification
Authors:	Moez Hamdi, Sonia Bouzidi and Imen Jdey
Abstract:	In multi-robot systems, reliable depth-based classification directly influences object recognition, environmental awareness, and coordinated decision-making-making efficient and accurate models essential for robust autonomous behavior. Vision Transformers (ViTs) have demonstrated remarkable performance across various computer vision tasks, including scene classification. Fine-tuning these models is computationally costly. However, Parameter-Efficient Fine-Tuning (PEFT) methods address this limitation by fine-tuning a minimal number of parameters. Despite their efficiency, the success of these methods depends greatly on the hyper-parameter selection. In this paper, we propose an automated hyperparameter optimization framework for Bias-Efficient Fine-Tuning (BEFT) using Optuna and conduct comprehensive comparisons with Low-Rank Adaptation (LoRA). We evaluate our Auto-Optimized PEFT strategy on three state-of-the-art ViT architectures: ViT-Base, ViT-Medium, and Swin Transformer for scene classification on the NYU Depth V2 dataset. Our optimized approach achieves exceptional performance with ViT-Base reaching 99.96% accuracy, ViT-Medium achieving 99.62% accuracy, and Swin Transformer attaining 94.13% accuracy. The results demon-strate that PEFT methods combined with automated hyperparameter tuning can effectively adapt large vision models for scene classification tasks while maintaining computational efficiency and parameter economy.
Download

Paper Nr:	649
Title:	A Scenario Generation Framework for Targeting Emotional Vulnerabilities in Virtual Reality First Responder Training
Authors:	Jeanne Parisse, Domitile Lourdeaux and Mathieu Chollet
Abstract:	This paper proposes a novel approach for the dynamic generation of training scenarios in Virtual Reality (VR), in the context of first-responders training. Our proposed model leverages an ontology, Large Language Models (LLMs) and Fuzzy Cognitive Maps (FCM) in order to produce scenarios meant to train first responders emotion regulation skills to be better prepared for complex, high-risk situations. These scenarios take into account users’ personal profiles, containing notably individual emotional vulnerabilities, i.e. which situational elements are susceptible to stress them more than others, in order to produce personalized, effective training experiences. The Fuzzy Cognitive Maps allow the system to model a degree of uncertainty when choosing story events, with probabilities for choosing the next scenario event being influenced by several factors, mainly the user’s profile and the virtual world state. Our evaluation shows that for randomized initial situations and user profiles, our system generates varied story events, offering diversity in generated scenarios.
Download

Paper Nr:	651
Title:	Dynamic Weighted Multimodal Fusion for Explainable Chest X-Ray Segmentation Using Integrated Gradients
Authors:	Ghazala Hcini and Imen Jdey
Abstract:	While multimodal analysis is becoming more popular in medical imaging, there are still two main obstacles to overcome. First, creating models that successfully combine heterogeneous data such as visual features from X-rays and structured patient metadata; and second, clinical systems must be explainable to foster user trust. Clinical adoption requires addressing both transparent decision-making and precise fusion. To this end, we propose a multimodal segmentation system that integrates chest X-rays with patient metadata using a Dynamic Weighted Concatenation Fusion module. The architecture combines a Transformer-based Metadata Expert with a Lightweight U-Net Image Expert, where image feature maps are dynamically weighted by a channel-wise attention gate derived from metadata. These gated features are concatenated with the original maps to selectively enhance or suppress visual information based on clinical context. Multimodal contributions are interpreted using pixel-level heatmaps and Integrated Gradients (IG) to provide explainable outputs. Experimental results on the Chest X-Ray dataset show a Dice Score of 90.93%, Validation IoU of 84.50%, and Specificity of 97.40%, while testing on a Brain Tumor dataset demonstrates robustness with 99.41% Accuracy, 80.65% Dice Score, and 70.01% Validation IoU. IG analysis confirms effective, interpretable use of both modalities. This approach addresses key clinical challenges and provides a foundation for accurate, transparent, and context-aware multimodal segmentation.
Download

Paper Nr:	657
Title:	Ordering Queries in Constrained Partial CP-nets
Authors:	Sultan Ahmed and Malek Mouhoub
Abstract:	In multi-attribute preference-based reasoning, the CP-net is an intuitive graphical model to represent user’s conditional ceteris paribus (all else being equal) preference statements in a compact form. The partial CP-net is an extension of the CP-net allowing a partial order over some of the variable’s values. In this paper, we address the problem of ordering queries in a partial CP-net, i.e., if an outcome is consistently orderable over another outcome. We propose a linear time method that can answer at least one of the ordering queries between two outcomes with respect to an acyclic partial CP-net. Then, we show that using only ordering queries, we can construct a consistent preference order over the outcomes induced by the partial CP-net. Finally, we apply the ordering queries in preference-based constrained optimization problems. In fact, we extend the current search algorithms for finding the Pareto optimal outcomes in constrained optimization with partial CP-nets. In our proposed method, ordering queries are considered instead of dominance testing. Our method is efficient to obtain the Pareto solutions; however, the method does not always return the entire Pareto set.
Download

Paper Nr:	659
Title:	Self-Supervised and Contrastive Learning for Audio-Based Traffic Congestion Classification in Smart Cities
Authors:	Mouna Zouari Mehdi, Youssef Salhi, Jihen Frikha Elleuch and Dorra Sellami
Abstract:	Traffic congestion is one of the most pressing challenges faced by modern cities, causing longer travel times, wasted fuel, and higher levels of pollution. Traditional camera-based monitoring systems, while common, often come with high costs, privacy issues, and limitations under poor lighting or weather conditions. In this work, we explore a different approach: using sound to monitor traffic. However, what is crucial in audio spectral analysis is the judicious selection of temporal windowing for an efficient spectral feature discernment. Prolonged windows invariably conflate distinct frequencies, thereby diminishing temporal acuity, whereas truncated windows inherently preclude comprehensive spectral representation. Consequently, an adaptive windowing strategy is mandatory for aligning with intrinsic signal modulations and content. This paper presents an optimal time-partitioning strategy for audio samples from traffic records, specifically designed to align with congestion contexts for robust classification. Accordingly, we evaluate several deep learning models, including residual convolutional neural networks (CNNs) and contrastive self-supervised learning methods. Our findings show that an adequate sample elaboration with contrastive pretraining, followed by fine-tuned classification, delivers outstanding results outperforming both traditional MFCC+SVM approaches and fully supervised CNNs. These results highlight the potential of audio-based analysis as a practical, low-cost, and scalable solution for real-time traffic congestion monitoring in smart cities.
Download

Paper Nr:	660
Title:	Can Generative Adversarial Networks Compete against Diffusion Models at Generating Time Series?
Authors:	Philipp Engler, Ludger van Elst, Sheraz Ahmed and Andreas Dengel
Abstract:	Diffusion models have become the state-of-the-art at synthetic data generation, widely replacing generative adversarial networks (GANs) in recent years. GANs often suffer from issues like training instabilities and mode collapse. However, they also offer advantages. They are computationally efficient, as they only require a single forward pass for generating a synthetic data sample. Thus, we ask the question whether GANs can still compete with diffusion models. Specifically, we tackle the task of generating conditional time series-one of the most common data modalities in industrial applications. Making a model follow conditions appropriately is not always trivial. In this work, we leverage a structured noise space (SNS) to implicitly condition GANs for generating time series data. The SNS allows to use an unconditional generator architecture for conditional data generation by training it on a certain latent space structure. We extensively evaluate the method on time series datasets from the UCR classification archive, comparing to other GAN, variational autoencoder and diffusion approaches. We find advantages over explicit conditioning regarding fidelity of generated samples and training stability. With our improved architecture, we show that SNS-GAN is able to compete with diffusion at a fraction of the sampling time. Additionally, we examine limitations and propose alternatives of structuring the noise space so that the model can effectively learn the condition in varying scenarios.
Download

Paper Nr:	661
Title:	Layered Rule-Based Approach: An Advanced Requirement Processing Producing an Unified Requirements According to Proposed Template
Authors:	Mariem Abdouli and Wahiba Ben Abdessalem Karaa
Abstract:	Effective requirements engineering is critical for the success of software development projects. In this work, we propose a Layered Rule-Based Approach for advanced requirement processing, designed to generate unified requirements by applying a set of rules and predefined template. Our method uses the technique that is known for its precision, we applies syntactic, semantic, linguistic and heuristic rules organized in multiple layers, allowing us to gradually improve and change informal requirements written in natural language. Each layer focuses on improving the requirement quality, such as removing ambiguity, avoiding long sentences, and organizing the structure. This ensures that the final requirements written according to our proposed GGFS-Template are clear and unified. Our proposed template is not specific to any particular domain, so, it can easily be adapted to different domains and handle new types of requirements. Overall, this research introduces a clear and structured way to improve requirement quality. It presents a new rule-based method that connects informal user needs with precise and unified requirements. Our layered rule-based approach provides a strong base for extracting knowledge. Validation using informal texts and accuracy scores shows that informal requirements can be successfully transformed into consistent ones, with a good accuracy rate of 0.89.
Download

Paper Nr:	664
Title:	ViRA-Thyroid: Multimodal Vision-to-Reasoning Augmentation for Interpretable Thyroid Nodule Diagnosis
Authors:	Noura Aboudi and Nawres Khlifa
Abstract:	Ultrasound is the primary modality for thyroid nodule evaluation, but diagnostic accuracy depends strongly on clinician experience. Although deep learning has improved TI-RADS classification, most models remain black boxes and offer limited clinical transparency and interpretability. We propose an explainable CAD framework that combines deep learning–based classification with a Retrieval-Augmented Generation (RAG) module to provide accurate and guideline-consistent explanations. A CBAM-enhanced ResNet50 is fine-tuned for multi-class TI-RADS classification, with th integration of DCGAN data generation model to address limited and data imbalance. A lightweight TI-RADS knowledge base supports explanation retrieval. The system outputs both the predicted TI-RADS category and a concise clinical explanation. Experiments demonstrate improved performance over baseline CNNs and meaningful explanations without loss of accuracy, enhancing transparency and trust in AI-assisted thyroid nodule assessment.

Paper Nr:	666
Title:	Comparison of Text-Based and Image-Based Retrieval in Modern Multimodal Retrieval Augmented Generation Large Language Model Systems
Authors:	Elias Lumer, Alex Cardenas, Matt Melich, Myles Mason, Sara Dieter, Vamse Kumar Subbiah, Pradeep Honaganahalli Basavaraju and Roberto Hernandez
Abstract:	Recent advancements in Retrieval-Augmented Generation (RAG) enable Large Language Models (LLMs) to access multimodal knowledge bases containing text and visual information in financial documents. However, existing multimodal RAG systems rely on LLM-based summarization to convert images into text, causing loss of visual context for retrieval and question answering. We present an analysis of two retrieval approaches for multimodal RAG systems. The first approach uses text-based chunk retrieval where images are summarized into text before embedding. The second uses direct multimodal embedding retrieval where images are stored natively in vector space and passed to vLLMs during generation. We evaluate both approaches across 6 LLM models and 2 multimodal embedding models on a financial earnings call benchmark comprising 40 question-answer pairs. Results show that direct multimodal embedding retrieval outperforms LLM-summary-based approaches, achieving absolute improvements of 13% in mean average precision and 11% in normalized discounted cumulative gain. These gains correspond to relative improvements of 32% in mAP@5 and 20% in nDCG@5. Direct multimodal retrieval also produces more accurate answers by LLM-as-a-judge pairwise comparisons, demonstrating that direct embeddings preserve visual context and reduce information loss.

Paper Nr:	672
Title:	Hybrid CNN-GNN Model for Retinopathy Classification Using Geometric and Topological Invariants
Authors:	Nader Belhadj, Mohamed Amine Mezghich, Jaouher Fattahi, Ridha Ghayoula and Lassaad Latrach
Abstract:	Diabetic retinopathy (DR) is a major cause of preventable blindness, motivating the development of reliable and interpretable automated screening methods. Although convolutional neural networks (CNNs) achieve strong performance in DR classification, they primarily model local appearance cues and remain limited in capturing the global geometric and topological organization of retinal vasculature. We propose a hybrid CNN– GNN framework that explicitly incorporates geometric and topological invariants to address these limitations. A pre-trained EfficientNet-B0 extracts visual features, while a Graph Neural Network (GNN) models relational vessel structure. Curvature-based geometric descriptors and persistent homology–derived topological invariants encode multi-scale vascular connectivity and morphology, providing complementary structural information beyond standard CNN features. The proposed model is evaluated on the Kaggle Diabetic Retinopathy dataset under two clinically relevant settings: five-class severity grading and binary screening (No DR vs. DR). Experimental results demonstrate consistent improvements over CNN-only and CNN–GNN baselines, achieving up to 89.34% accuracy for multi-class classification and 95.10% for binary screening. In addition to performance gains, the topology-aware design enhances interpretability through curvature maps and persistence diagrams, offering clinically meaningful structural insights. These results indicate that integrating geometric and topological descriptors into hybrid deep learning architectures provides a principled and effective enhancement for robust and explainable DR screening.

Paper Nr:	673
Title:	A Native Supervision Approach to Arabic VLM: Overcoming Transliteration Bias for Semantic Accuracy
Authors:	Monia Mahmoudi, Wided Moulahi and Imen Jdey
Abstract:	Current Arabic Handwriting Recognition (AHWR) systems suffer from semantic limitations primarily due to the use of transliterated (Romanized) ground truth labels in major public datasets. This forces the models to learn a character sequence without leveraging the rich linguistic context inherent to the Arabic language, significantly hindering error correction. This paper addresses this fundamental challenge by introducing a novel Vision-Language Model (VLM) architecture that combines a powerful Swin Transformer V2 encoder for robust visual feature extraction with an Arabic-native, pre-trained BERT decoder. Critically, we introduce the ”Native Ground Truth Restoration” methodology, a preprocessing step that systematically reconstructs authentic, fully-vowelized Arabic script from the Romanized labels found in datasets like KHATT. This restoration process enables the BERT decoder to use its deep understanding of Arabic grammar and morphology, allowing it to apply linguistic context for semantic error correction. Our proposed system achieves a state-of-the-art Character Error Rate (CER) of 0.7%, Word Error Rate (WER) of 0.9% and BLEU Score of 0.03% on the KHATT dataset. The resulting system transforms highly variable and chaotic handwriting into semantically accurate digital text, serving as a crucial assistive technology. By facilitating readable formats and enabling high-fidelity Text-to-Speech (TTS) integration, this work provides substantial support for students with learning disabilities, such as Dyslexia and Attention Deficit Hyperactivity Disorder (ADHD).

Paper Nr:	674
Title:	Diagnosis and Treatment Support as AI-Driven Assistant System for TeleNotary Emergency Care
Authors:	Md Faisal Kabir, Konstantin Piliuk and Sven Tomforde
Abstract:	Medical emergencies present significant temporal challenges, as clinicians are required to make multiple critical decisions within extremely limited timeframes. Effective and rapid clinical decision-making is essential to optimise patient outcomes and support healthcare practitioners in high-pressure scenarios. Currently, there is a notable absence of artificial intelligence (AI)–based decision-support systems specifically designed to assist physicians in real-time diagnostic and therapeutic reasoning in pre-hospitalised settings. To address this gap, we propose an AI-assisted TeleNotary decision-support framework that generates probabilistic rankings of potential diagnoses, therapeutic interventions, and required physiological measurements. The system generates the most probable diagnostic and therapeutic recommendations by leveraging a deep, multimodal learning architecture that integrates heterogeneous clinical data sources across three distinct data modalities: time-series signals, tabular records, and unstructured free-text. Performance evaluation of the recommendation system was conducted using standard metrics, including recall@k, macro-averaged AUC, weighted AUC, and coverage ratio, to comprehensively assess both predictive accuracy and recommendation completeness.
Download

Paper Nr:	677
Title:	A Cognitively Plausible Neural-Symbolic Model for Syntactic Disambiguation in Minimalist Grammar
Authors:	Carlos A. Leyva-Capote and Hilton Alers-Valentín
Abstract:	Language comprehension requires resolving structural ambiguity: hierarchical syntactic representations are compressed into linear strings, creating systematic many-to-one mappings where a single utterance admits multiple grammatical interpretations. While large language models demonstrate fluency, they fundamentally cannot guarantee coverage of all grammatically valid analyses-they are generators optimized for surface strings, not selectors operating over grammar-delimited structure spaces. We argue for a neural-symbolic architecture that implements a principled competence-performance separation: a symbolic Minimalist Grammar parser exhaustively enumerates the finite set of grammatically convergent derivations, while a neural performance module assigns preference probabilities by integrating explicit cognitive priors with learned lexical-structural representations via Tree-LSTM composition. This approach transforms the intractable problem of grammatical generation from massive token spaces into tractable selection over small candidate sets, promising data efficiency and interpretability through ablatable processing principles. The architecture makes falsifiable predictions about human sentence processing and provides a computational framework for testing minimalist syntactic theory. We envision integration with compositional semantics and LLMs to enable AI systems to operate in humans’ logical form space-grounded in shared formal representations rather than latent approximations. This synthesis of grammatical precision and statistical learning offers a path toward explainable, cognitively plausible language technologies essential for human-AI alignment.

Paper Nr:	678
Title:	AI-Driven Personalized Shopping: A Mobile Multimodal Platform to Study Affective and Behavioral Responses in a Shopping Context
Authors:	Marcel Antunes Raposo, Luciano Silva and Victor Matheus Batista Nascimento Sedovim
Abstract:	Understanding how consumers emotionally and implicitly interact with brands is essential for designing personalized and adaptive shopping experiences. Traditional approaches rely mostly on explicit self-reports, which often fail to capture the nonconscious mechanisms that drive preference formation and purchase intention. This study proposes and experimentally evaluates an intelligent mobile system that integrates eye tracking, heart-rate sensing, and a semi-generative chatbot to study behavioral and affective indicators related to self–brand congruence during simulated shopping tasks. The tool was implemented as an Android application combining curated visual stimuli, controlled brand-personality manipulation, and multimodal data logging through a hybrid database architecture. Thirteen participants completed three experimental conditions: a baseline session using generic products, a session incorporating a socialization questionnaire, and a session featuring AI-mediated chatbot interaction. Results reveal complementary multimodal patterns across interaction modes. Eye-tracking analysis showed statistically significant differences in visual engagement across conditions, while heart-rate measures captured heterogeneous physiological responses indicative of individual affective dynamics. Behavioral logs indicated a preference for navigation guided by brand-personality profiles, and the chatbot reached a 69% acceptance rate for its suggestions. Post-test questionnaire results demonstrated high usability and positive experiential evaluations, with reliability metrics confirming the internal consistency of the measured factors. The findings suggest that multimodal sensing combined with AI-mediated interaction offers a promising approach for exploring implicit signals in consumer–brand relationships and outlines a research direction toward future personalized commerce systems. Future work includes expanding the participant sample, refining affective conversational strategies, and exploring machine-learning models for automated inference of consumer profiles.

Paper Nr:	682
Title:	Explaining Decisions One Conversation at a Time: Opportunities and Risks of LLMs as Explainability Assistants
Authors:	Filip Cano
Abstract:	Modern AI systems increasingly rely on opaque, highly complex models whose inner workings remain inaccessible even to experts. This opacity creates challenges for trust, accountability, and compliance with emerging regulatory expectations such as the “right to an explanation”. While traditional explainability methods-feature attributions, counterfactuals, surrogate models-and interpretable model classes provide valuable insights for engineers, they often fall short of delivering the contextual, conversational explanations that real users expect. Large Language Models (LLMs) offer a promising new avenue for explanation due to their ability to engage interactively, adapt to user needs, and translate technical outputs into more accessible reasoning. However, their tendencies toward hallucination, conflict avoidance, and oversimplification introduce serious risks when used as explanatory agents. This paper analyzes these opportunities and limitations, examines verification strategies for ensuring explanation fidelity, and situates LLM-generated explanations within broader concerns about public trust. The paper concludes by outlining best practices and future research directions for building robust, verifiable, and human-aligned explanation systems.

Paper Nr:	683
Title:	Backbones in Pseudo-Boolean Optimization: Extraction and Analysis
Authors:	Matías Francia-Carramiñana, Bryan Alvarado-Ulloa, Dorit S. Hochbaum, Bistra Dilkina, Ricardo Ñanculef and Roberto Asín-Achá
Abstract:	Backbone, the set of variables that are fixed across all optimal solutions, captures key structural properties of combinatorial optimization problems. The backbone provides interpretable indicators of problem hardness and, as demonstrated by recent research in SAT, valuable supervision targets for learning-based methods. Yet, the Pseudo-Boolean Optimization (PBO) domain has lacked dedicated tools for its extraction. Here we introduce three new backbone extractors for Pseudo-Boolean Optimization (PBO): ROUNDINGBACK, a PBO-native extractor built on top of RoundingSAT; GUROBACK, an extractor built on top of the MILP solver Gurobi; and NAPBACK, a pipeline that converts PBO instances into SAT using NaPS and delegates the backbone extraction to CadiBack, a backbone extractor for SAT. Additionally, we propose two heuristic variants of ROUNDINGBACK: RB-WP (weighted-propagation ordering) and RB-PG (diversity-driven phase guidance). Each extractor demonstrates distinct strengths tailored to different domain applications. On the PBO Competition 2024 OPT-Lin benchmark, RB-PG achieves the highest extraction coverage (223 of 335 instances, 67%), outperforming GUROBACK and NAPBACK. Our experimental evaluation also shows that ROUNDINGBACK, GUROBACK and NAPBACK are complementary, and it is often the case that when an extractor fails, another is fit. Beyond algorithmic performance, a large-scale analysis over more than eight thousand instances shows a bimodal distribution distinguishing flexible and rigid problem classes. We also show that the correlation between backbone density and problem hardness is domain-specific. All extracted backbones are publicly released to foster future research in learning-based PBO and structural instance analysis.

Paper Nr:	688
Title:	Scalable BDI Logic Inference for Contradictory Knowledge Reasoning
Authors:	Hiraku Gondo, Hiroki Sakaji and Itsuki Noda
Abstract:	We present GrAss-Box Prover, a novel BDI logic inference system that supports higher-order reasoning at a practical computational cost. Real-world knowledge sources inevitably contain contradictory information that conventional first-order logic solvers cannot handle, while LLMs struggle with logical consistency despite their reasoning capabilities. GrAss-Box Prover enhances BDI logic reasoning by introducing three breakthrough optimization techniques: (i) hill-climbing approximation for confidence-level assignment, (ii) bit operations, and (iii) set-based pruning, collectively reducing computational complexity. Our system handles contradictory information and ambiguous beliefs-scenarios where traditional logical systems fail completely. Rigorous evaluation on the BoardgameQA benchmark reveals that GrAss-Box Prover outperforms several prompting methods and neuro-symbolic method while providing fully interpretable proof trees. This represents a practical advancement in automated reasoning, enabling practical deployment of BDI logic in real-world applications requiring both logical rigor and computational efficiency.

Paper Nr:	692
Title:	NPC: Automated Tool for Detecting and Explaining ChatGPT-Generated Programs
Authors:	Pachanitha Saeheng, Napat Boongaree, Chutweeraya Sriwilailak, Chaiyong Ragkhitwetsagul, Teeradaj Racharak and Ekapol Chuangsuwanich
Abstract:	The adoption of Large Language Models (LLMs) is rising in programming education, which raises concerns regarding academic dishonesty and a lack of trustworthiness in students’ programming submissions. There are recent automated techniques and tools for classifying submitted code as generated by LLMs or created by students. However, they lack an explanation of their decision, which educators often require to make informed decisions. This paper presents NPC, an approach for detecting and explaining code snippets generated by ChatGPT, employing machine learning and our proposed local neighborhood sampling strategy to build post-hoc explainability in artificial intelligence (AI). We develop our approach as a web application that not only detects ChatGPT-generated code but also provides educators with explanations in an easy-to-understand format for each classification. The evaluation found that the explanations were clear and helpful, reinforcing the tool’s potential to support academic integrity in programming education. The video demonstration of the tool is available at https://bit.ly/ase25-npc-demo. The tool’s source code is publicly available at https://github.com/pachanitha/NPC Project.

Paper Nr:	693
Title:	SHAP-LLM: Explainability-Guided Synthetic Tabular Data Generation Using Large Language Models
Authors:	Akshata Tejas Karandikar, Christoph Schommer and Salima Lamsiyah
Abstract:	Synthetic tabular data is essential for machine learning applications in settings where real data cannot be shared due to privacy, fairness, or access restrictions. Recent Large Language Models (LLMs) offer a flexible way to generate such data using prompt-based mechanisms, but current LLM-based approaches treat all features equally during generation. This limits their ability to capture the feature dependencies that are most relevant for downstream predictive tasks. In this paper, we present SHAP-LLM, an explainability-guided method that incorporates SHAP (SHapley Additive exPlanations) feature importances into the conditional prompting process. By prioritizing influential features through single-feature, Top-K, and multi-feature conditioning strategies, SHAP-LLM improves the preservation of task-relevant structure while maintaining distributional realism. Experiments on ten benchmark datasets using the Train-on-Synthetic, Test-on-Real (TSTR) protocol show that SHAP-LLM matches or surpasses a strong LLM baseline in predictive accuracy, feature-level fidelity, and label consistency. These results indicate that integrating explainability into LLM-based tabular data generation offers an effective and model-agnostic approach for producing task-aware synthetic data. https://github.com/akshatak1308/SHAP-LLM.git .

Paper Nr:	694
Title:	Machine Learning for Redistricting Analytics: Ensemble Aggregation and Outlier Analysis
Authors:	Sara Asgari, Farzane Asgari and Sadegh Asgari
Abstract:	Redistricting ensemble methods generate thousands of alternative redistricting plans to assess whether an enacted map is an outlier relative to the state’s underlying political and geographic structure and detect gerrymandering. However, ensemble analyses are typically limited to numerical measures of fairness, which do not generalize to all scenarios, overemphasize specific notions of fairness, are dependent on election data, or overlook high-dimensional spatial data. This paper introduces a machine learning framework for ensemble aggregation that identifies latent patterns in redistricting plans through unsupervised clustering of precinct-level district assignments. Using a weighted-graph representation of plans, we cluster large ensembles to construct “representative” or “average” redistricting plans and visualize the dominant structural features of feasible maps. We then evaluate the enacted plan’s placement within this learned “consensus” plan using visualizations. Experiments on redistricting datasets of two real U.S. states with differing characteristics show that clustering reveals stable structural modes within the ensemble, isolates atypical plans, and highlights whether the enacted plan departs from typical district configurations expected under neutral criteria. Our results demonstrate that machine learning aggregation provides a principled and interpretable extension to ensemble-based redistricting analytics, enabling more robust and interpretable assessments of fairness. Our work is applicable in geospatial analytics beyond redistricting and can also be generalized to not only help detect insightful demographic and political patterns and communities of interest but also integrate communities of interest within ensemble generation.

Paper Nr:	698
Title:	KAHLT: Kolmogorov-Arnold Hybrid Latent Transformer - A Hybrid Transformer For Classification Tasks
Authors:	Sanyam Sanjay Jain, Siddharth Khare, Paarth Goyal and Jagat Sesh Challa
Abstract:	We propose the Kolmogorov-Arnold Hybrid Latent Transformer (KAHLT), a hybrid architecture that integrates the superior expressivity of Kolmogorov-Arnold Networks (KANs) with the linear scalability of Perceiver-style attention. Addressing the high computational cost of KANs and the limited inductive bias of standard Multi-Layer Perceptrons (MLPs), KAHLT employs a fixed-size latent bottleneck populated by heterogeneous processing units. We introduce Hybrid Projections and a Grouped Hybrid Feed-Forward Network, which partition feature processing into parallel Chebyshev Polynomial based KAN and linear MLP expert lanes. We hypothesize that this architecture learns to separate features based on how difficult they are to model, allocating high-complexity features to KAN experts while reserving efficient linear lanes for simpler transformations. Empirical results on the arXiv benchmark demonstrate that KAHLT outperforms both pure MLP and fully KAN-based baselines, offering a superior balance of parameter efficiency, interpretability, and performance for sequence modeling.
Download

Paper Nr:	699
Title:	BirdCallNet: Joint Species and Call-Type Classification
Authors:	Paria Vali Zadeh and Sven Tomforde
Abstract:	Passive acoustic monitoring enables large-scale, non-invasive observation of ecosystems and animal behaviour. We conduct experiments on WiWa, a newly annotated bioacoustic dataset of forest soundscapes that provides both species labels and fine-grained vocalisation (call-type) annotations, addressing a common gap in existing bird-sound corpora. We benchmark four pretrained encoders-ConvNeXtBS, EAT, BirdMAE, and ProtoCLR-covering convolutional and transformer architectures trained on large-scale bird sound data. We compare a single-task formulation in a combined label space (species#call type) against a multi-task learning framework with parallel heads that jointly predict species and call types, and study three adaptation strategies: linear probing and an attentive pooling head over frozen encoders, and full fine-tuning as a reference. Across experiments, Multi-Task learning consistently improves both species and call-type prediction, with particularly strong gains on call types, indicating that shared representations benefit from cross-label supervision. Under matched budgets, linear probing remains a strong and often competitive baseline with markedly fewer trainable parameters, while an attentive head on BirdMAE achieves the best overall performance on WiWa. Full fine-tuning of the backbone and task heads, performed with a shared optimisation setting across backbones for a controlled comparison, yields mixed gains and does not consistently outperform probing, suggesting sensitivity to per-backbone hyperparameter choices. These results highlight the value of jointly annotated labels for bioacoustic analysis and provide a simple yet strong benchmark for future work on multi-label bird-sound classification.

Paper Nr:	700
Title:	Rental Price Prediction with Machine Learning Using Future Population Statistics
Authors:	Goei Sato, Ryo Hatano and Hiroyuki Nishiyama
Abstract:	This study aims to clarify how external public datasets, such as official land price information and future population projections, can be effectively incorporated into rent prediction models in Japan. Using approximately 9.18 million rental listings from 2018 to 2024, we constructed a nationwide unified model and eight region-specific models based on LightGBM, and evaluated three strategies for using external data-(i) no external data, (ii) adding external variables as features, and (iii) using them as post-prediction correction signals-under multiple forward time-series forecasting settings. The experimental results showed that incorporating land price data as features led to modest and context-dependent accuracy improvements, while future population statistics had only a small impact when directly used as features. In contrast, using population projections as post-prediction correction signals consistently reduced prediction errors, and the region-specific model with correction achieved the highest overall accuracy. These findings indicate that the effectiveness of external public data critically depends on how they are integrated, and that post-prediction correction provides a practical and scalable approach for nationwide rent prediction.

Paper Nr:	705
Title:	Assessing Moral Frameworks for Autonomous Vehicle Decision-Making
Authors:	Amir Rafiee
Abstract:	The prospect of fully autonomous vehicles (AVs) has intensified longstanding questions about how moral principles should guide decisions in situations where harm cannot be entirely avoided. Unlike human drivers, whose responses in emergencies are often inconsistent and opaque, AVs must be programmed to respond in systematic and reproducible ways. This paper examines three major normative traditions, utilitarianism, deontological ethics, and Rawlsian justice, as candidate foundations for ethical decision-making in AVs. Each framework is analysed in terms of its central commitments, its implications for fairness and responsibility, and the challenges of translating it into algorithmic form. Using stylised crash scenarios and thought experiments, the paper shows how these theories can yield diverging recommendations regarding harm minimisation, rights protection, and the treatment of vulnerable road users. The analysis reveals that while each approach captures important moral intuitions, none on its own offers a complete or uncontroversial basis for programming AVs in life-and-death situations. The paper concludes that progress will likely require hybrid architectures that combine optimisation techniques with rule-based constraints, fairness protections, and, in rare cases, carefully limited randomisation as a procedural tie-breaker. These reflections aim to support developers, ethicists, and policymakers in articulating transparent and publicly defensible principles for future AV behaviour.

Paper Nr:	708
Title:	Zero-Shot Anonymization of Customer Communications in German Energy Sector: A Comparative Study of On-Premise Open-Source LLMs
Authors:	Rico Herlt, Maximilian Orlowski and Florian Marquardt
Abstract:	This work presents a zero-shot prompt engineering approach evaluated across multiple open-source Large Language Models (LLMs) to identify and anonymize Personally Identifiable Information (PII) in real-world Customer-to-Business (C2B) communications from network operator processes in the German Energy Sector. This study addresses the requirement to work with anonymized real-world data rather than synthetic data. Such data is essential for business and software development processes, enabling analysis of first-level support scenarios, capturing authentic linguistic variations and human communication errors that are critical for representative test data. A practical approach for generating anonymized Customer-to-Business communications is demonstrated that enhances data privacy and ensures regulatory compliance in daily operations with real-world customer data. Five open-source LLMs are compared for performance on a manually annotated test set, demonstrating that GPT-OSS (20B parameters) achieves the highest success rate of 99% with the zero-shot prompt.

Paper Nr:	711
Title:	Modeling and Predicting Trust Dynamics in Manipulator Actions
Authors:	Sota Kaneko and Seiji Yamada
Abstract:	This study looks at how human trust changes during repeated interactions with a robot. It proposes a framework based on dynamic structural equation modeling (DSEM) to predict future trust states. Two hundred participants watched five pick-and-place actions in one of four performance scenarios defined by different success and failure sequences. A manipulation check showed that participants noticed a difference in task difficulty between two object types, and these difficulty ratings were used in the model as an observed variable. Task outcomes were also treated as observed variables, while trust was modeled as a hidden cognitive construct. This trust was measured indirectly through responses based on the MDMT, collected after each action. The resulting model captured patterns of trust over time in all scenarios and successfully predicted hidden trust at the next time step on the basis of only observed variables. These findings show that trust can be seen as a changing process influenced by performance sequences and perceived task difficulty. This provides a basis for robotic systems that can actively adjust their behavior to support the right level of human trust.

Paper Nr:	712
Title:	The Green Computing Infrastructure & Reporting Ontology: A Modular Ontology for Sustainable IT Management
Authors:	Ahmed Dridi, Aroua Taamallah, Salma Sassi and Richard Chbeir
Abstract:	This paper presents the Green Computing Infrastructure & Reporting Ontology (GCIRO), a modular ontology for modeling and managing the environmental footprint of ICT infrastructures. GCIRO unifies representations of ICT assets and facilities with multi-dimensional sustainability metrics (energy, carbon, water, circularity), governance and lifecycle processes, interventions, and evidence. Built using the NeOn methodology and reusing established vocabularies (PROV-O, QUDT, SOSA/SSN, BOT), it provides a coherent semantic backbone for integrated sustainability analysis and regulatory-aligned reporting. An OWL implementation and validation through reasoning and SPARQL competency questions on realistic data center scenarios show that GCIRO can support monitoring, optimization, and auditable green computing management.

Paper Nr:	715
Title:	Rethinking Retrieval: From Traditional Retrieval Augmented Generation to Agentic and Non-Vector Reasoning Systems in the Financial Domain for Large Language Models
Authors:	Elias Lumer, Matt Melich, Olivia Zino, Elena Kim, Sara Dieter, Max Harberg, Pradeep Honaganahalli Basavaraju, Vamse Kumar Subbiah, James A. Burke and Roberto Hernandez
Abstract:	Recent advancements in Retrieval-Augmented Generation (RAG) enable Large Language Models to answer financial questions from SEC filings, earnings reports, and regulatory documents. However, existing work lacks systematic comparison of vector-based and non-vector RAG architectures for financial documents, and its impact on retrieval accuracy and latency remains unclear. We present an evaluation comparing vector-based agentic RAG against hierarchical node-based systems that traverse sections without embeddings. We evaluate two enhancement techniques applied to the vector-based architecture, including reranking and small-to-big retrieval for answer quality. Across 1,200 SEC 10-K, 10-Q, and 8-K filings on a 150-question benchmark, we measure retrieval metrics (MRR, Recall@5), answer quality through pairwise LLM-as-a-judge, latency, and cost. Agentic RAG achieves a 68% win rate over hierarchical node-based systems with comparable latency (5.2 compared to 5.98 seconds). Reranking achieves a 59% absolute improvement at optimal parameters (10, 5) for MRR@5. Small-to-big retrieval achieves a 65% win rate over baseline chunking with only 0.2 seconds additional latency. Results show that applying advanced RAG techniques based on query complexity optimizes cost-performance tradeoffs for financial question answering.

Paper Nr:	16
Title:	Hallucination Mitigation with Agentic AI NLP-Based Open-Floor Standard
Authors:	Diego Gosmar and Deborah A. Dahl
Abstract:	One of the main challenges in AI foundation model pretraining, as well as in fine-tuning transfer learning, is hallucinations. In this paper, we examine how orchestrating multiple specialized agents can reduce such hallucinations, with an emphasis on systems that employ NLP (Natural Language Processing) to coordinate agent interactions. We test a pipeline that introduces three hundred and ten prompts, specifically engineered to induce hallucinations, into a front-end agent. This agent’s output is then reviewed and refined by second-and third-level agents, each of which employs different large language models and strategies to flag unverified claims, provide explicit disclaimers, and clarify any speculative elements. Key Performance Indicators (KPIs) are collected to measure hallucination-related behaviors with evaluations performed by a fourth-level agent. Our findings demonstrate the feasibility of multi-agent orchestration for hallucination mitigation and highlight the value of maintaining a structured exchange of meta-information.

Paper Nr:	19
Title:	BioAbbreviate: A Biomedical Dataset for Abbreviation Expansion and Disambiguation
Authors:	Mouheb Mehdoui, Amel Fraisse, Widad Mustafa El Hadi and Mounir Zrigui
Abstract:	The BioAbbreviate dataset addresses the challenges of abbreviation expansion and disambiguation in biomedical texts, a critical task for improving health literacy and text simplification. Abbreviations in medical literature often introduce ambiguity, translation issues, and information loss, complicating comprehension for non-experts. This paper presents a large-scale dataset curated from open-access biomedical repositories, including PubMed, PLOS, and PMC Europe, leveraging automated extraction and validation techniques to capture diverse abbreviation-expansion pairs. The dataset supports the development of robust NLP systems for abbreviation detection and expansion, with evaluations demonstrating the superior performance of domain-specific models like BioBERT. Key contributions include scalable data collection, pattern-based abbreviation identification, and contextual disambiguation strategies. The dataset aims to enhance readability, support downstream NLP tasks, and bridge gaps in biomedical text processing.

Paper Nr:	25
Title:	Leveraging Generative AI Models for Multidomain Network Security
Authors:	Rasa Brūzgienė, Šarūnas Grigaliūnas, Ilona Veitaitė, Renata Danielienė, Kęstutis Driaunys, Paulius Astromskis, Živilė Nemickienė, Dovilė Vengalienė, Rokas Stankūnas, Ieva Andrijauskaitė and Neringa Gaubienė
Abstract:	The rise of Generative Artificial Intelligence (GenAI) presents new opportunities for enhancing network security across technical, operational, human, and physical domains. This paper proposes a GenAI-driven network security framework that integrates OSI-layer–aware threat analysis with mandatory human oversight to support CISO decision-making. The framework is empirically evaluated using NetFlow datasets representing port scanning, ICMP flooding, and SPAM attacks. Six GenAI models are assessed using explainability index, hallucination rate, and token efficiency metrics. The results demonstrate that while certain models achieve high analytical accuracy and explainability, performance varies significantly in efficiency and hallucination behavior. The study further discusses legal and regulatory implications of deploying GenAI in security-critical environments, highlighting the necessity of human-in-the-loop control for accountable and reliable network security operations.

Paper Nr:	27
Title:	CVaR-SHAP: A Comprehensive Framework for Feature Attribution with Integrated Risk Quantification
Authors:	Marwa Thabet and Brahim Hnich
Abstract:	The interpretability of machine learning models, such as SHAP, is crucial for understanding model behavior. SHAP, in particular, calculates the contribution of each feature -referred to as SHAP values- using Shapley value estimates. However, since these estimates rely on sampling, they are subject to uncertainty. In this paper, we propose a novel framework based on SHAP that quantifies the risk associated with feature attribution. Specifically, we leverage the core feature sampling algorithm to measure the Conditional Value at Risk (CVaR). We demonstrate these risk assessments through a series of experiments, which reveal variations in risk across different features and their impact on model interpretability. This approach supports more informed decision-making in machine learning models. It leads to a more robust evaluation of interpretability and provides valuable insights into the stability and reliability of feature attributions.

Paper Nr:	40
Title:	Predicting Conductivity of Typical Nanocomposite Materials Using Extreme Gradient Boosting Machine Learning: A Comparative Analysis between Optuna and Grid Search Cross-Validation Hyperparameter Tuning
Authors:	Ryan Tyler, Masengo Ilunga, Arjun Maity and Pramod Sinha
Abstract:	Accurately predicting the electrical conductivity of nanocomposites is crucial for their design and use in electronics, energy storage, sensing, and civil engineering. Conductivity depends on multiple interacting material and processing parameters, making it a highly non‑linear property. This study applies an Extreme Gradient Boosting (XGBoost) model to predict the conductivity of a polymer-graphene nanocomposite using five features: nanomaterial size, tensile strength, flexural strength, temperature, and concentration. A key focus was comparing two hyperparameter tuning strategies, namely the traditional Grid Search Cross‑Validation (GridSearchCV) and the Bayesian optimization framework Optuna. Both models were trained and evaluated on a publicly available dataset reflecting known physical behaviours in nanocomposites. Performance was assessed using R2, RMSE, MAE, and computational cost. The Optuna‑tuned XGBoost model achieved slightly better accuracy, with an R2 of 0.740, RMSE of 26.84, and MAE of 19.67, compared to 0.737, 26.97, and 20.01 for GridSearchCV. However, this improvement required 56% more computation time. Overall, the results showed that XGBoost is a robust predictor of nanocomposite conductivity, and that Bayesian optimization methods such as Optuna can identify more effective hyperparameter configurations than exhaustive Grid Search, offering advantages for data‑driven materials design.

Paper Nr:	47
Title:	Detecting AI-Generated Answers in Software Engineering Assessments
Authors:	Ta Nguyen Binh Duong
Abstract:	We consider the problem of detecting AI-generated answers from students in higher education, in particular software engineering assessments. Existing detectors are mostly general-purpose, which might not perform well when the application domain changes or when the AI answers are paraphrased. We implement LLM-as-judge approaches and evaluate the latest reasoning and non-reasoning LLMs on their detection performance in our course. We then propose a new scoring-based detection approach to deal with the token cost issue and common paraphrasing attacks. The approaches are evaluated using several real-world datasets, which consist of AI-generated and paraphrased answers, as well as actual student answers to quizzes over three consecutive semesters in a software engineering course at our institution. The results demonstrate that reasoning LLMs provide stronger detection performance even when compared to more expensive non-reasoning models. At the same time, our cost-effective scoring-based detector is shown to be more resilient compared to reasoning LLMs and other existing detectors in paraphrasing attacks.

Paper Nr:	48
Title:	Combination of Ontology-Based Retrieval Augmented Graphs and Inductive Reasoning for Knowledge Fusion of Intelligent Systems
Authors:	Hien D. Nguyen, Duc M. Truong, Sang Vu, Diem Nguyen, Tai Huynh and Vuong T. Pham
Abstract:	This paper presents a novel e-learning knowledge retrieval method that integrates ontology-based retrieval-augmented graphs with inductive reasoning to enable intelligent knowledge fusion. Ontologies provide structured domain semantics, while knowledge graphs represent dynamic relations between concepts. By applying inductive reasoning over this integrated structure, the system infers contextually relevant and semantically enriched answers, effectively bridging gaps in learners’ understanding. Evaluated in the context of a Fundamentals of Database Systems (FDS) course course, the proposed method significantly outperforms traditional approaches in both precision and completeness of information retrieval. The integration enhances query interpretation and supports personalized, adaptive learning experiences. This approach demonstrates strong potential to elevate e-learning platforms into intelligent systems capable of delivering richer, more relevant content through semantic awareness and reasoning.

Paper Nr:	49
Title:	An Application to Automatically Detect Track-Limit Violations in Car Races
Authors:	Morgana Bellacci, Eduard Brahas, Marco Prepi, Fabio Seghetta, Francesco Santini and Michele Vantaggi
Abstract:	This paper presents an automated system to assist race stewards in detecting track-limit violations during motorsport events. Using monitoring cameras and YOLO-based computer vision, the system automatically identifies cars that cross predefined track boundaries and generates short video clips for human review. After comparing two implementation frameworks, we developed and deployed a dedicated race-car detection model trained on multiple vehicle categories to improve accuracy over standard pretrained models. The system is currently used in official races and significantly reduces the stewards’ manual workload.

Paper Nr:	84
Title:	Case-Based Prediction Using a Continuous Compatibility Measure
Authors:	Chunyang Fan, Fadi Badra and Marie-Jeanne Lesot
Abstract:	Case-based prediction (CBP) addresses classification or regression tasks applying the analogical transfer principle (ATP) that infers information from similar known situations provided in the so-called case base: ATP leverages the assumption that similarity in some aspects, e.g. data features, implies similarity in others, e.g. data class. This paper considers the family of transfer by global optimization and complexity-based approaches. It proposes a continuous compatibility measure which quantifies the extent of ATP violation observed in the case base rather than merely counting the number of such violations. This new compatibility measure leads to a new classifier called CCoAT (Continuous Complexity-based Analogical Transfer). Furthermore, it allows for gradient-based optimization of the considered similarity measures: the paper proposes a metric learning method, called CCoSiL, to optimize the similarity measures for case-based prediction. The comparative experimental study conducted on several benchmark data sets shows the relevance of both the proposed classifier and metric learning method: CCoAT combined with CCoSiL achieves higher accuracy than when combined with the state-of-the-art metric learning method LMNN. In addition, CCoSiL can be applied to several CBP algorithms, it generalizes well to k-NN.

Paper Nr:	88
Title:	Human-in-the-Loop: Quantitative Evaluation of 3D Models Generation by Large Language Models
Authors:	Ahmed R. Sadik and Mariusz Bujny
Abstract:	Large Language Models (LLMs) are increasingly capable of interpreting multimodal inputs to generate complex 3D shapes, yet robust methods to evaluate geometric and structural fidelity remain underdeveloped. This paper introduces a human-in-the-loop framework for the quantitative evaluation of LLM-generated 3D models, supporting applications such as democratization of CAD design, reverse engineering of legacy designs, and rapid prototyping. We propose a comprehensive suite of similarity and complexity metrics-including volumetric accuracy, surface alignment, dimensional fidelity, and topological intricacy-to benchmark generated models against ground-truth CAD references. Using an L-bracket component as a case study, we systematically compare LLM performance across four input modalities: 2D orthographic views, isometric sketches, geometric structure trees, and code-based correction prompts. Our findings demonstrate improved generation fidelity with increased semantic richness, with code-level prompts achieving perfect reconstruction across all metrics. A key contribution of this work is demonstrating that our proposed quantitative evaluation approach enables significantly faster convergence toward the ground truth, especially compared to traditional qualitative methods based solely on visual inspection and human intuition. This work not only advances the understanding of AI-assisted shape synthesis but also provides a scalable methodology to validate and refine generative models for diverse CAD applications.

Paper Nr:	96
Title:	Cooperative Agents Based Federated Learning for Ransomware Detection and Response in Cloud Systems
Authors:	Abderrahmen Chaieb, Wafa Mefteh and Ali Frihida
Abstract:	This paper proposes a novel defense strategy using a Cooperative Multi-Agent System and Federated Learning that will help enhance ransomware detection in cloud environments. Cloud systems are faced with the threat of being attacked by sophisticated ransomware for which traditional security methods are clearly insufficient. The solution employs intelligent agents over distributed nodes to monitor, detect, and respond to the threats collaboratively. In enhancing the detection models locally with preserving data privacy for improved security, the agents leverage federated learning. The proposed system is scalable, allowing for realtime collaboration between agents to detect ransomware more effectively. The paper further presents discussions on cloud system security issues, various security approaches, and performance evaluations of the proposed system in real-world scenarios, showing its efficiency in ransomware detection.

Paper Nr:	115
Title:	ZQBA: Zero Query Black-Box Adversarial Attack
Authors:	Joana C. Costa, Tiago Roxo, Hugo Proença and Pedro R. M. Inácio
Abstract:	Current black-box adversarial attacks either require multiple queries or diffusion models to produce adversarial samples that can impair the target model performance. However, these methods require training a surrogate loss or diffusion models to produce adversarial samples, which limits their applicability in real-world settings. Thus, we propose a Zero Query Black-box Adversarial (ZQBA) attack that exploits the representations of Deep Neural Networks (DNNs) to fool other networks. Instead of requiring thousands of queries to produce deceiving adversarial samples, we use the feature maps obtained from a DNN and add them to clean images to impair the classification of a target model. The results suggest that ZQBA can transfer the adversarial samples to different models and across various datasets, namely CIFAR and Tiny ImageNet. The experiments also show that ZQBA is more effective than state-of-the-art black-box attacks with a single query, while maintaining the imperceptibility of perturbations, evaluated both quantitatively (SSIM) and qualitatively, emphasizing the vulnerabilities of employing DNNs in real-world contexts. All the source code is available at https://github. com/Joana-Cabral/ZQBA.

Paper Nr:	121
Title:	Innovative Multi-Agent Architectures Propelling Genomic Analysis in General and Oncology Domains
Authors:	Zayneb Mannai, Nizar Omheni and Ramzi Mahmoudi
Abstract:	The explosion of genomic and multi-omics data is creating an urgent need for analysis platforms that are both intelligent and flexible. Multi-agent systems (MAS) are well suited to this challenge because they can distribute tasks across autonomous agents and adapt to complex workflows. Recent progress in large language models (LLMs) adds new possibilities, enabling cognitive agents that can reason, explain results, and perform adaptive quality control. In this paper we present a comprehensive quantitative review of twenty-five multi-agent genomic analysis systems reported between 2022 and 2025, including twenty designed for general genomics and five focused on oncology. Using clear comparison criteria, we examine architectural styles, task execution modes, automation levels, agent types and LLM adoption. Our findings reveal clear maturity gaps between general genomics and oncology, show that more than sixty percent of recent systems already integrate LLMs, and highlight opportunities for future work such as multi-omics reasoning agents and explainable clinical decision support.

Paper Nr:	125
Title:	Sign Language Datasets for Machine Learning: Evaluation of Datasets for ASL and Selected Languages
Authors:	Pawel Antonowicz, Marek Baranowski and Michal Podpora
Abstract:	The advancements in Sign Language Recognition (SLR) depend on having access to well-structured datasets. This review examines publicly available resources for American Sign Language (ASL) and selected other sign languages, covering datasets for letter, isolated word, and continuous signing tasks. Key aspects such as size, class balance, annotation quality, and general suitability for Machine Learning are assessed. Common issues, including data imbalance and availability limitations, are identified. Baseline training on selected SL datasets illustrates how these factors impact model performance. This review aims to support researchers in selecting appropriate datasets as well as recognizing their limitations in SLR applications.

Paper Nr:	132
Title:	Integration of Stacking Case-Based Reasoning with a Multi-Agent System Applied to Regression Problems
Authors:	Daniel Soto Forero, Marie-Laure Betbeder and Julien Henriet
Abstract:	This paper presents an integration model between CBR (Case Based Reasoning) and a MAS (Multi-Agent System) that improves our previously proposed model called ESCBR (Ensemble Case Based Reasoning). The integration model is designed to find approximate solutions to generic regression problems of one or more dimensions. The integration pretends to explore and exploit the solutions space better by using an individual cognitive process with social interactions and Bayesian reasoning. The proposed multi-agent system is composed of several heterogeneous agents pursuing a common goal, but in different ways. Each agent can choose an action from three possibilities and interact with other agents randomly. Bayesian learning allows each agent to learn using the information obtained from past iterations and to adapt its behavior according to the characteristics of the evaluated problem. The model has been evaluated in comparison with nine other regression models on eleven different regression databases extracted from the UCI site. The comparison between the models has been carried out with RMSE (Root Mean Squared Error) and MAE (Median Absolute Error) metrics. The results show that the proposed model obtains good results, globally among the three best models and effectively improves on the ESCBR base model.

Paper Nr:	135
Title:	A Frugal Heuristic for Guessing the Truth of Formulas
Authors:	Célia da Costa Pereira and Andrea G. B. Tettamanzi
Abstract:	Being able to tell if a formula is true or false given a knowledge base is an important task in Logic. Logic reasoners are designed to decide whether a formula is a logical consequence of the knowledge base, which is stronger than that and can be computationally expensive in some cases. In addition, under the open-world assumption, it may be impossible to infer a formula or its negation. In some practical situations, however, when beliefs about the world change rapidly and an agent has to make decisions in real time, it is acceptable to resort to heuristic methods to determine the probable veracity or falsehood of a formula, even in the absence of a guarantee of correctness. This is why we propose a fast heuristic to guess whether an arbitrary, unseen formula is true or false, without the need to train a model. Our method implicitly exploits a model-theoretic measure of semantic similarity among formulas. It repeatedly generates random interpretations and counts how often the truth value of the unseen formula aligns with the truth values of known, already labeled formulas under these interpretations. The results of the experiments show that the proposed heuristic has an accuracy comparable to more sophisticated methods, while being less computationally expensive, making it suitable for scenarios where an agent’s beliefs change frequently and quick decisions are required.

Paper Nr:	140
Title:	Comparative Analysis of LLMs for Software Quality Assessment via Code and Metrics
Authors:	Laura Cernău, Andrei-Paul Dobrescu, Ecaterina Cărbune and Georgiana Asandei
Abstract:	The use of large language models (LLMs) to analyse and identify errors in code is becoming increasingly common among developers. While many studies aim to improve the quality and effectiveness of LLM-generated code, this paper investigates how LLMs perceive real-world code compared to established code metrics. We evaluate two widely used models, OpenAI GPT-4o and Gemini 2.0 Flash, to determine whether their identification of architectural issues remains consistent when provided solely with code, code combined with metrics, or metrics alone. We asked the models to provide a brief assessment of potential problems in a file under each of these conditions. Our analysis shows that LLMs can often correctly identify errors exclusively based on metrics. Building on this finding, we further asked the models to assign a simple label-GOOD, MEDIUM, or BAD-reflecting their evaluation with minimal context. While prior research has focused on LLMs as code generators or bug fixers, few studies have explored their ability to evaluate code quality using abstract indicators such as metrics, we combined the two approaches (hybrid between code and metrics). Our results suggest that LLMs can assess code quality even in the absence of a full code context.

Paper Nr:	141
Title:	T-BEx: Transformer-Bayesian Explanations for Mental Health Detection
Authors:	Sabrine Toumi, Hasna Njah and Salma Jamoussi
Abstract:	Transformer-based models achieve state-of-the-art performance in detecting mental health conditions from social media text, yet their limited interpretability hinders clinical adoption. Existing post-hoc explainers identify influential features but fail to model conditional dependencies, reducing clinical usefulness. We propose T-BEx (Transformer–Bayesian Explanations), a meta-explainer that integrates token-level attributions with a Tree-Augmented Naive Bayes (TAN) structure to capture conditional dependencies among linguistic signals. This formulation yields context-aware, probabilistically structured explanations aligned with clinical reasoning. Experiments on four Reddit and Twitter datasets show that T-BEx improves explanation stability and sensitivity over baseline explainers while maintaining high fidelity and sparsity. These results demonstrate that T-BEx produces explanations that are faithful, robust, and clinically meaningful, advancing trustworthy AI for mental health.

Paper Nr:	143
Title:	Retrieval-Augmented Generation for Procurement Validation
Authors:	Daiga Deksne, Raivis Skadiņš, Mārcis Pinnis, Andris Hohbergs, Rūdolfs Jaunzars, Andrejs Petrovs and Justīne Rūdule
Abstract:	Retrieval-augmented generation (RAG) is applied across diverse domains, allowing to employment of large language models (LLMs) for in-domain question-answering without the need for fine-tuning of the generative LLMs with in-domain data. In this work, we analyse the applicability of RAG for procurement validation. We compare various configurations of different methods involved in the RAG process and find the best-performing methods for procurement validation. We analyze the impact of various text extraction libraries, segmentation strategies with different segment sizes, embedding model selection, and prompt construction methods. Furthermore, we show that recall of document retrieval can be improved by fine-tuning the embedding model with in-domain data - a collection of procurement documents in Latvian. Our best-performing configuration achieves a procurement validation accuracy of 70.73% on a publicly available procurement validation dataset for Latvian.

Paper Nr:	148
Title:	BondBERT: What We Learn when Assigning Sentiment in the Bond Market
Authors:	Toby Barter, Zheng Gao, Eva Christodoulaki, Jing Chen and John Cartlidge
Abstract:	Bond markets respond differently to macroeconomic news compared to equity markets, yet most sentiment models are trained primarily on general financial or equity news data. However, bond prices often move in the opposite direction to economic optimism, making general or equity-based sentiment tools potentially misleading. We introduce BondBERT, a transformer-based language model fine-tuned on bond-specific news. BondBERT can act as the perception and reasoning component of a financial decision-support agent, providing sentiment signals that integrate with forecasting models. We propose a generalisable framework for adapting transformers to low-volatility, domain-inverse sentiment tasks by compiling and cleaning 30,000 UK bond market articles (2018–2025). BondBERT’s sentiment predictions are compared against FinBERT, FinGPT, and Instruct-FinGPT using event-based correlation, up/down accuracy analyses, and LSTM forecasting across ten UK sovereign bonds. We find that BondBERT consistently produces positive correlations with bond returns, and achieves higher alignment and forecasting accuracy than the three baseline models. These results demonstrate that domain-specific sentiment adaptation better captures fixed income dynamics, bridging a gap between NLP advances and bond market analytics.

Paper Nr:	149
Title:	Acoustic-Based System for Early Detection of Elephants in Forests
Authors:	Sandra Gathoni Kahoro and Henry Muchiri Muriithi
Abstract:	Elephants heavily rely on acoustic communication for survival yet human activities which lead to human-elephant conflict increasingly threaten or degrade their natural habitats, hence reducing their population. Human-elephant conflict is one of the greatest causes of elephant population decrease making early detection of elephants critical for conservation. Elephants communicate through vocalizations the most used are low rumbles which are very low frequency and can travel over long distances and are below human hearing level. This paper presents a lightweight machine learning model for the detection of elephant vocalizations, implemented and tested on a Raspberry Pi Pico microcontroller. A dataset consisting of elephant calls and forest sounds was pre-processed and trained using a Convolutional Neural Network (CNN) model. The model achieved a high accuracy during evaluation showing its suitability for distinguishing elephant calls from background forest noise. To test the feasibility, the model was deployed onto Raspberry Pi Pico, integrated with a GPS and PDM microphone and the system successfully processed acoustic signals in real-time and displayed detection outcomes through an SMS message, confirming the practicality of low-cost, low-power devices for monitoring and detection. The results demonstrate that machine learning models can be adapted for deployment in resource-constrained IoT environments.

Paper Nr:	150
Title:	Bridging Ethical Concerns and Explainability through Fuzzy LS
Authors:	Narek Andreasyan, Egzont Demiri, Edy Portmann and Luis Terán
Abstract:	The rapid expansion of digital services necessitates robust ethical assessment frameworks to address issues such as bias, privacy concerns, and unforeseen societal impacts. Traditional rule-based (crisp logic-based) methods often fail to accommodate the complexity and imprecision inherent in modern digital systems (e.g. systems that used artificial intelligence). This paper proposes an innovative approach that integrates fuzzy logic-based LS to enhance ethical assessments for the use case of electronic voting (e-voting), where ethical concerns such as security, transparency, and fairness are crucial. Fuzzy logic provides a flexible means to handle vagueness in ethical assessments, moving beyond rigid binary classifications. LS translates complex data into intuitive insights, making ethical considerations more interpretable for diverse stakeholders. By combining these techniques, ethical assessments become more transparent, adaptable, and accessible, ensuring better alignment with real-world decision-making. This project develops a web-based application to process user inputs through surveys, using fuzzy logic-based LS to generate explainable assessments. By bridging theoretical ethical frameworks with practical implementation, this design science research project provides a dynamic tool to evaluate and enhance ethical standards in digital services. The proposed approach fosters trust, accountability, and sustainable innovation across various sectors, ensuring that digital transformation aligns with ethical responsibilities.

Paper Nr:	163
Title:	TinyGraphEstimator: Adapting Lightweight Language Models for Graph Structure Inference
Authors:	Michal Podstawski
Abstract:	Graphs provide a universal framework for representing complex relational systems, and inferring their structural properties is a core challenge in graph analysis and reasoning. While large language models have recently demonstrated emerging abilities to perform symbolic and numerical reasoning, the potential of smaller, resource-efficient models in this context remains largely unexplored. This paper investigates whether compact transformer-based language models can infer graph-theoretic parameters directly from graph representations. To enable systematic evaluation, we introduce the TinyGraphEstimator dataset - a balanced collection of connected graphs generated from multiple random graph models and annotated with detailed structural metadata. We evaluate several small open models on their ability to predict key graph parameters such as density, clustering, and chromatic number. Furthermore, we apply lightweight fine-tuning using the Low-Rank Adaptation (LoRA) technique, achieving consistent improvements across all evaluated metrics. The results demonstrate that small language models possess non-trivial reasoning capacity over graph-structured data and can be effectively adapted for structural inference tasks through efficient parameter tuning.

Paper Nr:	168
Title:	Integrated AI Approaches for Forest Fire Prediction: A Comparative Study of Regression and Deep Learning Models
Authors:	Ioan Daniel Pop, Andrei Văran and Adriana Mihaela Coroiu
Abstract:	Forest fires have become an major ecological and socioeconomic problem, amplified by the effects of climate change, prolonged droughts, and anthropogenic factors. Predicting their risk and impact represents a strategic research direction, with the potential to directly contribute to saving exposed ecosystems, resources, and communities. In this context, the present work investigates and compares two complementary artificial intelligence (AI) paradigms applied to forest fire prediction: regression algorithms on historical meteorological data and convolutional neural networks applied to satellite images. For the first approach, historical datasets from the Montesinho Natural Park (Portugal) were used, with variables such as temperature, humidity, wind and fine fuel dryness indices, to estimate the potentially affected area. Random Forest and Gradient Boosting models were trained and optimized, applying cross-validation techniques and hyperparameter optimization via Grid-SearchCV. For the second approach, a set of labeled satellite images (fire/presence and absence), processed and normalized, was used to train convolutional neural networks, including a custom CNN architecture and the ResNet50 model adapted by transfer learning. The results obtained highlight the performances and limitations of each method: regression models demonstrated a moderate capacity to estimate the burned area, being influenced by data imbalance, while neural networks achieved high accuracy values and low losses, indicating a solid potential in visual identification of fire risk. However, each single approach has constraints related to data volume, processing complexity and the risk of overfitting.

Paper Nr:	172
Title:	Identifying and Understanding Human Values in Text: A Tailorable LLM-Based Architecture
Authors:	Eduardo de la Cruz, Marcelo Karanik and Sascha Ossowski
Abstract:	As intelligent systems become more autonomous, the scientific community focuses on creating decision-making mechanisms that include ethical and moral considerations, unlike traditional utility-maximisation models. To achieve this, a key aspect is assessing how well these decisions align with human values. To this end, a promising line of research is centred on developing approaches based on Large Language Models (LLMs) to identify human values from text, whether explicit or implicit, enabling their recognition throughout. This paper introduces a LLM-based architecture to detect and quantify the intensity of human values in text, avoiding the limitations of previous approaches tied to specific value theory or complex prompt engineering. The architecture comprises three coordinated modules: one that generates structured value specifications from the foundational texts of any theoretical framework; one that labels texts using these specifications; and one that assigns graded support or resistance based on rhetorical and semantic evidence. This modular approach separates the tasks of conceptualising from detecting human values, creating a scalable and reproducible process driven by value specifications adaptable to various theories. The architecture was instantiated with multiple LLMs and evaluated using the ValueEval dataset. The experiments demonstrate good detection performance, confirming the generality of the pipeline.

Paper Nr:	193
Title:	FedCross: Cross-Layer Federated Learning for Reliable and Efficient IoT in Realistic Environments
Authors:	Rihab Saidi, Tarek Moulahi, Mounira Tarhouni and Salah Zidi
Abstract:	Federated Learning has emerged as a key solution for decentralized machine learning, particularly in IoT systems operating across cloud, fog, and edge layers. However, real-world IoT environments pose significant challenges, including communication delays, resource constraints, and device unreliability, which can degrade model accuracy and slow down learning convergence. To address these challenges, we propose FedCross, a novel cross-layer FL framework designed to optimize the success rate of information delivery, reduce communication latency, and maintain high model accuracy across heterogeneous IoT networks. FedCross integrates adaptive communication protocols, hierarchical model aggregation, and dynamic straggler mitigation techniques to ensure robust and efficient learning. Our experiments, conducted on a realistic IoT testbed using two different datasets, demonstrate that FedCross maintains high model accuracy (up to 98.25%), while achieving a 97% information success rate, an average latency of approximately 79.9 milliseconds (ms), and a local transmission delay of about 0.7 ms. These results highlight FedCross’s potential as a practical solution for scalable and reliable FL in complex IoT environments.

Paper Nr:	201
Title:	Ecosystem Simulation of the Amazon Rainforest Survey
Authors:	Shashank Desetti, Disha Deepak Shanbhag, Ananya Kini, Ajaybir Singh and Pooja Agarwal
Abstract:	The Amazon rainforest, the ”lungs of the Earth,” is under unprecedented threat from deforestation, habitat loss, and human activity. Though previous ecosystem simulation has yielded partial results, they tend to suffer from lack of scalability, adaptability, and integrating real-time feedback. This paper introduces the first MARL-based DT for the Amazon Rainforest, an intelligent self-improving ecological simulation that integrates Multi-Agent Reinforcement Learning (MARL), Deep RL, and high resolution satellite data ingestion to provide a scalable, real-time, and adaptive virtual ecosystem. In contrast with traditional static models, our method allows emergent, self-regulating dynamics that reflect actual ecosystem reactions to climate fluctuation, species movement, and human actions. We show how adaptive AI agents can maximize conservation planning by resolving trade-offs between biodiversity protection and sustainable land use. This research paves the way for next generation ecological DTs AI-driven, real-time, and interactive, shifting the way we predict and prevent danger to one of the planet’s most important ecosystems.

Paper Nr:	214
Title:	A Hybrid Architecture Combining Agentic AI and Deep Learning Techniques for Preventing of Malware Cybersecurity Attacks
Authors:	Juan Cañete and Eduardo Guzmán
Abstract:	As the world becomes more digital, cybersecurity threats are becoming more sophisticated and widespread. This poses significant challenges for individuals, organizations, and governments around the world. Among these threats, malware attacks represent a persistent and continuously evolving, capable of causing severe damage to systems, networks, and data, including ransomware with the aim of victim information. This paper investigates the transformative potential of agentic artificial intelligence (AI) in cybersecurity, with a focus on enhancing the detection and response to emerging threats. Our study explore the way to agentic AI can integrate deep learning (DL) techniques to find the best manner to autonomously identify anomalies in system logs and to take proactive decision such as terminate malicious processes if it is necessary. Therefore, this study seeks to optimize DL techniques within agentic AI framework to improve efficiency on detecting and mitigating malware attacks. To validate this approach, we have built processes which leverage empirical data from a real-world attack involving the WannaCry ransomware.

Paper Nr:	227
Title:	A Hybrid Approach to Specify and Evaluate non-Functional Requirements in Organizational-Centred Multi-Agent Systems
Authors:	Issam Bouslimi
Abstract:	The Organization-Centred approach dedicated to the design of Multi-Agent Systems (OCMAS) has been proposed as an alternative to the agent-centred approach to simplify the modelling and implementation tasks of complex systems. The introduction of high-level abstraction concepts such as roles and protocols has spared designers from worrying about implementation details at the design phase. Several Organizational Models (OMs) have been proposed, and various works have focused on specifying and evaluating the Functional Requirements (FR), such as achieving the overall system goal and preventing process blocking. However, few studies have placed the focus on the formalization and the evaluation of the Non-Functional Requirements (NFR) in multi-agent design and less on organization-based MAS, such as flexibility, robustness, efficiency, scalability and communication quality. In the context of our work, based on a Cooperative Information Gathering Simulator (CIGS), we will: i) propose a hybrid approach to specify and evaluate a set of NFRs, and ii) study how the quality of these NFRs evolves during scaling.

Paper Nr:	228
Title:	Optimized Coordination and Performance in UAVs-Cobot Systems for Missions in Dynamic Environment
Authors:	Hafedh Jouini, Hamza Gharsellaoui and Mohamed Khalgui
Abstract:	Unmanned Aerial Vehicles (UAVs) and Collaborative Robots (Cobots) offer a powerful synergy for tackling complex challenges in dynamic mission environments. UAVs excel in aerial surveillance but struggle with navigating complex terrains, while cobots improve ground operations but face coordination issues with UAVs. This paper proposes a hybrid coordination framework that integrates MAVLink2, dynamic scheduling, optimization, and a consensus algorithm to achieve real-time synchronization between UAVs and cobots. The system dynamically adjusts task allocation and coverage areas, improving operational efficiency in complex environments. Simulation results show significant improvements in coverage, reduced operational time, and enhanced task performance compared to traditional methods.

Paper Nr:	233
Title:	Simulating Speciation and Complex Trait Evolution in a Multi-Species Artificial Environment
Authors:	Jay B. Nash, Gary B. Parker, Jim O'Connor and Melanie Fernández
Abstract:	The evolution of new species through reproductive isolation and the development of complex traits are fundamental processes in understanding biodiversity. This research extends previous work on simulating allopatric speciation using genetic algorithms (GAs) with agents that rely exclusively on ”secondary character-istics”-traits unrelated to fitness-to select reproductive partners. Our new simulations introduce multiple species evolving in isolated regions for extended time frames, allowing the evolution of complex traits, such as sight, through incremental mutations. The grid environment was divided into four isolated areas with agents evolving independently for 200,000 time steps before the walls separating them were removed. Results showed that species did not interbreed after speciation, even with frequent interactions. Notably, one simulation resulted in a dominant invasive species out competing others due to food-gathering prioritization, while another run demonstrated long-term coexistence between two species. One species evolved sight, requiring five sequential mutations, but the associated energy costs limited its population size and prevented domination of the environment. These findings underscore the delicate balance between trait advantages and resource competition, demonstrating that speciation and adaptation are tightly interlinked and providing an interesting new look into the ways in which these processes can be simulated and observed.

Paper Nr:	236
Title:	Intelligent Firewall Rule Generation and Harmonization: A Multi-Agent Framework for Cybersecurity Using the NSL-KDD Dataset
Authors:	Noor Saud Abd and Kamel Karoui
Abstract:	The escalating complexity of cyber threats necessitates advanced, automated solutions for network security. This paper presents a novel multi-agent framework for intelligent firewall rule generation and harmonization, leveraging artificial intelligence and data science techniques to enhance cybersecurity. In the case of the NSL-KDD dataset, our approach employs local agents to mine firewall rules for various attack types, including DoS, R2R, and U2R, and a master agent that resolves the rules to remove conflicts and ensure consistency. The rules mined are classified into confident and dubious rules based on a dynamic confidence threshold, and the dubious rules are anonymized for privacy preservation during sharing. Our analysis of 68 rules reveals a dense population of TCP-based attacks, with ports 20, 21, and 23 most targeted, reflecting typical R2L and DoS trends. Graphical representations indicate the distribution of rules per attack type, protocol, and confidence scores, with an average confidence of 0.71. The system natively supports pfSense, enabling real-time rule deployment. Results demonstrate the efficacy of our approach in generating focused, high-confidence rules while addressing scalability and privacy concerns. On going work will incorporate LSTM-based intrusion detection for further refinement in rule accuracy.

Paper Nr:	251
Title:	EMOGAN: Emotion Modeling with Graph Attention Networks
Authors:	Fouad Oueslati, Amira Mouakher and Sahbi Bahroun
Abstract:	Emotion recognition in short and noisy social media text is a challenging yet crucial task for understanding human attitudes, opinions, and affective states in online communication. Such text is typically informal, fragmented, and highly context-dependent, limiting the effectiveness of sequential neural models that rely primarily on linear word order. In this paper, we propose EMOGAN (EMOtion Graph Attention Network), a graph-based multi-label emotion recognition framework that explicitly models relational dependencies in text using Graph Attention Networks. Each text instance is represented as a word-level graph, where nodes correspond to tokens and edges encode syntactic structure and local semantic proximity, enabling structured reasoning beyond sequential representations. To address label imbalance and calibration challenges inherent in emotion datasets, EMOGAN integrates an asymmetric loss function and class-specific threshold optimization. The proposed approach is evaluated on established benchmark datasets with fine-grained emotion labels, hierarchical emotion groupings, and sentiment-level abstractions. Experimental results demonstrate that EMOGAN effectively captures nuanced emotional dependencies, achieves competitive performance across multiple evaluation settings, and remains computationally efficient. Ablation studies further confirm the contribution of graph topology and data balancing mechanisms to robust emotion recognition in noisy short-text environments.
Download

Paper Nr:	276
Title:	Offensive and Defensive Evaluation of Soccer Players Using Tracking Data Derived from Match Videos
Authors:	Ryoya Maejima, Yasuyuki Tahara, Akihiko Ohsuga and Yuichi Sei
Abstract:	Player evaluation based on match data plays a crucial role in soccer analytics. In recent years, methods that assess player contributions by analyzing actions, such as passes and shots have gained considerable attention. However, these action-based approaches are limited to the ball holder, while tracking data that capture the movements of all players during a match is only available to a few professional teams and is generally difficult to obtain. To address this limitation, we propose a novel player evaluation model that generates tracking data from predicted player identification and locational information extracted from broadcast videos, and integrates this data with an existing action-value evaluation model. This approach enables the estimation of player contributions from both action-based and location-based perspectives. Experimental results demonstrate that our method achieves more comprehensive and accurate player evaluations compared to conventional models. Furthermore, validation experiments confirm that the tracking data generated by our framework provide more accurate predictions of player identities and locations than existing methods.

Paper Nr:	284
Title:	Machine Learning-Based Failure Analysis in Industrial Manufacturing: A Comparative Evaluation under Realistic Data Constraints
Authors:	Angela Klassen and Jivka Ovtcharova
Abstract:	Industrial failure analysis (FA) increasingly relies on data-driven approaches, particularly machine learning (ML), to enhance production efficiency, ensure process stability, and support decision-making of human operators. However, these methods face challenges in complex manufacturing environments, including limited data availability, data quality, process dynamics, and data variability across production stations. While prior research has demonstrated the potential of ML algorithms for FA, the impact of real-world data constraints on model performance remains insufficiently explored. This study presents a comparative evaluation of state-of-the-art ML models, including decision tree, random forest, logistic regression, and artificial neural networks, under realistic industrial limitations. Using a real-world dataset from automotive final assembly, we develop an intelligent FA system that leverages ML to assist assembly line workers in interpreting error protocols and identifying root causes. To emulate industrial conditions, targeted perturbation functions are implemented that simulate data scarcity, noise, concept drift, and increased analytical complexity. Across the analyzed perturbation settings, random forest shows highest robustness. Yet, all models degrade substantially under realistic challenges, with drops in the F1-score by over 50 % under heavy drift. These results quantify how real-world data-constraints affect different classifiers and provide guidance for stress-testing models before deployment.

Paper Nr:	288
Title:	Towards Hybrid Summarization: Integrating BART Generation with Genetic Algorithm Optimization
Authors:	Imen Tanfouri, Ghassen Tlik and Fethi Jarray
Abstract:	Automatic text summarization has become an essential task in Natural Language Processing (NLP) to address the increasing volume of digital information. While extractive methods preserve factual accuracy and generative models enhance fluency, combining both remains challenging. In this paper, we propose BART–GA–BART, a hybrid summarization framework that integrates generative diversity, evolutionary optimization, and generative refinement. The process begins with BART, which generates multiple diverse candidate summaries using controlled sampling strategies. These candidates are then optimized through a Genetic Algorithm (GA) that selects the most informative and least redundant combination of sentences based on a multi-objective fitness function combining ROUGE-L, BERTScore-F1, and a redundancy penalty derived from SBERT embeddings. Finally, the best optimized summary is refined again using BART via beam search generation to enhance fluency and coherence. Experimental results on the CNN/DailyMail dataset demonstrate the superior performance of the proposed model, achieving ROUGE-1, ROUGE-2, and ROUGE-L scores of 64.4, 45.6, and 57.6, respectively. These results significantly outperform several recent state-of-the-art models. The findings confirm that combining evolutionary optimization with generative modeling leads to high-quality summaries that are both informative and linguistically coherent, establishing BART–GA–BART as a strong hybrid baseline for future summarization research.

Paper Nr:	307
Title:	LLM-Assisted Modeling of SPARQL Behavior Tree-Based Agents with AJAN
Authors:	Hacane Hechehouche, André Antakli and Matthias Klusch
Abstract:	There are many established Semantic Web standards for implementing multi-agent driven applications. The AJAN framework allows to engineer multi-agent systems based on these standards. In particular, agent knowledge is represented in RDF/RDFS and OWL, while agent behavior models are defined with Behavior Trees and SPARQL to access and manipulate this knowledge. However, the appropriate definition of RDF/RDFS and SPARQL-based agent behaviors still remains a major hurdle not only for agent modelers in practice. For example, dealing with URIs is very error-prone regarding typos and dealing with complex SPARQL queries in large-scale environments requires a high learning curve. In this paper, we present an integrated development environment to overcome such hurdles of modeling AJAN agents and at the same time to extend the user community for AJAN by the possibility to leverage Large Language Models for agent engineering.

Paper Nr:	315
Title:	Deep Feature Fusion Framework with Uncertainty Estimation and Explainable AI for Enhanced Lung Disease Detection
Authors:	Yasmine Ammar, Rim Walha and Fadoua Drira
Abstract:	Early and accurate detection of lung diseases from chest X-ray images remains a critical challenge in medical imaging. In this study, we evaluate a hybrid deep learning framework where the combination of deep models has proven highly effective across multiple applications. Our main contribution lies in the proposed methodology for feature fusion, which leverages the global attention of Vision Transformer with the local representation power of VGG16 to build a more robust diagnostic model. In addition, we integrate uncertainty quantification as a measure of diagnostic reliability, enabling the system to flag low-confidence predictions. To enhance transparency, we incorporate state-of-the-art Explainable AI (XAI) techniques, providing clinically meaningful visual explanations aligned with radiological patterns. Overall, the proposed framework demonstrates that hybrid feature fusion, coupled with uncertainty-aware decision-making and XAI, can significantly improve both the accuracy and interpretability of automated lung disease detection.

Paper Nr:	329
Title:	AutoScene: Automated 3D Urban Scene Generation from Satellite Imagery Using Image Segmentation and Depth Estimation
Authors:	Sameer Beedi, Sanjiv Raghunandan, Jayanth Srinivasan, Kokila P., Richa Sharma and Deepa S.
Abstract:	Modern applications demand large, realistic 3D environments for accurate real-world modeling and visualization. Existing methods for manual 3D environment generation from satellite imagery are slow, costly, non-scalable, and computationally intensive. To address these challenges, this work proposes AutoScene, an automated pipeline for extracting urban features and generating 3D scenes from high-resolution (sub-meter) monocular satellite imagery and Digital Elevation Model (DEM) data. The pipeline employs two fine-tuned multi-encoder U-Net models with pretrained ImageNet weights for instance segmentation, producing georeferenced masks that are converted into vector maps. For 3D reconstruction, a depth estimation model approximates feature heights, which are integrated with DEM data to extrude features and generate realistic 3D scenes. AutoScene was evaluated using multiple datasets: the Massachusetts Roads Dataset (1171 aerial images of 1500×1500 pixels covering 2.25 km² per image), the DeepGlobe Dataset (RGB color-coded class labels), and the AIRS Building Dataset (857 high-resolution 7.5 cm aerial images preprocessed into 512×512 pixel tiles from 10,000×10,000 pixel areas). AutoScene achieved mean Intersection over Union (mIoU) score of 0.75 on the Aerial Imagery for Roof Segmentation dataset and 0.70 on the DeepGlobe and Massachusetts road segmentation datasets using the weighted multi-encoder approach. These results demonstrate significant improvements in segmentation accuracy and reductions in reconstruction time. AutoScene thus enables accurate, efficient, and scalable 3D urban modeling, offering valuable applications in urban planning, large-scale simulation, and disaster management.
Download

Paper Nr:	348
Title:	Thermal Diabetic Foot Ulcer Image Classification with Hybrid Deep Learning Methods
Authors:	Naima Boubaker, Leila Baccour, Boudour Ammar and Ali Wali
Abstract:	Thermal images present important details for detecting minute temperature differences.Their effectiveness lies on representation of ulcer before their appearance on foot skin. This can save many people from diabetic disease by early detection, thus early treatment. Thermal imaging is a low-cost, non-invasive tool of medical diagnostics. In this paper, Efficientnet B0 and deit-tiny transformer are applied to classify two thermal diabetic foot ulcer image data sets. Therefore, hybrid method fusing them is proposed to enhance results. This hybrid approach takes advantage of the CNN’s power in extracting local features and the Transformer’s ability to detect global dependencies. Importantly, the gain comes with minimal computing overhead, making the method ideal for clinical usage. The proposed hybrid achieved high accuracy, 99.73% for the first data and 96.52% for the second data set. Found results show the potential of lightweight hybrid EfficientNetB0 and DeiT-Tiny Transformer for thermal imaging tasks, and they encourage additional research on larger datasets and explainability methodologies to enhance clinician trust.

Paper Nr:	354
Title:	Synergizing Lexical and Semantic Representations for Enhanced Arabic Information Retrieval
Authors:	Khadouj Chadha Amraoui, Yassine Saoudi and Mohamed Mohsen Gammoudi
Abstract:	We present a unified, reproducible evaluation of lexical, dense, and hybrid retrieval methods for Arabic question answering using a normalized FAAQA corpus (extended split, 1,500 Q–A pairs). We compare TF–IDF and BM25 baselines with zero-shot dense retrievers, contrastively fine-tuned dense retrievers, and hybrid (lexical + dense) fusion across three Arabic transformer backbones: AraBERT, ARBERT, and CAMeLBERT. Results show that contrastive fine-tuning is essential: zero-shot dense encoders underperform BM25 (Top-1 in the single digits), while fine-tuned dense retrievers reach substantially higher scores (Top-1 up to ≈ 0.76 and MRR ≈ 0.86 for CAMeLBERT). Hybrid fusion improves robustness for particular query types, but on normalized text, well-fine-tuned dense models often match or exceed hybrid Top-1 performance. The experimental pipeline, evaluation scripts, and results artifact are documented and available upon reasonable request.
Download

Paper Nr:	356
Title:	Deep Hybrid Conv2D–Swin Transformer Network for Resilient and High-Fidelity Video Watermarking
Authors:	Souha Mansour, Saoussen Ben Jabra and Ezzedine Zagrouba
Abstract:	The exponential growth of digital video content on online platforms raises critical challenges for copyright protection and source authentication. In this work, we propose a novel frame-by-frame video watermarking approach that combines 2D convolutional networks with the Swin Transformer. While Conv2D layers efficiently capture local spatial structures for watermark embedding, the Swin Transformer models hierarchical global dependencies, enabling more robust extraction under complex distortions. The proposed method embeds a binary watermark into individual frames and extracts it using a symmetric hybrid architecture. To ensure both invisibility and robustness, a hybrid loss function is employed, integrating perceptual similarity and reconstruction fidelity. Experimental evaluations on benchmark video datasets demonstrate that our approach achieves high imperceptibility (PSNR/SSIM) while maintaining strong resilience against compression, noise addition, geometric attacks, and scaling. These results highlight the potential of hybrid Conv2D–Transformer architectures for advancing deep learning-based video watermarking, while serving as a proof-of-concept toward practical copyright protection scenarios.

Paper Nr:	360
Title:	Learning What Matters for Effective Historical Manuscripts Recognition
Authors:	Asma Kharrat, Fadoua Drira, Franck Lebourgeois and Bertrand Kerautret
Abstract:	Historical Handwriting Text Recognition (HHTR) has long posed significant challenges due to the high variability in writing styles, the degradation of historical documents, and the scarcity of labeled training data. While commercial systems have made notable progress in addressing some of these issues, their performance often remains limited when applied to diverse historical scripts or low-resource scenarios due to a passive randomly sampled dataset for learning. In recent years, efforts to enhance data efficiency have led to the development of smarter annotation strategies and the application of active learning techniques, enabling models to learn effectively from fewer annotated examples. At the same time, the emergence of Large Language Models (LLMs) has opened new avenues for contextual understanding and intelligent error correction support. This paper provides a comprehensive overview of the evolution of HHTR systems, with a particular focus on methods aimed at minimizing annotation effort. We examine selective annotation methodologies, explore Continual learning strategies, and analyze the growing role of LLMs in supporting transcription and recognition tasks. Finally, we highlight ongoing challenges and outline promising future directions for building intelligent, low-resource, and interactive systems for historical document analysis.

Paper Nr:	365
Title:	Is Your AI Model Secure? A Study of Data Poisoning Attacks on AI Models
Authors:	Saad El Jaouhari, Nouredine Tamani and Lina Ferial Benassou
Abstract:	As Artificial Intelligence (AI) systems have become increasingly integral to applications spanning healthcare, finance, autonomous vehicles, and critical infrastructure, securing models against adversarial threats has become essential. Among these, data poisoning-where adversaries tamper with training data to corrupt or subvert model behavior-poses a distinctive risk by targeting the learning phase rather than inference, degrading accuracy, manipulating predictions, or implanting future backdoors. Given the safety-critical nature of many deployments, understanding and mitigating poisoning risks is paramount for reliable Machine Learning (ML). This paper surveys data poisoning attacks across supervised, unsupervised, and semi-supervised learning, alongside corresponding defenses, and consolidates the literature into a taxonomy to aid practitioners. In addition, we propose a model-agnostic audit methodology for assessing robustness to data poisoning that operationalizes threat modeling, clean baselining, a representative attack suite, robustness metrics with confidence intervals, detectability screening, defense evaluation, and governance-oriented reporting. Together, the taxonomy and audit methodology provide a practical foundation for auditing AI models and selecting defenses tailored to specific use cases, while highlighting open gaps, particularly the limited effectiveness of current defenses and the field’s focus on neural classifiers at the expense of other model families, that motivate the development of more robust, broadly applicable protections.

Paper Nr:	366
Title:	An Ontology-Based Model for the Validation and Integration of Patient-Generated Health Data (PGHD) into Clinical Workflow
Authors:	Ahmed Dridi, Aroua Taamallah, Salma Sassi and Richard Chbeir
Abstract:	Integrating Patient-Generated Health Data (PGHD) into clinical workflows is challenging due to heterogeneous formats/units and the lack of systematic validation. VIDA-PGHD is a modular ontology that turns PGHD into trustworthy, workflow-ready information via three linked modules: Validation, Decision, and Action. It reuses HL7 FHIR, W3C PROV-O, SSN/SOSA, and ODRL, adding explicit semantics for dataquality checks, multi-criteria decision logic, and policy-governed action orchestration. Implemented in OWL 2 (≈180 classes, 255 object properties) with 57 monotonic, DL-safe SWRL rules, it preserves full provenance from observation to intervention and outcomes. Evaluation in Proteg´ e with the Pellet reasoner and SPARQL competency questions confirms logical consistency and that required inferences (e.g., outcome aggregation, decision triggers, policy authorization) are correctly derived. Overall, VIDA-PGHD elevates PGHD from raw device streams to explainable, actionable clinical intelligence.

Paper Nr:	370
Title:	From First Price to All Pay: Smooth Transfer Learning in Sealed-Bid Auctions
Authors:	Luis Eduardo Craizer, Edward Hermann and Moacyr Alvim Silva
Abstract:	We investigate smooth transfer learning in multi-agent reinforcement learning, applied to sealed-bid auctions. Instead of abruptly switching environments, we gradually vary them so that agents can adapt as incentives change over time. As a concrete setup, we move from the first price auction to all-pay auction along a payment parameter t, using the fact that both formats share the same sealed-bid structure and differ only in how non-winners pay. We first train to convergence in the first price auction and then increase t in small steps toward all-pay, so the reward changes smoothly and the policies adapt continuously. Across private value settings, smooth transfer speeds up convergence in the target format and improves policy quality relative to training from scratch, yielding policies closer to theoretical best responses, mitigating local equilibrium failures common in all-pay when trained from scratch, and achieving faster stabilization and higher returns. Gains persist for moderately large agent counts, although margins decrease as N increases. These results support smooth transfer as a practical route to scalable learning in auction environments and complement prior work on stepwise transfer over the number of agents.
Download

Paper Nr:	374
Title:	LLaVA-EX: Cross-Domain Fine-Tuning and Explainable Multimodal AI Framework for Robust and Trustworthy Social Media Understanding
Authors:	Feriel Gammoudi and Mohamed Nazih Omri
Abstract:	Multimodal large language models (MLLMs) such as LLaVA achieve strong performance on curated vision– language benchmarks, yet their effectiveness degrades on real-world social media data characterized by noise, platform-specific styles, and distributional shifts. In addition, their predictions are often difficult to interpret, limiting their applicability in domains where transparency is required. To address these challenges, we propose LLaVA-EX, an extension of LLaVA that jointly improves robustness across domains and multimodal explain-ability. LLaVA-EX introduces two lightweight and complementary components. First, a Domain Adaptation Layer (DAL) aligns visual–textual representations across social media platforms using contrastive learning and Maximum Mean Discrepancy (MMD) regularization. Second, an Explainability Module (EM) combines attention rollout and Grad-CAM-based attribution to produce coherent multimodal explanations. Both modules are compatible with parameter-efficient fine-tuning and can be integrated without modifying the core LLaVA backbone. We evaluate LLaVA-EX on three heterogeneous datasets-Twitter100K, Flickr30K, and Instagram-UserSet-under single-domain and cross-domain settings. Experimental results show consistent improvements over fine-tuned LLaVA baselines in BLEU, ROUGE-F1, and CLIPScore. In addition, explanation quality is analyzed using a Visual Explainability Index (VEI) and a complementary human evaluation. These results suggest that robustness to domain shifts and interpretability can be jointly addressed within a unified multimodal framework.

Paper Nr:	388
Title:	CANDACE: A Context-Aware Email Generation Framework
Authors:	Francisco Cardoso, Eva Maia and Isabel Praça
Abstract:	Phishing attacks remain a significant security challenge, with email as the primary attack vector. Conventional detection models are increasingly failing as attackers shift from generic emails to highly personalized and context-specific messages. The performance of these systems is limited by the scarcity of specialized, domain-specific training data for recognizing such threats. To address this issue we introduce CANDACE, a modular framework designed to generate context-aware synthetic email messages to improve these detection systems. The main innovation of CANDACE comes from its dual Knowledge Graph (KG) architecture, which gives the generation process a contextual foundation. The first KG maps external, real-world information about an organization, while the second models its internal structure, such as employees and projects. A Small Language Model (SLM) then uses the information of these KGs, with other important components, such as URL, to generate an email message that is contextually relevant to the domain of the organization. Our work shows that by using the dual KG architecture of CANDACE, it is possible to generate high-fidelity, context-aware synthetic data. This approach not only addresses the scarcity of specialized training data but also significantly enhances the capability of detection systems to identify and neutralize sophisticated, personalized phishing attacks.
Download

Paper Nr:	396
Title:	Uncertainty Calibration of Multi-Label Bird Sound Classifiers
Authors:	Raphael Schwinger, Ben McEwen, Vincent S. Kather, René Heinrich, Lukas Rauch and Sven Tomforde
Abstract:	Passive acoustic monitoring enables large-scale biodiversity assessment, but reliable classification of bioacous-tic sounds requires not only high accuracy but also well-calibrated uncertainty estimates to ground decision-making. In bioacoustics, calibration is challenged by overlapping vocalisations, long-tailed species distributions, and distribution shifts between training and deployment data. The calibration of multi-label deep learning classifiers within the domain of bioacoustics has not yet been assessed. We systematically benchmark the calibration of four state-of-the-art multi-label bird sound classifiers on the BirdSet benchmark, evaluating both global, per-dataset and per-class calibration using threshold-free calibration metrics (ECE, MCS) alongside discrimination metrics (cmAP). Model calibration varies significantly across datasets and classes. While Perch v2 and ConvNeXtBS show better global calibration, results vary between datasets. Both models indicate consistent underconfidence, while AudioProtoPNet and BirdMAE are mostly overconfident. Surprisingly, calibration seems to be better for less frequent classes. Using simple post hoc calibration methods we demonstrate a straightforward way to improve calibration. A small labelled calibration set is sufficient to significantly improve calibration with Platt scaling, while global calibration parameters suffer from dataset variability. Our findings highlight the importance of evaluating and improving uncertainty calibration in bioacoustic classifiers.
Download

Paper Nr:	405
Title:	Self-Healing OCR Pipelines: A Modular Defense Framework against Visual and Semantic Attacks
Authors:	Eya Ben Hmida, Sarra Abidi and Leila Ben Ayed
Abstract:	Optical Character Recognition (OCR) pipelines—both classical (Tesseract) and Large Language Model-based (Donut, TrOCR)—remain vulnerable to coordinated attacks: visual degradation (blur, low-ink, print-scan), structural manipulation (layout-swap), and semantic threats (prompt-injection). This paper proposes a modular defense framework comprising five orthogonal agents: Preflight (image quality analysis), Layout-Guard (schema validation), Semantic-Verifier (rule-based and Retrieval-Augmented Generation), LLM-Shield (defensive prompting), and Arbiter (multi-OCR voting). Evaluated on SROIE (receipts) and FUNSD (forms), the framework achieves: (1) Attack Success Rate (ASR) reduction from 95% to 22% (-73 percentage points), (2) field-level F1 improvement of +8–15 points, (3) acceptable latency trade-off of 2.85 s per document for critical workflows. Ablation studies confirm modularity and orthogonality. We release reproducible artifacts and provide the first systematic integration of prompt-injection defenses for LLM-OCR.

Paper Nr:	413
Title:	A Novel Hybrid Framework Leveraging SDN, PPO, and Game Theory for Bandwidth and Congestion Management in VANETs
Authors:	Nouha Alyaoui, Awatef BenFradj Guiloufi, Ghada Tlili and Karim Chabir
Abstract:	The proliferation of autonomous vehicles demands real-time, intelligent mechanisms for bandwidth and congestion management in highly dynamic Vehicular Ad-Hoc Networks (VANETs). However, system performance is often degraded by the selfish behavior of individual vehicles, which optimize local utility at the expense of global network efficiency and fairness. Existing solutions based on Software-Defined Networking (SDN) or Deep Reinforcement Learning (DRL) typically overlook these strategic interactions, leading to suboptimal and inequitable resource allocation. This paper introduces a hybrid SDN–PPO–GT framework that models bandwidth and congestion management as a non-cooperative game and embeds this formulation into a Proximal Policy Optimization (PPO) reward function combining Aggregate Travel Time (ATT), Jain’s Fairness Index, and a congestion cost. A centrally located PPO agent within the SDN controller learns to adjust packet prioritization, traffic shaping, and load-balancing rules to steer the system away from inefficient Nash equilibria toward socially optimal operating points. Large-scale VANET simulations show that the proposed SDN–PPO–GT approach reduces ATT by more than 10 % and raises fairness above 0.92 compared with a pure SDN–PPO baseline, while maintaining a Packet Delivery Ratio above 0.96 across a wide range of traffic densities.

Paper Nr:	424
Title:	A New AI System for Elderly Human Activities Analysis Based on Vision Transformer
Authors:	Neziha Jaouedi, Noureddine Boujnah, Mounira Hmayda and Med Salim Bouhlel
Abstract:	The global population is aging rapidly, increasing the demand for technologies that support independent living for the elderly. Automated monitoring using Artificial Intelligence (AI) offers a promising solution. This paper presents a novel vision-based architecture for elderly activity analysis. Our approach leverages a dual-feature strategy: we extract spatio-temporal features using the TimeSformer model and enhance human representation with 2D skeleton features obtained via pose estimation. These combined features are utilized then by a Vision Transformer (ViT) model for robust action recognition and classification. The proposed system is evaluated on the large-scale ETRI-Activity3D dataset, containing 55 daily activities. Experimental results demonstrate the effectiveness of our method, achieving a competitive average accuracy of 95.4% and outperforming several state-of-the-art approaches in recognizing critical activities for assisted living.

Paper Nr:	430
Title:	Explainable Graph-Neural Architectures for ICU Mortality Prediction Using Logic Explained Networks
Authors:	Dan-Stefan Damian
Abstract:	The increasing digitalisation of intensive care has led to vast repositories of patient data that remain underexploited for transparent clinical decision support. In this work, we present an explainable deep-learning framework that combines Logic Explained Networks (LENs) with Graph Neural Networks (GNNs) for predicting 28-day mortality in mechanically ventilated ICU patients. Each patient is represented as a node in a similarity graph constructed from physiological and clinical variables, enabling the GNN to capture relational dependencies between comparable cases. The LEN component augments the neural model with logic-based reasoning layers, producing human-interpretable first-order rules that approximate network predictions. Experiments on the MIMIC-II clinical dataset demonstrate that integrating relational structure improves mortality prediction accuracy (≈ 0.9) compared to traditional baselines, while LEN-derived rules achieve high fidelity (≈ 0.7) and align with known clinical risk factors such as age, organ dysfunction, and comorbidities. These results suggest that symbolic interpretability and deep relational learning can coexist in a unified framework, offering both strong predictive performance and transparent reasoning for critical care decision support.
Download

Paper Nr:	442
Title:	Restaurant Add-on Order Recommendation Using Dynamic Item Bias Modeling Based on Ordering Time Context
Authors:	Atsuko Mutoh, Naoya Yoshida, Kosuke Shima, Koichi Moriyama, Tohgoroh Matsui and Nobuhiro Inuzuka
Abstract:	In restaurant environments, customers often place multiple orders during a single visit, such as additional drinks or desserts after the main meal. However, most existing recommender systems treat these orders as a single transaction, overlooking the temporal and contextual dynamics within a dining session. This study proposes a novel add-on order recommendation method that models item bias dynamically based on the ordering time context. The proposed model adjusts reference data and item relevance according to the ordering position (e.g., first, second, or third order) and the elapsed time within the session, thereby capturing phase-dependent customer behavior. Using real-world restaurant data, we conducted experiments that varied contextual factors such as meal progress and order timing. The results show that the proposed dynamic item bias model outperforms baseline collaborative filtering and static context models in predicting add-on orders, particularly in later phases of dining sessions. These findings demonstrate that explicitly modeling intra-session temporal structure provides meaningful insight into the temporal shift of customer needs and supports more adaptive real-time recommendation in restaurant systems.
Download

Paper Nr:	443
Title:	AISTIP: AI Security Threat Intelligence Platform to Gather Knowledge from Technical Documents
Authors:	Kento Hasegawa and Seira Hidano
Abstract:	As AI research advances, the negative impacts, such as adversarial examples and data poisoning, are increasingly becoming apparent. Research on AI security is progressing rapidly, making it challenging for researchers and engineers to keep track of these developments. In this paper, we propose AISTIP, an AI Security Threat Intelligence Platform that collects and analyzes technical documents related to AI security. AISTIP features a crawling function that automatically gathers papers from the internet and a labeling function that annotates the collected papers. It employs a two-pronged approach for labeling: one based on known labels from a database and the other based on new labels derived from large language models (LLMs). This enables accurate labeling and adaptability to new labels in a rapidly evolving field. Evaluation experiments demonstrate that it can effectively handle new labels compared to existing labeling methods.

Paper Nr:	447
Title:	A Lightweight Solution for Pose-Based Recognition for Isolated Spanish Sign Language Using Recurrent Models
Authors:	Gerardo León-Quintana, José Salas-Cáceres and Javier Lorenzo-Navarro
Abstract:	This work addresses the problem of Isolated Sign Language Recognition (ISLR) in Spanish Sign Language (LSE) from a pose-based perspective. The proposed approach relies on 3D landmark extraction using Google’s MediaPipe framework to obtain face, hand, and upper-body keypoints, which are then normalized and transformed into spatial–temporal feature sequences. Two temporal alignment strategies, average sampling and max-length padding, were implemented to ensure uniform input dimensions across samples. Bidirectional recurrent neural networks (Bi-LSTM and Bi-GRU) were evaluated to capture the temporal dependencies inherent to signing. Experimental results on the LSE-Health-UVigo dataset show that the Bi-LSTM architecture combined with the Focal Loss function (γ = 3) achieved the highest performance, reaching 79.8% unweighted accuracy. The proposed model has an average response time of approximately 1 ms, making it suitable for deployment in real-time scenarios. These results highlight the effectiveness of pose-based recurrent architectures for ISLR and demonstrate the potential of lightweight models for robust sign language understanding.
Download

Paper Nr:	466
Title:	Blockchain and IoT for Transparency and Traceability in Supply Chain Management
Authors:	Elias-Nelu Dejan, Mara Hajdu Măcelaru and Petrică Pop
Abstract:	Clarity, reliability, and ease of traceability of the supply chain are major concerns in the global transportation and distribution of goods. People want to know that products are being sourced correctly, sustainably, and safely. However, current digital systems have problems: data is often scattered, systems do not communicate well with each other, and it is difficult to verify the authenticity of information.In this paper, we propose a system that uses blockchain technology and the Internet of Things (IoT) together to make the supply chain more transparent and secure. The model uses a combination of data stored on the blockchain (for important events and digital certificates, which cannot be modified) and cloud databases (for daily information).IoT sensors automatically store environmental data, such as temperature and location, to ensure that products are transported and stored correctly. The system was created as a web prototype using Node.js, MongoDB, smart contracts on Ethereum, and Vue.js. In conclusion, the combination of blockchain and IoT technologies provides a solid foundation for the next generation of supply chain monitoring systems. In the future, we will also add solutions based on artificial intelligence to automatically detect problems and improve logistics.

Paper Nr:	468
Title:	Don’t Compose Blindly: Adaptive GNNs for Opacity Verification in Modular Discrete-Event Systems
Authors:	Salwa Habbachi, Imene ben Hafaiedh and Zhiwu Li
Abstract:	Language-based opacity verification in modular discrete-event systems (DESs) is hindered by state-space explosion when constructing a global observer. Iterative compositional methods alleviate this issue but often rely on fixed composition sequences, leading to redundant computations when modules are locally non-opaque. We propose an adaptive framework using Graph Neural Networks (GNNs) to dynamically select optimal composition sequences. The GNN prioritizes compositions based on structural features, including inter-module coupling, transition patterns, and secret state distributions, reducing unnecessary computation. Experiments on synthetic and real-world benchmarks demonstrate an average speedup of 15.7×, a 68% reduction in compositions, and 99.8% verification accuracy. This approach enables scalable opacity verification for systems with up to 30 modules.

Paper Nr:	470
Title:	Voice Mitigates Expectation Deficiency: A Modality Comparison with Theory-Aligned Subjectivity Metrics in Co-Creative Image Generation
Authors:	Ryuki Matsuoka and Michita Imai
Abstract:	Recent advances in large language models (LLMs) and image generation technologies have enabled co-creation based on natural language input. However, particularly in co-creative tasks, verbalizing ambiguous and subjective expressions remains challenging, and a gap persists between generated results and users’ mental imagery. The originality of this study lies in conceptualizing this phenomenon as Agent Sensitivity Understanding Expectation Deficiency (ASUED) and empirically validating it through theory-aligned measurement metrics (SCA-L/SSI-L). We examine how the input modality difference between voice and text affects the use of subjective language in image generation tasks. Results from a co-creative experiment with 20 participants revealed that voice input significantly increased the use of subjective linguistic resources such as self-involvement, certainty, and degree adjustment compared to text input, with large effect sizes. This difference was consistently observed at both utterance (sentence) and participant levels. Furthermore, this difference can be interpreted as consistent with a causal effect whereby voice input increases subjective expression, yielding design implications that contribute to mitigating expectation deficiency in co-creation.

Paper Nr:	478
Title:	A Hybrid Lempel-Ziv-Welch Based Approach for Lossless-Lossy Color Image Compression Scheme
Authors:	Taif Ammash, Hamza Gharsellaoui and Leila Ben Ayed
Abstract:	In this paper, the authors present a color image compression hybrid approach that can be used either for lossless or lossy compression requirements. This proposed approach contains is based on two new contributions. The Lempel-Ziv-Welch approach which used as a lossless data compression algorithm that works by reading a sequence of symbols, grouping the symbols into strings, and converting the strings into codes. The Huffman technique, on the other hand, though efficient for images with uniform pixel distributions, leading to suboptimal compression results. Simulation results show that the proposed schemes provide significant improvements over previously published work for both the lossy and the lossless components.

Paper Nr:	484
Title:	LexiGen: A Fine-Grained Synthetic Corpus for GDPR Compliance Analysis
Authors:	Ali Bouhejba, Aroua Hedhili Sbaï, Souheib Yousfi and Layth Sliman
Abstract:	Privacy protection has become a critical global concern, leading to stringent regulations like the General Data Protection Regulation (GDPR). These rules force companies to document their data practices in privacy policies, but ensuring ongoing GDPR compliance is complex and expensive. To address this, we introduce the first synthetic corpus for detailed, article-level GDPR compliance analysis, generated using Large Language Models (LLMs). This corpus captures realistic legal language and obligations, providing a reusable benchmark for NLP tools. We demonstrate its utility through a comparative study of encoder-based models (BERT, Legal-BERT) and decoder-based models (Mistral-7B, its legal adaptations, and GPT-OSS-20), analyzing the impact of model scale on classification performance. Our results establish that encoder models deliver strong, stable performance, making them a reliable baseline when framing compliance as a three-class task (COMPLIANT, NON COMPLIANT, NOT APPLICABLE). This work positions the corpus as a practical foundation for future research in automated legal compliance.

Paper Nr:	493
Title:	LoDA-BERTopic: Domain-Adaptive Neural Topic Modeling with Low-Rank Fine-Tuning for Social Web Discourse
Authors:	Wadie Kadri, Souhir Bouaziz and Maha Charfeddine
Abstract:	Understanding evolving discussions on the Social Web is central to search and data mining applications, particularly in domains where terminology and semantics are highly specialized, such as climate change, public health, and technological innovation. Traditional topic modeling methods, including LDA, SeqLDA, and standard BERTopic, often struggle to capture the nuanced, domain-specific language in these contexts. We present LoDA-BERTopic, a domain-adaptive neural topic modeling framework that integrates Low-Rank Adaptation (LoRA) for fine-tuning the all-MiniLM-L6-v2 sentence transformer. By enhancing semantic representations through parameter-efficient adaptation, LoDA-BERTopic generates more coherent and contextually meaningful topics while keeping computational overhead low. Comparative experiments against BERTopic variants using all-mpnet-base-v2 and all-distilroberta-v1 embeddings, as well as traditional baselines, demonstrate significant gains in topic coherence and diversity. Evaluations on Reddit discussions of climate change reveal the effectiveness of LoDA-BERTopic for mining high-quality topics from noisy, user-generated content. Our results highlight the value of parameter-efficient fine-tuning in advancing topic modeling for complex social web data, offering both principled and practical improvements for web mining and information retrieval tasks.
Download

Paper Nr:	495
Title:	A Fuzzy Intelligent Adaptive Tutoring System with Explainable Recommendations: Case of Math Learning for Elementary Students
Authors:	Mounira Chkiwa and Mohamed Moez Chkiwa
Abstract:	This paper presents a novel Fuzzy Intelligent Adaptive Tutoring System (FIATS) designed to deliver personalized and explainable mathematics games to elementary school students aged 6–12. The system integrates a rich knowledge base of 520 mathematical questions and 180 interactive learning games, each semantically tagged with a topic (e.g., addition, fraction, angle) and a developmentally appropriate difficulty level (easy, medium, hard). A custom fuzzy inference system processes a child’s age and quiz performance to determine an optimal, continuous difficulty score. This score, combined with an analysis of the child’s failed topics, drives a transparent recommendation engine that suggests remedial activities. The system exemplifies core principles of Explainable AI (XAI) and Interpretable AI, as every recommendation is accompanied by a clear rationale linking the child’s errors to targeted practice. Implemented as a standalone, web-based application, this work demonstrates a practical, pedagogically sound, and transparent intelligent system at the intersection of Soft Computing, Knowledge Representation, and Intelligent User Interfaces.

Paper Nr:	497
Title:	Leveraging YOLO for AI-Powered Image-Based Plant Disease Detection in Sustainable Agriculture
Authors:	Dhouha Belghith, Abir Baâzaoui and Walid Barhoumi
Abstract:	To support environmental sustainability and ensure agricultural security, early detection and diagnosis of plant diseases are vital. Conventional methods, which rely heavily on human diagnosis, are often inadequate due to their labor-intensive, subjectivity, and time-consuming nature. To mitigate these limitations and protect crops effectively, artificial intelligence offers a powerful solution for early and accurate detection of plant diseases using image-based analysis. This paper contributes to sustainable agriculture by deploying an optimized YOLO V11 model for plant disease detection. The designed model was refined using the prominent PlantVil-lage dataset through a dedicated pipeline of balanced data augmentation and fine-tuning. This domain-specific calibration effectively addresses dataset imbalance and enhances generalization, particularly for rare disease categories, yielding a robust mean average precision (mAP@0.5) of 99.2%, even with limited and heterogeneous data. Comparative analysis shows that the proposed balanced training approach yields 19% and 15% improvements compared to ResNet50 and EfficientNetV2B0, respectively. Thus, this study emphasizes the domain-driven adaptation and calibration of YOLO V11 for a sustainable and a scalable plant disease detection.
Download

Paper Nr:	498
Title:	PACO-FS: Pareto-Based Ant Colony Optimization for Feature Selection
Authors:	Ines Alaya, Jamila Sassi and Jihen Balhoudi
Abstract:	Data classification is fundamental in data analysis, but increasing data volume often leads to more complex classification tasks. To tackle this, effective dimensionality reduction techniques are essential to identify the most relevant features and simplify the input space. In this work, we introduce PACO-FS, a novel multi-objective feature selection method based on Ant Colony Optimization (ACO) that leverages Pareto optimality to balance multiple criteria simultaneously. Our approach is designed to be flexible and applicable across various domains. We conducted extensive experiments on multiple datasets, and the results consistently demonstrate that PACO-FS achieves strong performance in selecting representative feature subsets, leading to improved classification outcomes.

Paper Nr:	501
Title:	Ontology-Based Framework for Real Estate Investment Recommendation through Expert Knowledge and Market Data
Authors:	Sang Vu, Quang Thi, Thuan Nguyen, Diem Nguyen and Hien D. Nguyen
Abstract:	This paper introduces an ontology-based real estate recommendation framework that integrates knowledge representation with machine learning to deliver personalized and context-aware property suggestions. The proposed system, named REstInves-model, is developed upon the relational knowledge framework, Rela-model, to combine expert knowledge and real estate market data for intelligent decision-making. An ontology is constructed to formally represent essential real estate concepts such as location, financial characteristics, environmental conditions, enabling structured knowledge modeling and intelligent inference. The combination of survey-based customer behavior data and real estate listings is utilized for analysis. Customer groups are identified using the K-Means clustering algorithm, while inference rules are employed to generate property recommendations tailored to each segment. Experimental results, based on over 7,000 properties collected from real estate platforms in Vietnam, demonstrate the system’s capability to provide suitable suggestions with users’ financial situations and preferences. The model bridges the gap between data-driven and knowledge-based approaches, contributing an interpretable framework for real estate investment decision support.

Paper Nr:	506
Title:	Asynchronous Signaling in Spiking Neural Networks: Enabling On-Chip Learning with Built-In Temporal Dynamics
Authors:	Hao Yu, Nancy Fulda and Jordan Yorgason
Abstract:	High-fidelity modeling of complex neural processes requires algorithms that mimic brain properties including spike-time dependent plasticity, temporal attributes in the synaptic cleft, and the influence of modulatory neurotransmitters such as dopamine and GABA. In support of this, we propose a novel machine learning algorithm, Asynchronous Spike Neural Network (A-SNN), with temporal learning capacities and a biologically inspired weight update algorithm. The algorithm is appropriate for on-chip training, with low space consumption while maintaining latency. We validate its effectiveness on two small-scale learning tasks and show that it achieves better performance than the well-known LSTM temporal learning algorithm, and also outperforms a network trained using Spike Timing Dependent Plasticity (STDP). Our algorithm shows promise on small-scale tasks, however, additional development is needed before it can scale to more complex data.

Paper Nr:	515
Title:	Deep Learning Architectures for 3D Point Clouds
Authors:	Nour Neifar, Achraf Ben-Hamadou and Ahmed Rekik
Abstract:	Point cloud learning has gained increasing attention due to its wide range of applications across various fields. However, the irregular and unordered nature of point clouds presents unique challenges for deep neural networks. In recent years, various deep learning approaches have been proposed to address these challenges, resulting in significant progress in both methodology and performance. This paper presents a comprehensive review of deep learning methods for point cloud processing, as well as an overview of the underlying challenges associated with this representation. It analyzes the evolution from early architectures to recent state-of-the-art models, highlighting their core mechanisms, strengths, and limitations. Through this synthesis, we aim to guide researchers in understanding point cloud learning while highlighting current challenges and providing perceptive observations, and suggesting future research directions in 3D point cloud processing.

Paper Nr:	531
Title:	Learning User Similarity from Heterogeneous Implicit Feedback in Recommender Systems
Authors:	Ho Thi Hoang Vy, Tiet Gia Hong, Vu Thi My Hang, Cuong Pham-Nguyen and Le Nguyen Hoai Nam
Abstract:	In collaborative filtering recommender systems, accurately measuring user similarity is fundamental to producing relevant and personalized recommendations. This task becomes particularly challenging in digital platforms, where explicit feedback, such as numeric ratings, is sparse, and implicit feedback, such as views, add-to-carts, and purchases, serves as the primary indicator of user preference. However, conventional methods often treat these diverse types of implicit feedback uniformly, ignoring their varying indicative strengths. To address this limitation, we propose two complementary models for learning user similarity from implicit feedback. The first model transforms heterogeneous implicit feedback into inferred explicit feedback. The second model learns separate embeddings for each implicit feedback type and integrates them into unified user representations. In both models, implicit feedback types are weighted according to their conversion rates, effectively capturing their influence on user preferences. Experiments on three datasets demonstrate that our methods consistently improve recommendation accuracy over baseline approaches, and a clustering-based implementation ensures scalability in large-scale environments.

Paper Nr:	539
Title:	Meta Deep Reinforcement Learning Based on Supervised Learning of Correspondence between Model Parameters and Reward Functions as Externally Conditioned Queries
Authors:	Takumi Kuitani, Hiroyuki Sato and Keiki Takadama
Abstract:	This paper focuses on Meta-Reinforcement Learning (MetaRL) towards zero-shot learning adaptable to unseen tasks and proposes Reward to Parameter MetaRL based on Supervised Learning (R2P-MetaRL (SL)), which learns the correspondence between model parameters and reward functions as external conditions (queries) through a latent variable space. In particular, the supervised learning in R2P-MetaRL (SL) minimizes the loss of the difference in the latent space between a reward Ri in the i-th task and the model parameters Wi of RL in the i-th task. Applying R2P-MetaRL (SL) to a maze problem has revealed the following implications. (1) The number of learning epochs of R2P-MetaRL (SL) is markedly reduced, and the median number of learning epochs on test tasks becomes 0, indicating that the proposed method largely achieves zero-shot learning on unseen tasks. (2) A loss that minimizes the directional difference between Ri and Wi in the latent space reduces the number of learning epochs while suppressing extremely large epochs in specific tasks, compared with the loss that minimizes both directional and norm differences. (3) A loss that minimizes the difference between Ri and Wi while maximizing the difference between Ri and Wi̸=j in the latent space reduces the variance of learning epochs compared with the loss that only minimizes the difference between Ri and Wi in the latent space.

Paper Nr:	540
Title:	Intelligent Cybersecurity Assessment for Smart-Home IoT Systems: A Rule-Based Integrated Pentest Tool Approach
Authors:	Bilel Arfaoui, Hichem Mrabet and Abderrazak Jemai
Abstract:	The rapid expansion of smart-home IoT technologies has intensified cybersecurity concerns, highlighting the need for reliable assessment methods capable of detecting and addressing device vulnerabilities. In this study, we present an intelligent, rule-based cybersecurity assessment framework that consolidates three widely used penetration-testing tools, Nmap, Binwalk and John the Ripper, into a single, coherent workflow for evaluating IoT devices. The framework conducts structured network analysis, firmware examination and password testing, which together expose weaknesses across several potential attack vectors. Unlike improvised or manual testing approaches, the proposed solution automates the entire analysis process through a structured rule-based workflow, enabling faster and more consistent identification of security flaws while limiting human intervention. Each tool is integrated in a way that takes advantage of its specific capabilities, resulting in a more precise and efficient vulnerability detection process through cross-layer result correlation. The output of the assessment offers practical and actionable recommendations that can be applied directly to strengthen device configurations or adjust network settings. The framework is designed with the particular characteristics of IoT environments in mind, including complex firmware structures and heterogeneous communication protocols, which makes it especially suitable for smart-home systems. Its relevance lies in its ability to bridge the gap between academic security research and the operational challenges faced in real-world smart-home deployments. Experimental testing confirms that the combined toolset is effective in uncovering critical vulnerabilities, improving detection coverage and reducing assessment time compared to standalone tool execution, ultimately contributing to a stronger security posture within smart-home ecosystems. More broadly, this work advances ongoing discussions on IoT security by providing a scalable, automated and tool-agnostic assessment workflow that delivers prioritized and actionable remediation guidance.

Paper Nr:	555
Title:	Dual-Objective Federated Learning Strategy for Lung X-Ray Analysis in Healthcare Systems
Authors:	Antoni Jaszcz, Katarzyna Prokop, Piotr Żerdziński, Dawid Połap and Jakub Siłka
Abstract:	Federated Learning has emerged as a promising approach to collaboratively train machine learning models across decentralized healthcare systems with big data while maintaining the privacy of patient data. In this paper, we propose using the joint model to segment and classify lung X-ray images with a federated learning strategy. By eliminating the need for two distinct models, we aim to improve the system’s maintainability while also directly using the output and encoding backbone of the segmentation model in the classifier segment. The presented experiments describe the initialization, training and fine-tuning of the model, with thorough analysis. The results obtained demonstrate the effectiveness of the proposed dual-objective learning strategy, with a segmentation accuracy of 99.15% and 98% mean dice coefficient. The model achieved a score of 94% f1-score in the classification task, considering 4 different possible patient outcomes.

Paper Nr:	558
Title:	Current Trends and Insights in Bug Reporting: A Systematic Literature Review
Authors:	Andreea Galbin Nasui
Abstract:	Bug reporting and software testing are essential elements of software quality assurance. This study provides a structured overview and in-depth analysis of current bug reporting and testing methodologies through a Systematic Literature Review (SLR) based on publications from multiple digital libraries. It focuses on three key aspects: bug reports, testing techniques, and the use of Machine Learning in bug report analysis. The research investigates strategies to enhance the quality, reproducibility, and usefulness of bug reports, emphasizing how detailed reporting and testing improve bug detection, triage, and resolution. The findings reveal a predominance of foundational studies that examine the motivations, challenges, and benefits of bug reporting and testing, with many proposing automated tools and frameworks. This review identifies critical research gaps and highlights future directions for advancing software engineering practices to support better software quality assurance.

Paper Nr:	566
Title:	Enhanced Multi-Class Ultrasound Thyroid Nodule Classification Using CBAM-ResNet50 with Multi-Scale Feature Extraction and DCGAN-Based Data Augmentation
Authors:	Noura Aboudi and Nawres Khlifa
Abstract:	Thyroid nodule classification in ultrasound images remains challenging due to the limited availability of large annotated datasets, class imbalance, and hard-to-detect inter-class differences. This paper proposes a novel framework for multi-class thyroid nodule classification that integrates advanced feature extraction, attention mechanism, and data generation techniques. The proposed method is based on a fine-tuned ResNet50 architecture enhanced with Convolutional Block Attention Modules (CBAM) inserted after each convolutional block to improve feature representation. Multi-scale and directional information is captured using the Shearlet Transform, enriching the extracted features. To address restricted data availability and class imbalance, Deep Convolutional Generative Adversarial Networks (DCGANs) are employed to generate realistic synthetic images, with a dedicated model trained for each TI-RADS category. Experimental results on a benchmark dataset demonstrate that the proposed approach outperforms baseline methods, improving both classification accuracy and generalization performance.

Paper Nr:	573
Title:	Enhancing Emotion Recognition in Manga Dialogues by Leveraging Speech Balloon Shapes
Authors:	Thanh-Lam Bui, Thien-Bao Nguyen, Hoai-Thuong Dang and Xuan-Nam Cao
Abstract:	Emotion recognition in manga dialogue is a complex, unsolved task. While traditional NLP methods focus on text, they ignore crucial visual cues. This paper presents two core contributions. First, we introduce a new benchmark for this task by annotating the standard Manga109 dataset. We provide novel emotion labels for its textual dialogues, generated using a text-only, multi-LLM consensus pipeline. This creates a strong, realistic baseline. Second, we propose a novel multimodal architecture that fuses text features (via BERT) with visual features from the existing speech balloon masks (via ResNet50). We hypothesize that the balloon’s shape contains vital emotional information that text alone lacks. Experiments show our multimodal model significantly outperforms the text-only baseline. This work validates that balloon morphology is a critical feature for semantic analysis in manga.
Download

Paper Nr:	585
Title:	Temporal Visual Explanation for Depression Screening from Facial Expressions
Authors:	Marcelo Simas Mattos, Jose Manuel Molina López and Ana Cristina B. Garcia
Abstract:	This paper presents a video-based framework for temporally dynamic visual explanations of facial behavior in automated depression detection. The system extracts Action Units (AUs), gaze, head pose, and facial landmarks from videos using OpenFace, producing multivariate time series processed by a temporal convo-lutional network with self-attention. To interpret model predictions, a Temporal Dynamic Explanation Model (TDEM) maps feature relevance to anatomically defined facial regions, generating animations that illustrate how key AUs evolve over time. SHAP (SHapley Additive exPlanations) values are used as a post-hoc attribution mechanism to quantify feature contributions. The results demonstrate that the proposed approach enables interpretable insights into expressive dynamics, highlighting facial regions most relevant to depressive patterns and enabling clinicians to assess their temporal evolution.

Paper Nr:	596
Title:	Polarity Related Influence Maximization through Multi-Agent Reinforcement Learning
Authors:	Anikó Kopacz and Camelia Chira
Abstract:	Influence maximization is a network optimization problem, which consists of selecting nodes as sources while maximizing the spread of information. The source nodes that are initially activated form the seed set. Polarity-related influence maximization accounts for having both positive and negative types of relationships between nodes. In this paper, we propose a two-stage multi-agent reinforcement learning based approach to address the polarity-related influence maximization problem. In the first stage, Louvain community detection is applied to find interlinked groups of nodes, and we assign agents to a subset of the communities. The responsibility of the agents is to select a strategy to determine a node from the community as the next seed-node. In the second stage, a deep reinforcement learning model is trained to select the strategies that maximize the information spread in the network. The nodes determined by the agents are aggregated and form the seed set. The proposed approach is validated on the Bitcoin OTC, WikiElec and Slashdot directed signed networks, and the results show that community-based reinforcement learning agents are able to optimize the positive influence spread.

Paper Nr:	608
Title:	Neural Network–Based Speedup of Vehicle-Group Assignment for Dial-a-Ride Problem
Authors:	Adéla Kubíková, Michaela Urbanovská, David Fiedler and Jiří Vokřínek
Abstract:	The Dial-a-Ride Problem (DARP) is a pickup-and-delivery problem with capacity and timing constraints, whose exact solution becomes computationally expensive for large instances. In this work, we propose a neural network–based classifier to accelerate a DARP solution method, the Vehicle-Group Assignment (VGA), by predicting the feasibility of vehicle-group assignments. Two classifier variants are introduced: a per group size model for fixed group sizes and a transformative model for arbitrary sizes. Experiments on a Manhattan instance show that both models significantly decrease the computational time of the VGA, with the per group size model maintaining near-optimal solutions and the transformative model offering faster runtime with minimal cost increase. Additionally, the computational time and solution quality are compared against heuristic methods: Insertion Heuristic (IH) and Adaptive Large Neighborhood Search (ALNS). This comparison demonstrates that learned feasibility prediction can effectively enhance exact DARP solvers and provide better performance-cost tradeoffs than heuristic solutions.

Paper Nr:	609
Title:	ECLARISS: Security-by-Design Smart-Home Surveillance System with Edge Dynamic Biohashing
Authors:	Amira Henaien, Hadda Ben ElHadj and Lamia Chaari Fourati
Abstract:	Deep learning is rapidly evolving and brings major benefits to smart-home surveillance, yet frequent model updates introduce compatibility and security risks when systems depend on fixed feature extractors. This creates a strong need for model-agnostic biometric protection, especially in edge-IoT environments handling sensitive data. ECLARISS meets this need through a Security-by-Design framework that preserves template security across continuous DL evolution using dynamic biohashing and NIST-compliant ECC. The architecture applies Zero Trust with TPM/TEE-rooted trust, AES-GCM encryption, and adaptive credential renewal, while Security Hubs maintain resilient local access control during outages. We formally prove, using ProVerif, that ECLARISS provides unlinkability, irreversibility, replay resistance, and forward secrecy, demonstrating that strong cryptographic design can safely support evolving AI-based IoT systems.

Paper Nr:	613
Title:	Emotion-Conditioned 3D Human Motion from Monocular Video: A Mocap-Free Diffusion Pipeline
Authors:	Ciprian Paduraru, Alexandru Sasu and Alin Stefanescu
Abstract:	Generating realistic 3D human motion conditioned on emotional labels is challenging because behavioral cues are kinematically subtle and difficult to capture at scale. This paper presents a mocap-free, monocular pipeline for emotion-conditioned motion synthesis. A hybrid inverse-kinematics network is used to lift strictly curated 2D video into 3D sequences for a new whole-body dataset, and the resulting kinematics are used to train a diffusion-based motion generator. To compensate for the lack of ground-truth contact annotations in video-derived data, geometric auxiliary losses on foot velocity and floor support are incorporated into the training objective. The framework is evaluated on emotion-conditioned generation with an action-to-motion–style protocol, including comparisons against alternative generative baselines and an ablation of geometric and lifting components. A user study indicates that diffusion-based synthesis better preserves emotional style and achieves high perceived realism, while still falling short of real motion. Finally, an analysis of standard action-based automatic metrics reveals weak correlation with human judgments for emotional motion, highlighting the need for emotion-aware quantitative benchmarks.

Paper Nr:	616
Title:	Artificial Intelligence and Big Data in Social Media Analytics: A Comprehensive Review of Methods, Applications and Trends
Authors:	Hichem Dabbèchi, Nahla Haddar, Haytham Elghazel and Kais Haddar
Abstract:	The rapid proliferation of social media has created vast, dynamic, and heterogeneous datasets that offer unprecedented opportunities for businesses, researchers, and policymakers to extract actionable insights and gain competitive intelligence. Leveraging these data requires advanced artificial intelligence (AI) and big data techniques, including machine learning, deep learning, and graph neural networks, capable of analyzing large-scale, temporally evolving social networks. This paper presents a comprehensive review of recent research (2020–2025) on AI-driven social media analytics, focusing on methodologies, applications, and emerging trends. We systematically categorize the literature into four main areas: (1) graph-based models and graph neural networks for network analysis, (2) temporal and dynamic community detection, (3) natural language processing for textual analytics, and (4) applications in business intelligence. Key challenges, such as data quality, scalability, real-time processing, interpretability, and ethical considerations, are critically discussed. Finally, we highlight future research directions, including hybrid AI frameworks, multi-modal analytics, ex-plainable AI, and the integration of social media insights into business intelligence pipelines. This review provides a structured foundation for researchers and practitioners to harness AI and big data for strategic decision-making, predictive analytics, and the advancement of social network analysis and business intelligence.

Paper Nr:	619
Title:	HybV-Search: An Intelligent Video Retrieval Engine with Multimodal Hybridization and LLM-Driven Adaptive Control
Authors:	Amal Trifa and Salma Mehrez
Abstract:	Locating precise temporal segments in video based on complex multimodal queries is a significant chal-lenge. Existing end-to-end models often apply a rigid fusion strategy, failing to adapt to diverse user intents. We introduce a hybrid, multi-stage search architecture that decouples query understanding from retrieval. Our framework leverages a Large Language Model (LLM) for initial intent detection, classifying queries as dialogue-centric or scene-centric to dynamically guide the search strategy. Video content is indexed using a rich tri-modal representation (visual features, descriptive captions, and dialogue transcripts), enabling a highly efficient coarse-to-fine search. This hierarchical process first identifies candidate videos and then performs a precise, query-adaptive localization on their segments. Our approach demonstrates significant improvements in search granularity and efficiency, accurately locating both visual events and specific spoken utterances.

Paper Nr:	620
Title:	Non-Intrusive Acoustic Monitoring of Bridge Expansion Joints: Multi-Class Defect Detection Using Hybrid Ensemble Machine Learning and Synthetic Audio Augmentation
Authors:	Mohammad Reza Mohebbi, Manu Gupta and Mario Döller
Abstract:	A non-intrusive structural health monitoring system for Bridge Expansion Joints (BEJs) was developed using acoustic sensing and Machine Learning (ML). Vehicle-induced impacts at bridge joints produce characteristic bump sounds, whose spectral and temporal properties change in the presence of structural anomalies such as delamination, cracks, or loose components. This study introduces a hybrid ensemble learning method consisting of Random Forest (RF), Gradient Boosting (GB), Support Vector Machines (SVM), and K-Nearest Neighbors (KNN) for binary and multi-class defect classification. For improving the model’s robustness and handling class imbalance, real audio data captured on an Austrian bridge were increased by overlaying defect waveforms generated from synthetic simulations and real-world impact-echo signals, in which 10 varying defect conditions were simulated. Short-Time Fourier Transform, Mel-Frequency Cepstral Coefficients (MFCCs), spectral and time-domain features, and perceptual measures were extracted, while feature selection was performed using mutual information and minimum Redundancy Maximum Relevance (mRMR). The proposed system achieved 98.9% accuracy for binary defect detection and 82% accuracy with a macro-F1 of 0.88 for ten-class defect recognition. Feature reduction and downsampling to 22.05 kHz made efficient implementation on edge devices possible. In this work, it is shown that acoustic analysis combined with ML provides a practical and scalable early warning system for bridge maintenance, with potential application in infrastructure monitoring and smart city development.

Paper Nr:	639
Title:	Let Us Help Each Other: A Mutualistic Framework for Self-Assisted Quality Control Automation
Authors:	Matteo Tschesche, Stefan Schiffer, Abhirup Das, Alexander Ferrein and Ingo Elsen
Abstract:	This paper introduces the concept of self-assisted automation and proposes a mutualistic framework for quality control automation. Unlike conventional automation, which relies on fixed logic and static setups, self-assisted automation builds on the idea that the automation process itself becomes the subject of automation. The motivation for this paradigm arises from the growing need for flexible, data-driven manufacturing. While modern AI offers adaptability, its deployment in industrial environments is hindered by the absence of mechanisms for efficient retraining and adaptation to data drift. Current automation architectures and robotic inspection systems often depend on manually taught poses or CAD-based matching, making them labor-intensive and inflexible. As a result, many manufacturers remain unable to exploit AI-enabled automation fully. At its core, self-assisted automation leverages human expertise while minimizing manual workload through AI-based assistance. To realize this, we propose a framework that integrates robotic inspection hardware with anomaly detection, Human-in-the-Loop retraining, active learning, coverage path planning, and a novel 3D labeling tool designed to reduce expert effort. The framework was evaluated for applicability in several industrial case studies, demonstrating its potential for adaptive, cost-effective, and sustainable quality control automation.

Paper Nr:	642
Title:	Multilingual Minimal Contrastive Editing
Authors:	Domingo Benoit Cea, Ricardo Ñanculef and Joaquín De Ferrari
Abstract:	We introduce MMiCE, a multilingual domain-agnostic method for generating contrastive explanations via minimal edits to multiclass and multilabel inputs. MMiCE builds on this by fine-tuning large language models with LoRA adapters and guiding edits with attribution and distance constraints, MMiCE produces fluent, faithful edits that flip model predictions. We demonstrate effectiveness across English and Spanish datasets in both social media and clinical domains, achieving a 99% average of flipped labels across datasets. We also propose a new method for counterfactual edit generation in multilabel settings through an inverse gradient attribution scheme and demonstrate it’s fluency performance improvement for the multilabel setting.
Download

Paper Nr:	668
Title:	A Dual-Task U-Net Framework for Efficient Breast Ultrasound Diagnosis
Authors:	Ikram Ben Ahmed and Chokri Ben Amar
Abstract:	As one of the main causes of cancer deaths worldwide, breast cancer has become a serious cause for concern in the modern world. Thus, early diagnosis is essential for successful treatment because it could help avoid complications and complicated pathological treatment. Several CAD systems were established for this reason. However, to produce more accurate results, these systems must continue to evolve by integrating novel techniques. Ultrasound imaging is a potential method for diagnosing and treating many diseases. However, because manual ultrasound image interpretation is time consuming and prone to error, the creation of reliable and effective automated segmentation and classification models is crucial. In this paper, we suggest XAI-CAD, an explainable multitask transfer-learning framework that couples a U-Net segmentation backbone with pretrained encoders to jointly segment lesions and classify pathology on breast ultrasound images, for quick, efficient and highly accurate diagnosis. Using two publicly accessible datasets, BUSI, and Database B, the model was developed, tested, and evaluated against other approaches. Quantitative criteria for evaluating segmentation outcomes, such as Dice coefficient, precision, and recall, are all above 90%, demonstrating that the proposed architecture system can differentiate functional tissues in breast ultrasound images. As a result, our proposed architecture has the potential to offer the classification required to aid in the clinical detection of breast cancer while also improving imaging in other modalities of medical images.

Paper Nr:	680
Title:	A Retrieval-Augmented Generation Based Approach for Contextualized & Complete Explanations
Authors:	Salma Kishk, Nourhan Ehab and Mervat Abu-Elkheir
Abstract:	Explainable AI is now essential for building the trust required to integrate Artificial Intelligence into real-world applications. The explanation graph framework introduces explainability in a contextualized way. While prior work relies on manual modeling of stakeholder context and fixed explanation templates, we propose automating the stakeholder context extraction and introducing a Retrieval-Augmented Generation approach for explanation generation. Applied to a personality detection and entrepreneurship tendency use case, the framework integrates model outputs, expert knowledge, and extracted stakeholder context to produce factual, contextualized, and complete explanations. These improvements reduce manual effort, eliminate the need for rigid templates, and make explanation graphs more adaptable for real-world deployment.

Area 2 - Agents

Full Papers

Paper Nr:	197
Title:	How Can Beliefs Alter Opinions? Joint Opinion and Belief Evolution
Authors:	Hiro Kataoka, Jérôme Euzenat and Koji Hasebe
Abstract:	Multi-agent opinion dynamics and belief propagation have been studied independently. However, people beliefs may influence their opinions and beliefs should not be too dissonant with opinions. This paper considers how to combine opinions and beliefs so that they preserve the cognitive coherence of agents. For that purpose, it integrates social opinion and belief propagation using classical procedures and two new specific operations guided by values held by agents. In order to assess the effect of this model, we experiment with such agents, varying the processing workflow, graph topology, initial beliefs and opinions, and values. Results show that, by maintaining the coherence between beliefs and opinions of individual agents, the social beliefs and opinions resulting from their propagation are indeed affected. In particular, contrary to what happens with the classical opinion dynamics and belief propagation procedures, connecting opinions and beliefs makes them not necessarily converge and not even stabilize. This means that the model supports agents having polarized opinions and beliefs. The outcome of the propagation depends on the workflow, topics, values, and initial graph topology, beliefs and opinions.
Download

Paper Nr:	198
Title:	From Particle Swarms to Opinion Dynamics: A Novel Framework for Modeling Social Influence on YouTube
Authors:	Sabrine Toumi, Raïda Ktari, Hanen Ameur, Hasna Njah and Salma Jamoussi
Abstract:	Modeling opinion evolution in social networks remains challenging due to the complex, adaptive nature of social influence. This paper presents a novel framework that leverages Particle Swarm Optimization (PSO) principles to model opinion dynamics as a multi-agent optimization process. We establish formal correspondences between PSO mechanisms and social influence, developing a mathematical model that integrates individual stubbornness, social influence from neighbors, and informational influence from content features. The model parameters are estimated directly from observable social media behavior, enabling practical application to real-world data. We validate our approach on a dataset of YouTube comments about COVID-19 vaccination, demonstrating competitive performance against established baselines while providing unique and interpretable insights. Our analysis reveals that social influence dominates the informational influence in online discourse, accounting for 63% of opinion updates on average. This work bridges swarm intelligence and social dynamics, offering a powerful new paradigm for understanding and predicting opinion evolution in digital environments.
Download

Paper Nr:	231
Title:	Probabilistic Alternating-Time Temporal Logic with Stochastic Abilities
Authors:	Sarra Zaghbib, Gabriel Ballot, Vadim Malvone and Jean Leneutre
Abstract:	Model checking provides a rigorous means to analyze complex systems by ensuring properties hold across all executions. While originally applied to closed systems, model checking now extends to multi-agent systems, such as distributed protocols, communication systems, robotics, and cybersecurity. A recurring challenge across these domains is reasoning about agents with hidden and uncertain profiles; for example adversarial traders in markets, coordinated users in social media, or attackers in cybersecurity. Addressing this requires logics capable of capturing both probabilistic profiling and reasonig. Alternating-time Temporal Logic (ATL) offers a foundation for reasoning about strategic abilities in multi-agent systems. Extensions such as ATL with Stochastic Abilities (ATL-SA) incorporate stochastic capacities, but existing frameworks remain limited: they can model uncertainty over profiles or allow reasoning about capacities from observed actions, yet not both simultaneously. In this work, we extend ATL-SA with a probabilistic capacity operator, enabling the specification and verification of properties that combine stochastic profiling, inference of hidden information, and strategic reasoning. This framework broadens the scope of formal verification as a general methodology for adversarial and uncertain environments, with applications spanning markets, social platforms, distributed computing, and cybersecurity.
Download

Paper Nr:	241
Title:	Clinical Decision Support System for Pulmonary Embolism Case Management Using Large Language Model Agents
Authors:	Antonio Giménez-López, Andrés Piñeiro-Martín, Iago Mosquera-Fajardo, Laura Docío-Fernández and Carmen García-Mateo
Abstract:	Clinical decision-making can be highly complex due to the cognitive overload of processing extensive patient data while simultaneously considering the full set of recommendations and warnings outlined in official clinical guidelines. To support physicians in this process, we developed and evaluated a multi-agent Clinical Decision Support System (CDSS) that combines Large Language Model (LLM) agents with Retrieval-Augmented Generation (RAG) to autonomously interpret medical guidelines and patient information, to generate clinically relevant outputs. We evaluated the system in a pulmonary embolism (PE) management scenario using two custom-designed datasets: 45 realistic guidelines queries and 20 simulated patient cases. The system achieved 97.78% accuracy in guidelines management and mean experts ratings above 95% for clinical accuracy, risk score computation, and interpretability. These results suggest that agentic CDSSs have the potential to reason effectively over clinical data and guidelines, producing structured outputs that are generally consistent with clinical standards and expert expectations. This work provides a proof of concept for integrating agentic LLM architectures into CDSS workflows, highlighting their potential to augment clinical decision-making in PE and underscoring the need to assess their effectiveness in real-world clinical settings and their applicability to other disease domains.
Download

Paper Nr:	280
Title:	Decentralized Workload Scheduling in Edge-Fog-Cloud Continuum Using Hormone-Based Swarm Intelligence Algorithm
Authors:	Kefan Wu, Nadezhda Varzonova and Melanie Schranz
Abstract:	The growing complexity of workload allocation in the edge-fog-cloud continuum requires efficient scheduling mechanisms to decrease the dependency on central servers and communication costs. Agent-based principles of swarm intelligence offer a possibility for decentralized decision-making and optimization in such a distributed system from the bottom-up. This paper explores an innovative workload scheduler inspired by the hormone-based algorithm (HBA). We define a novel behavior of the resource agents by leveraging virtual hormones in the framework. The proposed HBA workload scheduler can direct requests (pods) to the best resource agent for execution. Moreover, we present a new simulation framework for the evaluation of this agent-based modeling approach, specifically designed for the edge-fog-cloud continuum. We explore the performance of the HBA under various configurations. The simulation case studies are designed to compare the HBA with an ant colony optimization (ACO) implementation. Our results show that the scheduling under HBA can reduce the communication costs in the entire continuum and additionally reduce the dependency on fog and cloud nodes.
Download

Paper Nr:	311
Title:	Learning Long-Horizon Multi-Agent Coordination from Temporal Logic Specifications
Authors:	Albin Larsson Forsberg, Alexandros Nikou, Aneta Vulgarakis and Jana Tumova
Abstract:	We study multi-agent reinforcement learning (MARL) under temporally extended Signal Temporal Logic (STL) objectives, which require reasoning over both long-horizon dynamics and inter-agent relations. We propose TD-MAT, a transformer-based architecture with multivariate positional encodings, causal temporal masking, and a decomposed reward based on arithmetic–geometric mean robustness with variance regularization. Experiments on coordination tasks ranging from unstructured multi-objective problems to strict temporal sequencing show that TD-MAT learns effective long-term behaviors and generalizes to heterogeneous agent settings. Ablation studies highlight the necessity of temporal masking, positional encodings, and reward decomposition, while comparisons to MAPPO, RMAPPO, and MAT reveal that transformers provide the greatest benefit on unstructured, long-horizon tasks.
Download

Paper Nr:	344
Title:	Towards Reliable Reasoning with Multi-Agent Cognitive Collaboration in Small Language Models
Authors:	Cristina Alameda, Gonzalo Méndez and Raquel Hervás
Abstract:	Small Language Models offer practical advantages but face critical limitations, including poor self-calibration and inconsistent reasoning. We present MACoMoE (Multi-Agent Cognitive Mixture-of-Experts), a modular framework that addresses these challenges via structured cognitive specialization and collective validation, without increasing model size. The architecture separates reasoning generation from quality assessment across three layers: specialized cognitive reasoners (deductive, inductive, abductive), multi-perspective evaluators (logical validity, epistemic soundness, empirical groundedness), and a decision aggregation layer comparing four consensus mechanisms. Evaluation on 256 reasoning problems on two mathematical datasets shows that voting-based aggregation substantially outperforms self-assessment: Simple Majority achieves 95.7% alignment with ground truth versus 51.4% for self-confidence scoring (+44.3pp). Chain-of-Thought prompting improves accuracy by 29.3pp on MR-GSM8K and 23.4pp on MR-BEN while reducing reasoning steps by 36%, indicating efficiency gains through structured exploration. These results demonstrate that compact models can effectively serve as peer evaluators, and that simple consensus protocols suffice when reasoning diversity is ensured, offering a computationally efficient alternative to individual SLM deployment and costly frontier model ensembles.
Download

Paper Nr:	367
Title:	Distributed Multi-UAV Partition-Based Patrolling with Fault Tolerance: A Study on Meeting-Based Coordination Strategies
Authors:	Puvikaran Santhirasegaram, Henrik Van Peet, Mads Beyer Mogensen, Giovanni Bacci, Timothy Robert Merritt and Michele Albano
Abstract:	Patrolling tasks in multi-robot systems are essential for applications such as surveillance and monitoring, where minimizing the time between visits to any given location is critical. This paper investigates fault-tolerant redistribution strategies for multi-agent patrolling systems in partitioned environments using Unmanned Autonomous Vehicles (UAVs). We propose Heuristic Meeting-based Patrolling (HMP), a novel distributed and fault-tolerant patrolling algorithm. Building on the Heuristic Conscientious Reactive (HCR) strategy and incorporating periodic synchronization meetings, HMP enables decentralized coordination and dynamic fault recovery through minimal communication. UAVs exchange information at shared meeting points, enabling detection of failures and redistribution of responsibilities. We evaluate HMP and its simplified variants in various simulated environments using the Multi-Agent Exploration and Patrolling Simulator (MAEPS). The results demonstrate that HMP offers strong performance under both normal and fault conditions, comparative to state-of-the-art patrolling strategies in terms of idleness metrics. However, we also identify limitations in meeting scheduling under certain fault conditions, which can cause cascading failures. Based on these findings, we discuss potential improvements for future work, including enhanced meeting scheduling and adaptive partitioning strategies.
Download

Paper Nr:	453
Title:	Online Learning of Object-Centric Symbolic Models in Partially Observable Environments
Authors:	Leonardo Lamanna, Luciano Serafini, Alessandro Saffiotti and Paolo Traverso
Abstract:	Object-centric world models are increasingly used in agents that reason and act within an environment. These models specify which objects exist, their properties, how they are changed by actions, and how they are perceived through the agent’s sensors. In this paper, we address the problem of how an agent can learn such models autonomously and online by executing actions and observing their effects. The agent models the environment as an object-centric, partially observable, relational MDP. We define an algorithm by which the agent incrementally learns the elements of such a model, namely: the signature for representing the objects and their properties, the observation function that links object states to observations, and a lifted specification of the transition function. We evaluate our approach by demonstrating how the agent can use the learned model to plan and solve a set of tasks not known a priori and compare these results with a set of Reinforcement Learning (RL) baselines. We show that our method learns an environment model that is effective for planning, while requiring significantly less training and outperforms the RL baselines.
Download

Paper Nr:	475
Title:	Perfecting AI Agent Frameworks through Unified Design Principles
Authors:	Sergio Lopez, Guillermo Ramos and Asier Arranz
Abstract:	Large language model (LLM) agents now power applications ranging from code generation to enterprise knowledge orchestration, yet available frameworks remain polarized. Research prototypes deliver transparent reasoning loops and pedagogical clarity, whereas production toolkits emphasize operational resilience, safety, and deployment tooling. This paper synthesizes these perspectives into a unified architecture that balances theory and practice. We articulate design principles centered on modularity, explainability, safety-by-default, and observability-first execution. Building on these principles, we propose a layered stack covering deliberation policies, orchestration engines, model abstraction, tool and service ecosystems, execution environments, and monitoring/governance planes. We illustrate how interoperable interfaces support cross-language implementations, how context propagation enables rich tool use and grounding, and how guardrails, evaluation pipelines, and deployment workflows can be embedded from the outset. The result is a pedagogical yet production-ready blueprint that educators, students, and developers can leverage to construct resilient agentic systems. We conclude with open research directions spanning evaluation benchmarks, self-improving policies, and multimodal orchestration, positioning the unified framework as a foundation for responsible innovation in the agent era.
Download

Paper Nr:	541
Title:	Block Stacking Problem by Group of Agents Using Particle Swarm Optimization
Authors:	Kotaro Yamada, Stephen Raharja and Toshiharu Sugawara
Abstract:	We propose a particle swarm optimization (PSO)-based method in a multi-agent system by identifying a particle as a drone agent to efficiently solve the block stacking problem (BSP), which involves stacking blocks to form walls or structures of a specified shape under certain constraints in an environment. A BSP can be considered as an extension of the multi-agent pattern formation problem in the sense that agents must reach specified locations; however, agents in the BSP are required to iteratively carry blocks to construct the specified structure at given locations. In addition, we assume that the agents are physical entities of a certain size and must individually consider the constraint on the stacking order to avoid collisions with stacked blocks. Therefore, conventional PSO-based methods for pattern formation problems cannot effectively solve the BSP. Using our proposed method, drone agents can iteratively carry blocks to specified locations for efficient BSP solutions. For this purpose, we introduce pheromone control mechanisms so that the agent can avoid unnecessary concentration or collisions while satisfying the BSP constraints. Experimental evaluation demonstrated that our method can solve a BSP more efficiently than the simple extension of methods for the pattern formation problem.
Download

Paper Nr:	599
Title:	SafeCityLearn: A Benchmark for Safety-Constrained Reinforcement Learning in Distributed Energy Systems
Authors:	Amal Nammouchi, Andreas Kassler, Arunselvan Ramaswamy and Andreas Theorcharis
Abstract:	Safe reinforcement learning (Safe RL) aims to optimize performance while respecting safety constraints, yet existing benchmarks largely focus on robotic or navigation tasks. We introduce SafeCityLearn, the first benchmark for Safe RL in distributed energy systems. Extending the CityLearn environment, SafeCityLearn introduces configurable constraint functions and cost signals reflecting realistic energy limits, such as state-of-charge, comfort, and grid import bounds. The benchmark integrates with OmniSafe, enabling standardized training and evaluation across major Safe RL algorithms aiming to avoid constraint violations, including PPO-Lagrangian, RCPO, and SAC-Lag. Experiments on a full-year battery scheduling task show that Lagrangian-based algorithms significantly reduce SoC violation rates often below 5% while achieving competitive energy cost performance, whereas unconstrained baselines showcase persistent infeasibility. SafeCityLearn provides a reproducible framework for advancing constraint-aware reinforcement learning in sustainable energy systems. We make our implementation available to the community as open source for reproducibility. and community The full benchmark, implementation, and experiment scripts are publicly available at: https://github.com/AmalNamm/Safe-CityLearn.
Download

Paper Nr:	605
Title:	Distributed Performability Optimization for Multi-UAV Road Traffic Monitoring
Authors:	Qingyang Zhang, Koji Noshiro, Mohammad Dwipa Furqan, Koji Hasebe and Fumio Machida
Abstract:	UAV-based monitoring is a promising solution to road traffic management and operation. UAVs process image data for traffic monitoring, which demands significant local computing resources. To reduce workload and conserve battery, it is effective to exploit computation offloading. However, the offloading decision is not trivial when multiple UAVs collaborate and share limited computing resources. It is a crucial challenge to maximize the total system performance and availability in such a multi-UAV-based distributed monitoring system. To address this challenge, we formulate the multiple-UAV offload decision problems as a Distributed Constraint Optimization Problem (DCOP) and solve it using effective heuristic algorithms. We consider performability as the optimization target, which is a composite measure of total system performance and availability. To derive quantitative performability estimates for given computation modes (local processing or offloading), we model the system using stochastic reward nets. The estimated performability values of UAVs are supplied to the DCOP individually, and the distributed algorithms determine their computation modes to maximize the total system performability. We conducted simulation-based experiments in which realistic computational demands were generated using a virtual road traffic simulator. The evaluation results show that the Max-Sum algorithm achieves the best performability across all evaluation scenarios, demonstrating the effectiveness of the DCOP-based solution approach.
Download

Paper Nr:	618
Title:	From Utility-Based Action Selection to the Emergence of Communication Chains in a Multi-Robot System
Authors:	Erwan Martin, Philippe Mathieu and Antoine Nongaillard
Abstract:	Multi-Robot systems (MRS) offer an effective alternative to human operation in hazardous or large-scale environments. However, ensuring efficient coordination when communication is constrained remains a significant challenge. This paper introduces an emergent coordination mechanism that enables each robot to make autonomous decisions based on locally acquired knowledge through perception and proximity communications. Every robot relies on a utility function that balances individual and collective efficiency. Unlike existing approaches, our method does not rely on global consensus or environment modification and demonstrates robustness to communication disturbances. We show that imposing specific constraints leads to the emergence of a communication chain towards a base station, and analyze the influence of various parameters on the dynamics of this phenomenon.
Download

Short Papers

Paper Nr:	31
Title:	Emotional Modulation in Swarm Decision Dynamics
Authors:	David Freire-Obregón
Abstract:	Collective decision-making in biological and human groups often emerges from simple interaction rules that amplify minor differences into consensus. The bee equation, developed initially to describe nest-site selection in honeybee swarms, captures this dynamic through recruitment and inhibition processes. Here, we extend the bee equation into an agent-based model in which emotional valence (positive–negative) and arousal (low–high) act as modulators of interaction rates, effectively altering the recruitment and cross-inhibition parameters. Agents display simulated facial expressions mapped from their valence–arousal states, allowing the study of emotional contagion in consensus formation. Three scenarios are explored: (1) the joint effect of valence and arousal on consensus outcomes and speed, (2) the role of arousal in breaking ties when valence is matched, and (3) the “snowball” effect in which consensus accelerates after surpassing intermediate support thresholds. Results show that emotional modulation can bias decision outcomes and alter convergence times by shifting effective recruitment and inhibition rates. At the same time, intrinsic non-linear amplification can produce decisive wins even in fully symmetric emotional conditions. These findings link classical swarm decision theory with affective and social modelling, highlighting how both emotional asymmetries and structural tipping points shape collective outcomes. The proposed framework offers a flexible tool for studying the emotional dimensions of collective choice in both natural and artificial systems.
Download

Paper Nr:	42
Title:	Robust Centralized Coordination in Multi-Agent Systems: Addressing Agent Diversity and Deception
Authors:	Pia Schweizer, Jonas Lange, Luna Kaendler, Sven Tomforde and Christian Krupitzer
Abstract:	The effectiveness of centralized control in multi-agent systems can be increasingly challenged by behavioral uncertainty as agents gain greater autonomy. Such agents may behave unreliably or even maliciously, compromising the central planner’s ability to maintain system-wide performance. This paper investigates how a central planner (i.e., Central Control Unit) can sustain robust coordination in the presence of uncooperative or even malicious agents. Using platooning as a representative use case, we introduce four distinct agent behavior profiles, two genuine ones and two malicious ones. In a simulation, we compare three coordination strategies-optimistic, fair, and trust-based-under varying distributions of agent types. We analyze how the agent types impact the global performance, fairness, and system stability. Our findings reveal that while optimistic strategies yield strong performance under ideal conditions, fairness and trust mechanisms significantly enhance resilience in adversarial settings. These strategies support equitable treatment of cooperative agents and mitigate exploitation by malicious ones.
Download

Paper Nr:	45
Title:	Evaluation of Diversity in LLM-Based News Discovery through an Agent-Based System
Authors:	Sergio Muñoz and Carlos Á. Iglesias
Abstract:	The proliferation of digital news platforms and sophisticated recommendation algorithms has fundamentally transformed information consumption patterns. These systems often prioritize engagement through preference reinforcement, creating echo chambers that limit exposure to diverse viewpoints and contribute to social polarization. While efforts to incorporate diversity into news recommendation systems have shown promise, existing solutions face several limitations, including data sparsity, limited semantic understanding, and difficulty in balancing relevance and diversity. These shortcomings highlight the need for innovative approaches that integrate relevance, diversity, and fairness in content selection. In this work, we explore the use of Large Language Models (LLMs) to address these challenges in digital news discovery. Specifically, we evaluate the capabilities of LLMs for retrieving and selecting news content from the web, with a focus on identifying relevant articles that represent a range of perspectives. Our study proposes an agent-based system that systematically compares different LLMs through a set of quantitative and normative metrics, evaluating their effectiveness in content retrieval and perspective diversity. The results offer practical insights into the strengths and limitations of current LLM-based approaches for digital news discovery.
Download

Paper Nr:	65
Title:	DualRAG Architecture: Structured Retrieval with Layered Knowledge Models
Authors:	Michal Podpora, Marek Baranowski, Aleksandra Kawala-Sterniuk and Mariusz Pelc
Abstract:	DualRAG is a novel architecture for Retrieval-Augmented Generation (RAG) that introduces a two-layer knowledge representation to improve relevance assessment and context construction. Each knowledge chunk is divided into two parts, separated by a delimiter ("+="): a concise, relevance-focused descriptor and a detailed, content-rich segment. During retrieval, only the relevance descriptor is used for scoring and ranking candidate chunks. After selection, both parts are combined to construct a prompt with greater contextual depth and precision. This decoupling of retrieval scoring from generative input enables more accurate filtering without compromising the expressiveness of retrieved content. The approach is particularly suited to domains requiring fine-grained distinction and layered factual detail, such as environmental regulations or technical documentation. DualRAG improves retrieval quality while maintaining compatibility with existing RAG pipelines and model infrastructures.
Download

Paper Nr:	86
Title:	Sparse Rewards as Preferences: Investigating Reward Shaping with Preference-Based RL Methods in Sparse Reward Domains
Authors:	Veerendrababu Vakkapatla, Shariq Faraz and Virendra Singh
Abstract:	Reinforcement learning (RL) agents struggle with sparse rewards. While reward shaping offers guidance, traditional preference-based RL (PbRL) learns rewards from human input when explicit reward functions are absent. In this work, we adapt PbRL for sparse-reward tasks by automatically generating preferences from the existing sparse reward signal, removing the need for human intervention. We evaluate recent PbRL methods on continuous control tasks (OpenAI Gym MuJoCo), focusing on how their distinct exploration or reward inference strategies contribute to agent learning under this self-supervised framework. Our findings reveal that these specific characteristics significantly enhance the dense reward functions inferred from self-supervised preferences. This leads to more sample-efficient policies that outperform those trained on original sparse rewards, highlighting the potential of leveraging specific PbRL features to automate reward design in sparse-reward problems.
Download

Paper Nr:	108
Title:	AI Agents with Decentralized Identifiers and Verifiable Credentials
Authors:	Sandro Rodriguez Garzon, Awid Vaziry, Enis Mert Kuzu, Dennis Enrique Gehrmann, Buse Varkan, Alexander Gaballa and Axel Küpper
Abstract:	A fundamental limitation of current LLM-based AI agents is their inability to build differentiated trust among each other at the onset of an agent-to-agent dialogue. However, autonomous and interoperable trust establishment becomes essential once agents start to operate beyond isolated environments and engage in dialogues across individual or organizational boundaries. A promising way to fill this gap in Agentic AI is to equip agents with long-lived digital identities and introduce tamper-proof and flexible identity-bound attestations of agents, provisioned by commonly trusted third parties and designed for cross-domain verifiability. This article presents a conceptual framework and a prototypical multi-agent system, where each agent is endowed with a self-sovereign digital identity. It combines a unique and ledger-anchored W3C Decentralized Identifier (DID) of an agent with a set of third-party issued W3C Verifiable Credentials (VCs). This enables agents at the start of a dialog to prove ownership of their self-controlled DIDs for authentication purposes and to establish various cross-domain trust relationships through the spontaneous exchange of their self-hosted DID-bound VCs. A comprehensive evaluation of the prototypical implementation demonstrates technical feasibility but also reveals limitations once an agent’s LLM is in sole charge to control the respective security procedures.
Download

Paper Nr:	130
Title:	The Effect of Belief Boxes and Open-Mindedness on Persuasion
Authors:	Onur Bilgin, Abdullah As Sami, Sriram Sai Vujjini and John Licato
Abstract:	As multi-agent systems are increasingly utilized for reasoning and decision-making applications, there is a greater need for LLM-based agents to have something resembling propositional beliefs. One simple method for doing so is to include statements describing beliefs maintained in the prompt space (in what we’ll call their “belief boxes”). But when agents have such statements in belief boxes, how does it actually affect their behaviors and dispositions towards those beliefs? And does it significantly affect agents’ ability to be persuasive in multi-agent scenarios? Likewise, if the agents are given instructions to be open-minded, how does that affect their behaviors? We explore these and related questions in a series of experiments. Our findings confirm that instructing agents to be open-minded affects how amenable they are to belief change. We show that incorporating belief statements and their strengths influences an agent’s resistance to (and persuasiveness against) opposing viewpoints. Furthermore, it affects the likelihood of belief change, particularly when the agent is outnumbered in a debate by opposing viewpoints, i.e., peer pressure scenarios. The results demonstrate the feasibility and validity of the belief box technique in reasoning and decision-making tasks.
Download

Paper Nr:	157
Title:	Social Norm Reasoning in Multimodal Language Models: An Evaluation
Authors:	Oishik Chowdhury, Anushka Debnath and Bastin Tony Roy Savarimuthu
Abstract:	In Multi-Agent Systems (MAS), agents are designed with social capabilities, allowing them to understand and reason about social concepts such as norms when interacting with others (e.g., inter-robot interactions). In Normative MAS (NorMAS), researchers study how norms develop, and how violations are detected and sanctioned. However, existing research in NorMAS use symbolic approaches (e.g., formal logic) for norm representation and reasoning whose application is limited to simplified environments. In contrast, Multimodal Large Language Models (MLLMs) present promising possibilities to develop software used by robots to identify and reason about norms in a wide variety of complex social situations embodied in text and images. However, prior work on norm reasoning have been limited to text-based scenarios. This paper investigates the norm reasoning competence of five MLLMs by evaluating their ability to answer norm-related questions based on thirty text-based and thirty image-based stories, and comparing their responses against humans. Our results show that MLLMs demonstrate superior performance in norm reasoning in text than in images. GPT-4o performs the best in both modalities offering the most promise for integration with MAS, followed by the free model Qwen-2.5VL. Additionally, all models find reasoning about complex norms challenging.
Download

Paper Nr:	183
Title:	Interface Assistant Agents: A General Segment-Everything Approach
Authors:	Gustavo Guidi Venâncio Martins, Antonio Carlos Sobieranski, Fabricio Ourique and Alison R. Panisson
Abstract:	We present a general-purpose approach for developing interface assistant intelligent agents based on a novel General Segment-Everything Approach (GSEA). Our method treats user interface understanding as a segmentation problem, leveraging computer vision techniques to isolate relevant components within diverse UI environments. By integrating GSEA with Large Multimodal Models (LMMs), the system extracts semantic information from segmented interfaces, enabling high-level interpretation of interface structures and elements. These capabilities are incorporated into a decision-making mechanism that allows the agent to interact autonomously with user interfaces to complete a wide range of tasks. We demonstrate the effectiveness of our approach towards a case study, highlighting its task performance, and potential for scalable interface automation. This work contributes toward the development of more adaptable and perceptually grounded assistant agents for real-world digital environments.
Download

Paper Nr:	184
Title:	Towards a BDI Architecture for Cooperative Agents on Resource-Constrained Microcontrollers
Authors:	Maurício Darabas Ronzani, Italo Firmino da Silva, Jim Lau, Roberto Rodrigues-Filho, Fabricio Ourique and Alison R. Panisson
Abstract:	This paper presents the design and implementation of cooperative multi-agent systems based on the BDI (Belief-Desire-Intention) architecture for resource-constrained microcontrollers. The proposed solution is validated on real hardware, demonstrating a complete and decentralized cooperation cycle over a low-power ESP-NOW network. At the core of our architecture is a belief synchronization loop, a mechanism that integrates network-state awareness into the BDI reasoning cycle. This enables a sensor agent to deliberate on its environment and successfully request actions from a remote actuator agent by forming dynamic cooperation plans. This work aims to contribute to an identified gap in the literature concerning real world applications of cooperative agents on low-cost hardware, as research has predominantly focused on simulation, single-agent optimization, or higher-power computing platforms.
Download

Paper Nr:	239
Title:	Analyzing Japan-Specific Macroprudential Policies for Variable-Rate Mortgages: An Agent-Based Approach
Authors:	Takahiro Obata and Takao Miyamoto
Abstract:	Following the Bank of Japan’s termination of its zero-interest-rate policy in March 2024, Japan has entered a period of rising interest rates after decades of near-zero levels. Among the markets most affected by interest rate increases is the housing market, where approximately 70% of outstanding housing loans are variable-rate mortgages (VRMs). This composition implies that higher rates could significantly impact household financial stability. To mitigate such adverse effects, Japan adopts macroprudential regulations specific to its housing market, namely the so-called Five-Year Rule and 125% Rule. However, the long-standing low-interest environment has made it difficult to empirically assess their effectiveness under conditions of rising rates. To address this issue, this study develops an agent-based model (ABM) of the Japanese housing market that explicitly incorporates variable-rate mortgages and Japan’s specific macroprudential rules. The model is calibrated using housing market data from Tokyo, where residential transactions are most concentrated. Based on this calibrated model, we simulate an interest-rate-hike scenario to evaluate the effects of the Five-Year Rule and 125% Rule. Simulation results indicate that the effectiveness of these policy measures depends on the specific threshold settings. In particular, tighter parameterization of these rules reduces the number of mortgage defaults among households, highlighting their potential as stabilizing mechanisms during interest rate upswings.
Download

Paper Nr:	242
Title:	Transformer-Based Action Prediction for Efficient Mutual Modeling in Multi-Agent Reinforcement Learning
Authors:	Kohei Suzuki and Toshiharu Sugawara
Abstract:	Although multi-agent deep reinforcement learning (MADRL) has made progress, a key challenge remains: non-stationarity of the environment caused by agent interactions, which undermines learning stability. Another issue is the difficulty in designing reward functions that promote agent cooperation in multi-agent systems. To address these challenges, approaches that model other agents’ behavior, called modeling other agents (MOA), have been explored. However, conventional methods rely on assumptions, such as modeling other agents as entities with fixed policies, which limits their applicability. Our study is based on continuous mutual modeling (CMM), in which agents continuously model others’ policies as they learn. To achieve an efficient CMM framework, we propose a QMIX with Transformer-based predictor, which uses the Transformer to predict other agents’ actions. Our experiments showed that our method enables agents to understand others’ actions more accurately, promoting better coordinated/cooperative behavior in complex multi-agent environments.
Download

Paper Nr:	243
Title:	A Natural Language Agentic Approach to Study Affective Polarization
Authors:	Stephanie Anneris Malvicini, Ewelina Gajewska, Arda Derbent, Katarzyna Budzynska, Jarosław A. Chudziak and Maria Vanina Martinez
Abstract:	Affective polarization has been central to political and social studies, with growing focus on social media, where partisan divisions are often exacerbated. Real-world studies tend to have limited scope, while simulated studies suffer from insufficient high-quality training data, as manually labeling posts is labor-intensive and prone to subjective biases. The lack of adequate tools to formalize different definitions of affective polarization across studies complicates result comparison and hinders interoperable frameworks. We present a multi-agent model providing a comprehensive approach to studying affective polarization in social media. To operationalize our framework, we develop a platform leveraging large language models (LLMs) to construct virtual communities where agents engage in discussions. We showcase the potential of our platform by (1) analyzing questions related to affective polarization, as explored in social science literature, providing a fresh perspective on this phenomenon, and (2) introducing scenarios that allow observation and measurement of polarization at different levels of granularity and abstraction. Experiments show that our platform is a flexible tool for computational studies of complex social dynamics such as affective polarization. It leverages advanced agent models to simulate rich, context-sensitive interactions and systematically explore research questions traditionally addressed through human-subject studies.
Download

Paper Nr:	245
Title:	A Simulation Model for Evaluating Walking Promotion Policies to Support Health-Promoting Cities
Authors:	Setsuya Kurahashi, Taisei Mukai, Fumiko Kumada and Hideyuki Nagai
Abstract:	In the context of a declining birthrate, an aging society, and a shrinking population, there is a growing demand for urban environments that promote residents’ health, well-being, and sense of security. This study develops and evaluates a simulation model to assess the effectiveness of walking promotion policies aimed at encouraging daily physical activity among residents. Enabling individuals to maintain healthy lifestyles within their communities as they age yields significant benefits for both individuals and society, including enhanced life satisfaction, improved quality of life, and reduced healthcare expenditures. To address these challenges, this paper introduces the concept of the Smart Wellness City, a framework designed to foster population-level health promotion through digital and spatial interventions. The study presents the development of a simulator that estimates residents’ step counts based on synthetic population and mobile spatial data. The simulator is intended to evaluate the impacts of various policies designed to increase daily physical activity across the local population.
Download

Paper Nr:	250
Title:	Visual Language State Machine Robot
Authors:	Elias Goldsztejn, Dan Rouven Suissa and Ronen Brafman
Abstract:	Autonomous mobile robots in human environments must balance efficiency with social compliance and safety. While Vision-Language Models (VLMs) enable semantic scene understanding, existing frameworks often lack dynamic reconfiguration in response to changing scenarios, gestures, and hazards. We present zero-shot VLM integration without fine-tuning using a modular state-machine architecture that dynamically adapts navigation, gesture-triggered transitions, and safety-critical overrides. Experiments show robust gesture recognition, object detection, and scenario adaptability, with seamless behavior switching at expected transition points.
Download

Paper Nr:	273
Title:	Generalized Welfare-Aware Matching with Multimodal Preferences and Criteria Strategizing of Autonomous Agents
Authors:	Peash Ranjan Saha, Salimur Choudhury and Kai Salomaa
Abstract:	A multimodal preference model deals with more than one criterion, where an agent maintains an independent preference list for each of the criteria outlined in a problem. A one-sided matching is a mechanism to pair a set of agents with a set of objects or goods where the agents have preferences over the objects. We introduce a multimodal preference model for one-sided matching, where an agent assigns strategic weight to each criterion. We design a generalized score-based approach to determine a generalized cost for each possible pairing of the matching. A generalized welfare-aware matching is introduced, where the objective is pairing the maximum possible number of agents with the minimum possible generalized cost. We provide the integer programming formulation of generalized welfare-aware matching, and an algorithm to determine that in polynomial time. Any matching is popular for a criterion if it receives no fewer votes than any other matching in a pairwise election. It is certainly popular when it is popular for all criteria in a multimodal preference model, which has been introduced recently in the literature. We established theoretical foundations to represent the relations between a generalized welfare-aware matching and a certainly popular matching. In the cases when a certainly popular matching does not exist, we empirically evaluated the matching with the popular matching for each criterion.
Download

Paper Nr:	279
Title:	Learning from Failure: Towards Balance-Aware Robotaxi Fleet Deployment via Multiagent Reinforcement Learning
Authors:	Jiyao Li and Vicki H. Allan
Abstract:	Robotaxi services represent a promising direction in the rapidly growing field of intelligent transportation systems. A critical challenge lies in maintaining a balance between supply (available AVs) and demand (ride requests) across diverse urban regions. While most previous research achieves strong performance in demand-intensive areas such as downtown, it often struggles in demand-sparse or remote regions like suburbs and airports. To address this limitation, we propose Unserved-Request Experience Generation (UEx-Gen), which generates additional training experiences from unserved passenger requests to improve model knowledge of demand-sparse regions. Furthermore, we introduce the Priority Bootstrapping Actor–Critic (PB-AC) algorithm, which prioritizes experiences from under-served areas through bootstrapped sampling, enabling AVs to learn more effectively about supply–demand imbalances. Experiments on a city-scale dataset show that our approach achieves nearly 100% service rate with about 1 minute waiting time in downtown, and up to 30% higher service rate with around 8 minutes less waiting time at the airport compared to state-of-the-art baselines.
Download

Paper Nr:	281
Title:	On Theoretically-Driven LLM Agents for Multi-Dimensional Discourse Analysis
Authors:	Maciej Uberna, Michał Wawer, Jarosław A. Chudziak and Marcin Koszowy
Abstract:	Identifying the strategic uses of reformulation in discourse remains a key challenge for computational argumentation. While LLMs can detect surface-level similarity, they often fail to capture the pragmatic functions of rephrasing, such as its role within rhetorical discourse. This paper presents a comparative multi-agent framework designed to quantify the benefits of incorporating explicit theoretical knowledge for this task. We utilise an dataset of annotated political debates to establish a new standard encompassing four distinct rephrase functions: Deintensification, Intensification, Specification, Generalisation, and Other, which covers all remaining types (D-I-S-G-O). We then evaluate two parallel LLM-based agent systems: one enhanced by argumentation theory via Retrieval-Augmented Generation (RAG), and an identical zero-shot baseline. The results reveal a clear performance gap: the RAG-enhanced agents substantially outperform the baseline across the board, with particularly strong advantages in detecting Intensification and Generalisation context, yielding an overall Macro F1-score improvement of nearly 30%. Our findings provide evidence that theoretical grounding is not only beneficial but essential for advancing beyond mere paraphrase detection towards function-aware analysis of argumentative discourse. This comparative multi-agent architecture represents a step towards scalable, theoretically informed computational tools capable of identifying rhetorical strategies in contemporary discourse.
Download

Paper Nr:	380
Title:	Towards Human–AI Co-Creation in Urban Design Using Participatory Multi-Agent Simulation
Authors:	Soma Endo, Mamoru Yoshizoe, Hiromitsu Hattori, Luis Alberto Alonso Pastor and Kent Larson
Abstract:	Against the backdrop of increasing complexity in social issues, including lifestyle changes resulting from the spread of infectious diseases and heightened environmental awareness, as well as the rapid advancement of AI, there is a growing demand for the redefinition and redesign of societies and cities. Participatory Multi-Agent Simulation (PMAS) is a promising approach for engaging diverse stakeholders in the processes of social and urban design. However, non-expert users who lack specialized knowledge in urban design or information technology often struggle to understand and effectively utilize such simulations. We propose an information support agent powered by large language models to assist PMAS users. The agent provides support by identifying key points and potential problems related to the matters under consideration and offering advice on managing and intervening in the simulation environment. By enhancing users’ understanding of complex issues and supporting their decision-making, the agent enables a PMAS environment in which a wide range of stakeholders can collaboratively address societal and urban issues.
Download

Paper Nr:	390
Title:	SAT-Based Large Neighborhood Search for Multi-Agent Pathfinding
Authors:	Max Frommknecht and Pavel Surynek
Abstract:	We propose a novel hybrid algorithm, LNS-SAT, that uses a Boolean Satisfiability (SAT) repair engine within a Large Neighborhood Search (LNS) to solve Multi-Agent Path Finding (MAPF) makespan-optimally. An initial solution is obtained from the agents’ shortest paths to their goal, where they wait for all agents to finish. Compact spatial-temporal zones are created around conflicts and solved using a lazy Multi-Decision Diagrams (MDDs) SAT encoding. If a zone proves unsatisfiable (UNSAT) time of agents who wait at their goal is reallocated to the zone. If no more waiting time (slack) is available, the zone is expanded and tried again until a local solution is found. In the limit, the zone encompasses the full MAPF instance. Increasing the makespan only when the global instance is UNSAT ensures that the first successful makespan is optimal. Experiments on MovingAI benchmark maps with 10–200 agents show substantially higher success rates and lower runtimes compared to a global MDD-SAT baseline. LNS-SAT scales the performance of MDD-SAT encodings to larger maps with more agents while preserving optimality.
Download

Paper Nr:	400
Title:	Bottom-Up Swarm Intelligence for Secure Cross-Cluster Orchestration in Edge Computing
Authors:	Marija Gojković, Mika Auer, Horia-Rares Vulcu, Fred Buining and Melanie Schranz
Abstract:	Edge computing enables delay-sensitive microservices to run near end applications, extending cloud capabilities to the edge. In multi-cluster environments, fluctuating workloads and resource availability, together with privacy constraints, complicate cross-cluster coordination. We propose pod orchestration using swarm intelligence, where decentralized agents achieve adaptive scheduling across distributed clusters, supporting rigid pods with strict deadlines and elastic pods exploiting residual resources. An Artificial Bee Colony (ABC)inspired mechanism performs emergent global optimization through local, bottom-up decisions with minimal shared state. Because swarm intelligence relies on local interactions, it naturally preserves privacy by limiting global data access. To further protect sensitive information, we integrate anonymization to mask quasi-identifiers and homomorphic encryption for secure computations. We also study how these privacy mechanisms affect coordination efficiency and system overhead. Evaluation shows that our orchestrator achieves scalable, adaptive, and privacy-secure scheduling while maintaining acceptable performance under dynamic, cross-cluster edge conditions.
Download

Paper Nr:	402
Title:	Agentic AI in Autonomous Driving: LLM-Enhanced RL with Reusable Skills and Adaptive Exploration for Enhanced Safety
Authors:	Sirine Maaroufi, Ikbal Chammakhi Msadaa, Dimeth Nouicer and Khaled Grayaa
Abstract:	Ensuring safety, adaptability, and interpretability in autonomous driving remains a central challenge for reinforcement learning (RL). Although RL agents can acquire complex control policies through interaction with the environment, they often struggle with sparse rewards, poor generalization, and unstable exploration in safety-critical domains. Large language models (LLMs), by contrast, exhibit strong reasoning and abstraction capabilities but lack real-time adaptability. This work bridges these paradigms within the emerging vision of Agentic AI by introducing a hybrid framework that couples a Proximal Policy Optimization (PPO) agent with an LLM-guided SkillBank and intrinsic motivation. The SkillBank serves as a dynamic repository of validated driving maneuvers that can be reused across contexts, enhancing sample efficiency, stability, and interpretability. Concurrently, curiosity-driven intrinsic rewards promote structured exploration and continual self-improvement. Empirical results show that the proposed LLM-SkillBank agent achieves a 50% reduction in collision rate and a 73% increase in mean episode length compared to a standalone PPO baseline, reflecting substantial gains in safety and driving stability. Beyond quantitative improvements, the framework yields interpretable decision-making, as SkillBank activations correspond to human-recognizable maneuvers such as adaptive deceleration and lane changing. By integrating reasoning-guided skill reuse with intrinsically motivated exploration, this study advances a step toward safer, more adaptable, and transparent autonomous driving systems, narrowing the gap between symbolic reasoning and embodied learning in complex dynamic environments.
Download

Paper Nr:	403
Title:	SemanTEXT: Semantic Recognition of Text Labels in Indoor Environments for Navigation of Visually Impaired Individuals
Authors:	Dharmateja Adapa, Nishant Shubham, Virendra Singh Shekhawat and Avinash Gautam
Abstract:	Semantic recognition of objects in indoor environments enhances choice of destinations for the Visually Impaired (VI) in navigating complex indoor spaces. However, much of the information in such spaces is provided in the form of visual cues such as text/symbols. Information in these visual cues remains inaccessible to VI individuals. Recent frameworks built for conveying semantic understanding to VI users ignore textual/symbolic cues and only provide details for complete objects. We propose a framework that embeds text detection and understanding in an underlying semantic navigation framework to associate textual context with object instances for better goal selection in complex indoor environments. Since no standard dataset on indoor environments provides text recognition and localization data, we created our own dataset using indoor spaces in our university buildings. We tested the proposed method on our dataset and achieve 80% accuracy in text detection and absolute localization error of 0.287m in identifying and attaching textual descriptions to existing objects in the environment. Further, we implement the system on a robotic framework to test the end-to-end system of mapping and navigation.
Download

Paper Nr:	407
Title:	Estimating the Individual Importance of Communication Links for Distributed Optimization in Virtual Power Plants
Authors:	Jörg Bremer and Sebastian Lehnhoff
Abstract:	Distributed optimisation based on multi-agent systems has become a qualified solution for a range of complex problems in energy management – particularly, but not exclusively, in the field of virtual power plants. Many factors influence characteristics such as solution quality, message volume, or convergence behaviour, and thus performance in these approaches. One factor that has received surprisingly little attention to date is the communication network used by the agents for exchanging information during negotiation. Previous research has only considered the network type, but not the exact distribution of communication connections based on connected plant types and their individual contribution to optimisation. This paper presents an approach for estimating the individual importance of individual connections based on their marginal contribution, analogous to the concept of Shapley values.
Download

Paper Nr:	450
Title:	APODSS: An Agentic Pediatric Oncology Decision Support System
Authors:	Marco Aruta, Ciro Listone, Shilpa Srinivasareddy and Aniello Murano
Abstract:	Pediatric oncology poses unique diagnostic and prognostic challenges due to the complexity and rarity of childhood cancers. To address these, we propose APODSS, an Agentic Pediatric Oncology Decision Support System that assists clinicians in real-time, data-driven decision-making. APODSS integrates three specialized agents: a convolutional neural network (L-CNN) for leukemia image classification, a Cox Proportional Hazards model for survival prediction, and a Retrieval-Augmented Generation (RAG) Large Language Model (LLM) for context-aware reasoning. These agents are orchestrated via LangGraph, enabling transparent, traceable, and extensible workflows. The system achieves strong performance, with an F1-score of 0.97 for image classification and a C-index of 0.742 for survival prediction. The LLM agent operates locally via Ollama, grounded in semantically retrieved clinical guidelines from ChromaDB. Evaluation highlights APODSS’s ability to deliver low-latency inference and interpretable outputs across modalities. Future developments aim to broaden diagnostic coverage, incorporate real-world data, and enhance personalization.
Download

Paper Nr:	451
Title:	Safe Reward Learning from Human Preferences and Justifications
Authors:	Ilias Kazantzidis, Timothy J. Norman, Yali Du and Christopher T. Freeman
Abstract:	We address the problem of learning autonomous safe agent behaviour with unknown dynamics and reward functions, where traditional Reinforcement Learning is impossible. We present DROPJ, a human-centred algorithm that maximises safety during both training and deployment. We first learn a world model (a learned simulation) from a set of past real-world trajectories. A user then plays the game in the simulation to draw several informative virtual trajectories. From these, we extract pairs of trajectory segments and present them to a user to elicit their preference over these segments and the reason (justification) for that preference. With this feedback, a reward model is trained, which is used to deploy the agent with Model Predictive Control. We find that generating trajectories from user trials significantly reduces the computational cost of training, and significantly improves performance during deployment. In that context, we show that the use of preferences rather than other types of feedback substantially improves the performance. We further demonstrate that the use of justifications associated with safety requirements results in safer policies.
Download

Paper Nr:	454
Title:	Towards Secure and Transparent Practical Assessments in Computer Science: A Multi-Agent and Blockchain-Based Architecture
Authors:	Sondes Hattab and Faten Ziadi
Abstract:	Ensuring integrity and transparency in computer science practical assessments remains a major challenge, particularly in remote and large-scale contexts where the risk of fraud is significant. This paper proposes a secure and explainable architecture that integrates multi-agent systems (MAS) with blockchain technology to specifically prevent and detect fraudulent behaviours during programming exams. The proposed framework comprises four cooperative agents: a supervision agent for real-time monitoring of student activities, an analysis agent for detecting code similarity and AI-generated content, an integrity agent that records critical events on the blockchain to provide tamper-proof evidence, and a coordination agent that manages communication and decision-making among modules. This distributed and proactive approach supports automated fraud detection, immutable evidence, and interpretable decision-making, while preserving the human dimension of pedagogical evaluation. The paper also discusses key ethical and operational challenges, including privacy protection, latency in distributed recording, and user acceptability. Future work focuses on developing a functional prototype and conducting empirical testing in real university settings to assess detection accuracy, system performance, and pedagogical impact. The integration of MAS and blockchain technologies thus provides a trustworthy, adaptive, and transparent solution to enhance academic integrity and reduce fraud in practical assessments.
Download

Paper Nr:	471
Title:	Learning Engagement Assistant (LEA): A Multi-Agent AI Framework for Adaptive and Personalized Learning with Simulated Student Agents
Authors:	Teri Rumble, Javad Zarrin, P. George Lovell and Ruth Falconer
Abstract:	Recent advances in Artificial Intelligence and Large Language Models (LLMs) are enabling adaptive agent systems for personalized learning. However, adoption in higher education remains constrained by variability in course design, learner diversity, and the need for pedagogical alignment and instructor oversight. This paper presents the Learning Engagement Assistant (LEA)-a tri-modal, adaptive AI agent that delivers individualized instruction through integrated Chat, Tutor, and Quiz modes. LEA combines course-specific Retrieval-Augmented Generation (RAG) and Knowledge Component (KC) models to provide contextually grounded instruction and assessment to support scalability across instructional domains. A multi-agent orchestration dynamically adjusts task difficulty and scaffolding using learner performance, cognitive load estimation, zone of proximal development inference, and motivation tracking. The contributions are: (1) a pedagogically grounded orchestration framework integrating knowledge modeling and mastery estimation with tri-modal content generation; (2) a scalable knowledge representation pipeline achieved through modular course RAG knowledge bases and KC models; and (3) an evaluation framework with mode-specific performance metrics and simulated learner agents. Simulation findings demonstrate robust retrieval accuracy, coherent multi-turn tutoring, and adaptive stability across learner profiles and domain content, indicating that LEA can support pedagogical consistency across subject areas and dynamic learner-responsive support.
Download

Paper Nr:	479
Title:	ODiN: Offline Reinforcement Learning with Diffusion Policies for Bilateral Negotiation Strategies
Authors:	Yuji Kobayashi and Katsuhide Fujita
Abstract:	Negotiation, in which multiple agents with conflicting interests seek agreement, is a canonical multiagent task that requires advanced decision-making. Automated negotiation agents that learn strategies to replace humans have attracted growing interest. Reinforcement learning (RL) has been explored for autonomous strategy acquisition due to its versatility; however, conventional online RL often converges early to weak proposals and requires costly, large-scale simulators. To address these limitations, we propose ODiN (Offline Diffusion Negotiator), an offline RL agent that employs a diffusion-based policy for robust offline optimization. Its core variant extends Diffusion Q-Learning-originally developed to integrate behavioral imitation with reward-driven optimization in continuous action spaces-to discrete negotiation tasks by leveraging pre-collected logs. Extensive experiments across five negotiation domains confirm that ODiN-Q achieves substantial performance improvements, outperforming the CQL baseline by approximately a factor of two across all domains. Furthermore, diffusion-policy-based imitation learning consistently surpasses conventional imitation learning methods, demonstrating the effectiveness of diffusion-based approaches in negotiation environments. These results provide the first empirical evidence that diffusion-policy-based offline RL can operate effectively in automated negotiation and contribute to the development of practical, high-performance negotiation agents.
Download

Paper Nr:	480
Title:	iAgent-Based Context-Aware Adaptive Thresholding for Fault Tolerance in Wireless Sensor Networks
Authors:	Mouna Ktari and Ahmed Hadj Kacem
Abstract:	Wireless Sensor Networks (WSNs) require robust fault-tolerance mechanisms to ensure network longevity under strict energy constraints. This paper extends the iAFTA agent-based framework with context-aware adaptive thresholding for Cluster Head (CH) monitoring and replacement. Parameters α, γ, and δ are dynamically tuned based on cluster energy, density, and CH history across 4 configurations (Table 3). Simulations across standard (200 nodes) and large-scale (900 nodes, 15,000 rounds) scenarios demonstrate significant gains over baseline iAFTA: Combination 3 achieves +15.93% FND (Scenario 1), while Combination 2 yields +11.05% FND and -3.71% AEC (Scenario 5). Certain configurations reduce energy use by up to 7.21% but incur higher runtime overhead (+243.18%) in dense networks. These results validate the approach’s flexibility and scalability, though simulation-only evaluation and empirical tuning represent current limitations. The adaptive mechanism enhances fault tolerance while highlighting the need for context-specific parameter optimization in practical WSN deployments.
Download

Paper Nr:	489
Title:	Centralized-to-Decentralized Knowledge Transfer in Multi-Agent Reinforcement Learning via a Hybrid Execution Paradigm
Authors:	Shota Takayama and Katsuhide Fujita
Abstract:	Despite the effectiveness of Centralized Training with Centralized Execution (CTCE) for multi-agent coordination, its reliance on global information limits real-world applicability. Conversely, the Centralized Training with Decentralized Execution (CTDE) paradigm is often ineffective for complex coordination tasks. Here, we bridge this critical gap by introducing the Centralized-to-Decentralized (CtoD) learning concept, formalized through a novel Centralized Training with Hybrid Execution (CTHE) paradigm. The developed CtoD Multi-Agent Transformer (CtoD-MAT) method achieves knowledge transfer using curriculum learning as the core mechanism for gradually shifting agents from centralized to decentralized control. A dynamic scheduling mechanism, featuring a mediator module, ensures robust and effective knowledge transfer. Using challenging StarCraft Multi-Agent Challenge benchmarks, we demonstrate that CtoD-MAT successfully produces competitive decentralized policies, notably solving complex coordination tasks that are intractable using standard CTDE methods.
Download

Paper Nr:	520
Title:	An Agentic Outlook on Neuro-Symbolic AI: Towards Modular, Adaptive, and Explainable AI Systems
Authors:	Nourhan Ehab
Abstract:	Building AI systems that can reason, adapt, and explain their decisions remains a central challenge. Neuro-symbolic (NeSy) approaches combine neural learning and symbolic reasoning, yet most existing systems rely on static, pipeline-based architectures with limited coordination and adaptability. We introduce ANSA (Agentic Neuro-Symbolic Architecture), a framework that reinterprets neuro-symbolic systems through the lens of intra-system agency. In ANSA, perception, reasoning, orchestration, verification, explanation, and adaptation are implemented as autonomous, communicating agents that cooperate toward shared decision-making goals. Explicit communication, arbitration, and feedback mechanisms enable continual adaptation and verifiable coordination beyond conventional hybrid designs. We formally characterize ANSA as a class of agentic neuro-symbolic systems and instantiate it across three heterogeneous domains: medical diagnosis, financial advising, and chess commentary. Empirical results and expert evaluations show that ANSA improves explainability, adaptability, and trustworthiness compared to conventional hybrid baselines. We conclude by distilling design principles for building robust agentic neuro-symbolic systems.
Download

Paper Nr:	538
Title:	Evaluating LLM Alignment with Human Trust Models
Authors:	Anushka Debnath, Stephen Cranefield, Bastin Tony Roy Savarimuthu and Emiliano Lorini
Abstract:	Trust plays a pivotal role in enabling effective cooperation, reducing uncertainty, and guiding decision-making in both human interactions and multi-agent systems. Although it is significant, there is limited understanding of how large language models (LLMs) internally conceptualize and reason about trust. This work presents a white-box analysis of trust representation in EleutherAI/gpt-j-6B, using contrastive prompting to generate embedding vectors within the activation space of the LLM for diadic trust and related interpersonal relationship attributes. We first identified trust-related concepts from five established human trust models. We then determined a threshold for significant conceptual alignment by computing pairwise cosine similarities across 60 general emotional concepts. Then we measured the cosine similarities between the LLM’s internal representation of trust and the derived trust-related concepts. Our results show that the internal trust representation of EleutherAI/gpt-j-6B aligns most closely with the Castelfranchi socio-cognitive model, followed by the Marsh Model. These findings indicate that LLMs encode socio-cognitive constructs in their activation space in ways that support meaningful comparative analyses, inform theories of social cognition, and support the design of human–AI collaborative systems.
Download

Paper Nr:	572
Title:	Computational Political Philosophy
Authors:	J.-Martín Castro-Manzano and Jorge-A. Cervantes-Urbiola
Abstract:	In this contribution we present some models of political philosophy by using traditional multi-agent simulation techniques in order to introduce the concept of computational political philosophy -the political philosophy done with computational tools. Through comparative political philosophy, we encode five typical philosophical models as agent traits and simulate their interaction in a grid-world. We evaluate the simulation through six social metrics -participation, mobility, integration, cohesion, efficiency, and sustainability- and then we compare both the philosophical presets and the emergent social regimes via normalized profiles and a regime transition matrix.
Download

Paper Nr:	602
Title:	Using Reinforcement Learning Techniques for Autonomous Vehicles Navigation in Urban Intersections
Authors:	Felipe Merenda Izidorio and Gleifer Vaz Alves
Abstract:	Reinforcement Learning (RL) is an effective tool for solving complex challenges, such as the development of autonomous vehicles (AVs). Based on the Markov Decision Model and the Bellman Equations, RL algorithms allow vehicles, through continuous interactions with the environment, to learn to develop policies that choose the best actions for different situations, resulting in safer and more reliable vehicles. The aim of this work is to develop urban traffic environments for AV navigation and to evaluate different RL techniques, focusing on the agents' ability to make decisions efficiently and safely. Four different scenarios were developed at an urban intersection using the Unity platform. In these scenarios, vehicles were trained using the Proximal Policy Optimization (PPO) algorithm in its original form and in combination with additional methods: Curiosity, Generative Adversarial Imitation Learning (GAIL) and Random Network Distillation (RND), all available in Unity's ML-Agents tool. The results were analyzed with a focus on efficiency and vehicle safety. In general, after training, the average reward graphs and the duration of the episode per iteration showed a similar behavior between the techniques. However, in the tests, the combination of PPO with GAIL showed a more reliable behavior than the others, proving to be a promising technique for future studies with AVs. It is important to note that the results were based on the specific scenarios developed for this study, and the comparison of the techniques was done impartially, with the same parameters applied to all of them.
Download

Paper Nr:	617
Title:	LATInsights: LLM-Powered Conversational Agents for Geospatial Climate Analytics
Authors:	Diletta Chiaro, Davide Piu, Antonio Elia Pascarella, Paolo De Piano and Giovanni Giacco
Abstract:	Urban climate adaptation planning requires synthesizing complex Earth Observation data, yet traditional Planning Support Systems face an “implementation gap” due to technical barriers preventing non-specialist planners from accessing specialized geospatial workflows. We present LATInsights, a conversational agent system that addresses this gap through natural language interaction, enabling municipal planners to perform sophisticated thermal analytics without remote sensing expertise. The system implements a novel two-tier architecture: a stateful LangGraph-based orchestration engine managing semantic reasoning and multi-turn dialogue, decoupled from a stateless Model Context Protocol Server exposing domain-specific geospatial tools. This separation enables independent scalability while ensuring analytical accountability through structured planning and three-tier error recovery with Large Language Model-driven re-planning. We demonstrate operational efficacy through Milan’s Via Neera cycling corridor case study, where the system autonomously orchestrates multi-scenario thermal simulations. This work demonstrates how conversational AI can democratize geospatial intelligence for evidence-based urban climate governance.
Download

Paper Nr:	635
Title:	Unity Is Strength: Hybridizing Neuro-Symbolic AI and Agentic AI through Symbolic Coordination Mechanisms
Authors:	Alessia Papale and Federico Cabitza
Abstract:	Over the past decades, Neuro–Symbolic AI has sought to unify the strengths of neural networks and symbolic reasoning, promising systems that combine pattern learning with interpretable logic. Building on the well–known classification of neuro–symbolic systems proposed by Kautz, we introduce a seventh novel configuration called Symbolic(Neuro) that reframes the symbolic component as a meta–process coordination scaffolding among neural and human actors. This position paper outlines the conceptual foundations of this approach, grounded in the idea that cognition can be a social, collective and emergent phenomenon, and explores its theoretical and practical implications for the future of Agentic AI and Hybrid Intelligence.
Download

Paper Nr:	645
Title:	CertAI: A Certification Framework for Trustworthy and Secure Autonomous AI Agents
Authors:	Faisal Anwer, Mohammad Nadeem, Mohammed Abdullah Tahir, Jaafar Gaber and Salman Ali
Abstract:	AI agents are rapidly evolving from generating passive text to autonomously executing complex real-world tasks, such as booking travel, managing social media, and performing enterprise operations. Leading organizations including OpenAI, Google, Microsoft, and Salesforce are driving this transformation, positioning AI agents as the next major wave of generative AI technologies. As these systems gain autonomy and capability, effective security measurements and governance becomes essential to ensure their actions promote beneficial outcomes and prevent harm. This paper introduces CertAI, a certification framework designed to ensure the security, privacy, ethics, robustness, transparency and fairness of autonomous AI agents. CertAI generates verifiable certificates from structured metadata, embedding ethical, security and other parameters to include trust scores, risk levels, incident counts, and domain-specific compliance status. The framework is complemented by CertAI-Bench, a benchmark that systematically populates metadata through probing AI agents across multiple dimensions such as security, ethics, fairness, transparency, and others. Together, these components enable scalable certification, transparent evaluation, and continuous monitoring of autonomous AI systems, establishing a foundation for secure, responsible and verifiable AI governance. The certification process determines numerical values of the dimensions utilizing different models and calculates trust score. The results show that the agents based on larger models provide stronger safety guarantees, but fairness and transparency remain the weakest dimensions across all AI agents.
Download

Paper Nr:	685
Title:	MAFIA-NeT: Multi-Agent Framework for Interactive Agricultural Negotiation and Trading Systems
Authors:	Dung T. Phan, Chi N. L. Phan, Long S. T. Nguyen, Phuc T. Dao, Quan M. Bui, Tin T. Ngo, Thi T. Nguyen and Tho T. Quan
Abstract:	Large Language Models (LLMs) have demonstrated strong reasoning capabilities, yet they remain limited in sustained coordination and reliable real-world operation. These limitations have motivated the development of LLM-powered agents and LLM-based Multi-Agent Systems (MAS), which are better suited for domains that require adaptive reasoning and iterative decision-making. E-Agri commerce is one such domain, as agricultural negotiation involves dynamic pricing, diverse buyer requirements, and rapidly changing market conditions. Existing approaches mainly focus on prediction rather than interactive negotiation, making them insufficient for multi-turn bargaining scenarios. To address this gap, we propose MAFIA-NeT, a multi-agent framework for end-to-end agricultural trade negotiation. The system coordinates specialized agents for structured data parsing, market-informed price reasoning, and strategic decision support. A key contribution is the LLM-Guided Negotiation Subspace (LGNS), which uses an LLM-based anchor mechanism to identify a compact negotiation region and a lightweight regression model to estimate feasible pricing trajectories, enabling efficient negotiation without exhaustive search. MAFIA-NeT was deployed in collaboration with Agri Sung Joint Stock Company and evaluated on authentic B2B negotiation transcripts. Results show improvements in negotiation accuracy, price convergence, and operational practicality, demonstrating the promise of LGNS-guided multi-agent negotiation for scalable agricultural trade automation.
Download

Paper Nr:	689
Title:	Fairness Driven Multi-Agent Path Finding Problem
Authors:	Aditi Anand, Dildar Ali and Suman Banerjee
Abstract:	The Multi-Agent Path Finding (MAPF) problem aims at finding non-conflicting paths for multiple agents from their respective sources to destinations. This problem arises in multiple real-life situations, including robot motion planning and airspace assignment for unmanned aerial vehicle movement. The problem is computationally expensive, and adding to it, the agents are rational and can misreport their private information. In this paper, we study both variants of the problem under the realm of fairness. For the non-rational agents, we propose a heuristic solution for this problem. Considering the agents are rational, we develop a mechanism and demonstrate that it is dominant strategy incentive compatible and individually rational. We employ various solution methodologies to highlight the effectiveness and efficiency of the proposed solution approaches.
Download

Paper Nr:	704
Title:	SHARP: A Complete Multi-Task MAPD Planner for Realistic Warehouse Scenarios
Authors:	Taisei Hirayama, Kohei Yoshida, Hiroki Sakaji and Itsuki Noda
Abstract:	Conventional multi-agent pickup and delivery (MAPD) formalizations have not comprehensively addressed the operational realities of automated warehouses, including order-driven task structures, narrow single-agent-width corridors, dead-end workstations, and kinematic constraints. We extend multi-task MAPD by explicitly modeling order-driven task generation with distinct outbound and inbound task types, and introduce domain-specific constraints: the outbound completion sequence constraint and the inbound–outbound alternation constraint. We propose SHARP (Safe HAven Retreat Planner), which combines the safe haven retreat mechanism with Safe Interval Path Planning (SIPP), offering two strategies: Coupled SHARP (C♯) for simultaneous task allocation and path planning to maximize solution quality, and Decoupled SHARP (D♯) for improved computational scalability. Comprehensive evaluation over 9,600 trials on both simplified and constrained warehouse layouts shows all methods successfully completed all tasks in every trial, providing the first empirical demonstration of completeness in environments with narrow single-agent-width corridors and dead-end workstations under realistic warehouse constraints. C♯ achieves up to 14% reduction in total task completion time compared to the existing anytime MT-MAPD method, delivering the best performance, while D♯ achieves up to 8% reduction for configurations with 10 or more agents with superior computational scalability
Download

Paper Nr:	721
Title:	Building Intelligent Agents Based on the Clarion Cognitive Architecture: Some Essential Principles
Authors:	Can Serif Mekik and Ron Sun
Abstract:	This paper analyzes computational requirements for implementing hybrid neuro-symbolic agents based on the Clarion cognitive architecture. Clarion integrates explicit symbolic rules with implicit neural representations through a dual-representational structure and a phasic activation-based processing cycle, enabling heterogeneous and synergistic forms of reasoning and decision making. We examine how the architecture gives rise to distinctive implementation demands, including tight coupling between symbolic and neural knowledge representations, dynamic acquisition of new symbolic structures, modular organization, and precise temporal control. From this analysis, we derive five core software design principles that collectively support coherent, flexible, and scalable Clarion agents. We then illustrate how these principles are realized in pyClarion, a lightweight open-source Python library employing sparse dynamic numerical data structures with structured indices and a discrete-event simulation environment for building Clarion agents. The resulting approach provides a principled foundation for constructing next-generation neuro-symbolic intelligent agents.
Download

Paper Nr:	34
Title:	An Online Mechanism for Transfer Learning in Reinforcement Learning Scenarios
Authors:	Menaxi J. Bagchi and Shivashankar B. Nair
Abstract:	Transferring knowledge in reinforcement learning can reduce training time and computation by leveraging pre-trained neural networks instead of learning from scratch. This paper introduces an algorithm, namely, Neural network Selection for Transfer (NST), for determining whether to select a suitable pre-trained neural network (DQN model) from a repository to transfer knowledge from or learn from scratch. This paper also proposes an algorithm, namely, Performance-Aware Knowledge Transfer (PAKT), that identifies the appropriate time to discontinue the use of the pre-trained network. Deciding whether to employ a pre-trained neural network for knowledge transfer, and, if so, selecting the appropriate one and deciding when to stop using it, helps prevent negative transfer. Knowledge transfer is performed using both policy reuse and value function based techniques, with the proposed algorithms operating in conjunction with these methods. The robots execute these methods autonomously without any centralized controller. Experiments in Webots show that reusing an appropriate pre-trained neural network speeds up learning on the new task. Experimental results against benchmark methods validate the effectiveness of the proposed approach.
Download

Paper Nr:	38
Title:	Comparative Analysis of Multi-Agent Architectures for Entity Relationship Classification
Authors:	Maryam Berijanian, Kuldeep Singh and Amin Sehati
Abstract:	Entity relationship classification remains a challenging task in information extraction, especially in scenarios with limited labeled data and complex relational structures. In this study, we conduct a comparative analysis of three distinct AI agent architectures designed to perform relation classification using large language models (LLMs). The agentic architectures explored include (1) reflective self-evaluation, (2) hierarchical task decomposition, and (3) a novel multi-agent dynamic example generation mechanism, each leveraging different modes of reasoning and prompt adaptation. In particular, our dynamic example generation approach introduces real-time cooperative and adversarial prompting. We systematically compare their performance across multiple domains and model backends. Our experiments demonstrate that multi-agent coordination consistently outperforms standard few-shot prompting and approaches the performance of fine-tuned models. These findings offer practical guidance for the design of modular, generalizable LLM-based systems for structured relation extraction. The source code and dataset are available at https://github.com/maryambrj/ALIEN.
Download

Paper Nr:	64
Title:	From Chatbot to Validator: A Dual-Agent Strategy towards Trustworthy On-Premise Conversational LLMs
Authors:	Michal Podpora, Marek Baranowski, Aleksandra Kawala-Sterniuk, Mariusz Pelc, Piotr Rogala, Piotr Kawa, Pawel Pirog, Anna Romaniewska and Wojciech Rogala
Abstract:	This paper aims to inspire and guide industry leaders and decision-makers in successfully implementing LLM-based agents utilizing additional case-relevant knowledge, beyond LLMs’ original training. The technique presented in this paper, Retrieval-Augmented Generation (RAG), has already proven its practical value in real-world applications. As the RAG offers the injection of use-case-specific knowledge into LLM-based agents, making it possible to design customized expert systems with non-public knowledge. By customizing the expertise area, RAG becomes a powerful tool for businesses to pursue new objectives, and provide new services to their clients. This study presents a structured approach for the Polish Bielik 2.3 model deployment, integrating a Dual-Agent mechanism where one agent manages user interaction and the other ensures response validity and API compliance. The use of embeddings within the retrieval pipeline improves the computational efficiency and response time. The main research contributions are the Dual-Agent architecture and systematic evaluation of response time, benchmarked across five hardware configurations, from consumer laptops to a high-end server equipped with Nvidia H100.
Download

Paper Nr:	71
Title:	Accelerated Simulation Using Dummy Agents for Emergency Vehicle Routing
Authors:	Atsuo Ozaki and Daiki Akutagawa
Abstract:	The annual increase in emergency vehicle operations along with longer response and transport times has become a significant social issue. This study proposes a new route search method based on the operational behavior of emergency vehicles. In contrast to heuristic approaches, this method employs multiple dummy agents to explore all possible routes simultaneously. Comparison with a brute-force search method on a simple road network confirmed that an increased probability of deriving optimal routes was achieved under uneven traffic conditions. Additionally, the proposed method demonstrated a performance improvement of over an order of magnitude compared to the brute-force search method, with further improvements as the problem scale increased.
Download

Paper Nr:	131
Title:	Agent-Based Detection and Resolution of Incompleteness and Ambiguity in Interactions with Large Language Models
Authors:	Riya Naik, Ashwin Srinivasan, Swati Agarwal and Estrid He
Abstract:	Large Language Models are considered as modern-day oracles, but for Long multi-turn interactions can get tedious if it is simply to clarify contextual information that can be arrived at through reasoning. In this paper, we examine the agent-based architecture to bolster LLM-based Question-Answering (QA) systems with additional reasoning capabilities. Specifically, we use LLM-based agents to automatically detect and resolve incompleteness or ambiguity in questions. Our study focuses on benchmark datasets containing such deficiencies, and equips GPT-3.5-Turbo and Llama-4-Scout with zero-shot ReAct agents that operate as transducers. Rather than producing an answer in a single step, the language model now decides between 3 actions: (a) classify a question as incomplete, ambiguous, or normal; (b) resolve deficiencies if present; and (c) answer the reformulated question. We compare the use of LLMs with and without agents that incorporate these components. Results show three main benefits: (1) A shortening of the length of interactions with human; (2) An improved answer quality; and (3) Explainable resolution of deficiencies in the question. The trade-off is additional LLM calls, sometimes increasing latency. But on tested datasets, the benefits outweigh the costs except when questions already have sufficient context. Suggesting the agent-based approaches offer a practical mechanism to enhance robustness and usability of LLM-driven QA systems.
Download

Paper Nr:	215
Title:	A Distributed and Automated System for DDoS Identification and Mitigation in UAV Networks
Authors:	Naser Abbas Hussein, Khadija Rammeh Houerbi and Hella Kaffel Ben Ayed
Abstract:	As Unmanned Aerial Vehicles (UAVs) are increasingly integrated into critical applications, their susceptibility to cybersecurity threats, particularly Distributed Denial of Service (DDoS) attacks, has escalated. This paper introduces a novel blockchain-based architecture employing Ethereum smart contracts to fortify UAV net- works. The system integrates five coordinated modules: secure UAV registration, reputation-driven trust eval- uation, real-time anomaly detection, dynamic threat mitigation, and centralized orchestration. In experimental simulations, the model achieved near-perfect detection accuracy in detecting varied DDoS patterns—40% bandwidth abuse, 30% packet floods, and 20% traffic anomalies, High Error Rate anomalies represented 10% of detections. UAVs retained stable reputation levels despite persistent attacks, with mitigation strategies scaled according to threat severity. This decentralized solution enhances network resilience by eliminating centralized points of failure and enabling autonomous, verifiable defense.

Paper Nr:	235
Title:	C-STEP: Continuous Space-Time Empowerment for Physics-Informed Safe Reinforcement Learning of Mobile Agents
Authors:	Guihlerme Daubt and Adrian Redder
Abstract:	Safe navigation in complex environments remains a central challenge for reinforcement learning (RL) in robotics. This paper introduces Continuous Space-Time Empowerment for Physics-informed (C-STEP) safe RL, a novel measure of agent-centric safety tailored to deterministic, continuous domains. This measure can be used to design physics-informed intrinsic rewards by augmenting positive navigation reward functions. The reward incorporates the agent’s internal states (e.g., initial velocity) and forward dynamics to differentiate safe from risky behavior. By integrating C-STEP with navigation rewards, we obtain an intrinsic reward function that jointly optimizes task completion and collision avoidance. Numerical results demonstrate fewer collisions, reduced proximity to obstacles, and only marginal increases in travel time. Overall, C-STEP offers an interpretable, physics-informed approach to reward shaping in RL, contributing to safety for agentic mobile robotic systems.

Paper Nr:	295
Title:	The Dynamics of Trustworthiness Evaluation in Multi Agent Systems
Authors:	Frédérique Lalieu, Tomasz Zurek and Tom Van Engers
Abstract:	The aim of this research is to develop a model for the evaluation of trustworthiness factors in a multi-agent system (MAS) in response to external signals. In our model, trustworthiness is composed of benevolence, competence and integrity. Messages from other agents and observations, collectively called ‘external signals’, can influence an agent’s belief about someone’s benevolence, competence and integrity. The model was implemented in the ASC2 framework, a domain-specific language (DSL) and cross-compiler created for modeling MAS’s consisting of BDI-agents.
Download

Paper Nr:	319
Title:	POSG–MARL for Corporate BTC Treasuries: Strategic Allocation under Informational Asymmetry
Authors:	Ioannis T. Thomaidis, Nikolaos P. Rachaniotis and Thomas K. Dasaklis
Abstract:	We develop a Partially Observable Multi-Agent Reinforcement Learning (POSG–MARL) framework to model strategic corporate Bitcoin (BTC) treasury allocation under informational asymmetry. Each quarter, firms jointly choose BTC exposure and disclosure opacity, knowing rivals will infer and react to public signals filtered through noise and belief inertia. This transforms treasury policy into a dynamic game of exposure vs. disclosure, where payoffs balance excess return over cash against volatility risk, trading frictions, and event-amplified costs of opacity. Policies are trained with Multi-Agent Proximal Policy Optimization (MAPPO) under Centralized Training, Decentralized Execution (CTDE) using squashed Gaussian actions and turnover caps, while a centralized critic consumes a compact public summary and last joint actions. The result is a practical POSG–MARL framework that maps volatility and disclosure–belief dynamics to rational disclosure–allocation rules. The framework provides paper-ready diagnostics for disclosure–belief coupling, volatility response, turnover behavior and reward decomposition, enabling transparent analysis of signaling and allocation choices in corporate BTC treasury management.
Download

Paper Nr:	350
Title:	Self-Organized Restoration of Multi-Energy Grids via Coupling Point Coordination
Authors:	Rico Schrage, Emilie Frost and Astrid Nieße
Abstract:	The resilience of multi-energy grids is increasingly critical due to the interdependence of electricity, heat, and gas infrastructures and their vulnerability to destructive events. This paper presents a self-organized restoration scheme based on multi-agent systems (MAS) for impaired energy systems, enabling decentralized coordination through local balancing and distributed optimization of coupling points. The approach uses gossip-based consensus for intra-sector balancing and a formulation based on the Alternating Direction Method of Multipliers (ADMM) for cross-sector coordination, explicitly modeling communication delays to capture real-world constraints. We evaluate the methodology on two benchmark grids under diverse failure and prioritization scenarios. Results demonstrate that the agent-based scheme achieves restoration performance comparable to a centralized baseline while maintaining robustness across failures and respecting sectoral priorities. These findings highlight the potential of self-organized systems for scalable, autonomous, and resilient restoration of multi-energy systems.
Download

Paper Nr:	423
Title:	Trust Calibration through Role-Aware and Causal Explanations: An Extension of HAExA for Smart Building Multi-Agent Systems
Authors:	Narendar Kumar, Yazan Mualla, Vincent Hilaire and Stéphane Galland
Abstract:	As AI-driven multi-agent systems (MAS) are increasingly deployed in smart buildings, transparent and trustworthy human–agent collaboration is essential. Building on the Human-Agent Explainability Architecture (HAExA), this position paper presents an extension tailored to smart buildings with three modules: role-aware filtering (adapting explanation content to occupants, facility managers, and security staff), privacy guardrails (redacting sensitive data while exposing policy-level signals), and an outcome feedback loop (linking ex-ante predictions to ex-post results to calibrate trust over time). This contribution enable explanations that are parsimonious, role-sensitive, and privacy-preserving, while embedding causal accountability into the interaction cycle. We outline open research challenges including role alignment, longitudinal evaluation, and LLM integration towards trustworthy human centered smart building ecosystems.
Download

Paper Nr:	491
Title:	Multi-Agent System for Collaborative Fault Diagnosis in Multi-Stage Manufacturing Environments
Authors:	Felipe Merenda Izidorio, Paulo Leitão, José Barbosa and Gleifer Vaz Alves
Abstract:	Modern multi-stage manufacturing demands sophisticated fault diagnosis to maintain quality across production stages. Traditional single-method approaches operate reactively, failing to exploit correlations between multiple variables or leverage complementary analytical techniques. This work presents a generic multi-agent system integrating multiple machine learning algorithms for collaborative diagnosis in geometrical inspection processes. The architecture comprises three layers: Point Agents applying complementary algorithms to individual measurements, Station Agents coordinating through correlation analysis and clustering, and an Inter-Station Agent providing cross-sectional perspective. The experimental validation using real automotive production data across two stations demonstrates the effectiveness across three configuration profiles. Results reveal fundamental trade-offs between detection sensitivity and false alarm rates, with conservative configurations prioritizing specificity while aggressive configurations maximize the early detection. The system demonstrates consistent early warning capabilities with 4-5 unit lead times and computational performance compatible with production cycle times in the validation scenario. This generic architecture also enables distributed collaboration between heterogeneous techniques, demonstrating broad applicability across diverse manufacturing contexts.
Download

Paper Nr:	503
Title:	Improving Implicit Hate Speech Detection via a Community-Driven Multi-Agent Framework
Authors:	Ewelina Gajewska, Katarzyna Budzynska and Jarosław A. Chudziak
Abstract:	This work proposes a contextualised detection framework for implicitly hateful speech, implemented as a multi-agent system comprising a central Moderator Agent and dynamically constructed Community Agents representing specific demographic groups. Our approach explicitly integrates socio-cultural context from publicly available knowledge sources, enabling identity-aware moderation that surpasses state-of-the-art prompting methods (zero-shot prompting, few-shot prompting, chain-of-thought prompting) and alternative approaches on a challenging ToxiGen dataset. We enhance the technical rigour of performance evaluation by incorporating balanced accuracy as a central metric of classification fairness that accounts for the trade-off between true positive and true negative rates. We demonstrate that our community-driven consultative framework significantly improves both classification accuracy and fairness across all target groups.
Download

Paper Nr:	557
Title:	Evaluation of Task Assignment Strategies for Capacitated Multi-Agent Pickup and Delivery in Automated Sortation Systems
Authors:	Evren Çilden and Faruk Polat
Abstract:	Automated sortation has emerged as a major trend in logistics, enabling scalable and time-efficient sorting of items through the deployment of robotic agents. This study examines the use of capacity-enhanced agents in automated sorting domain, by evaluating task assignment strategies that improve the Token Passing with Multiple Capacity (TPMC) algorithm for the Multi-Agent Pickup and Delivery with Capacities (MAPDC) problem. In a simulated sorting domain, eight methods leveraging agent waypoint information are evaluated against the commonly used nearest-pickup task selection method. The results show that task assignment heuristics incorporating Closeness Centrality, Hausdorff Distance, and cost-based measures significantly improve solution quality of path planning in sortation systems. Moreover, the gains are greater in the sortation domain with disjoint pickup and delivery areas than in automated warehouse settings without spatial constraints, highlighting the importance of environment structure in MAPDC.
Download

Paper Nr:	667
Title:	An Expertise-Aware Framework for Usefulness of Information in Education and Lifelong Learning
Authors:	Célia da Costa Pereira
Abstract:	We introduce a dynamic, expertise-aware framework that extends traditional usefulness measures to the educational and career development domains. By modeling expertise as a graded, evolving continuum-rather than a binary state-our approach enables a fine-grained assessment of how training opportunities align with users’ professional goals. This framework not only evaluates immediate skill acquisition but also accounts for long-term career progression, offering a personalized and adaptive tool for learners and organizations alike. We have demonstrated the functionality and validity of our framework through illustrative examples and a detailed case study, showcasing its practical applicability and effectiveness in real-world scenarios.
Download

Paper Nr:	679
Title:	Knowledge-Graph-Based Chatbot: An Application of Anonymized Electronic Health Records for Decision-Making in Maternal Care
Authors:	Daniel Izquierdo, Jairo J. Pérez and Juan C. Botero
Abstract:	Electronic Health Records (EHRs) are essential for documenting clinical care. Despite this, its data is fragmented across heterogeneous formats and large volumes of unstructured text hinder timely information obtention and decision-making in maternal contexts. This work presents a knowledge-graph–based question– answering system, the main part of a chatbot, that integrates Large Language Models (LLMs) with a clinical Knowledge Graph built from anonymized EHRs standardized to the OMOP Common Data Model (v5.4). The dataset includes 23,853 obstetric patients and over seven million clinical entities, transformed into nodes and semantic relationships within Neo4j AuraDB to ensure traceability and explicit representation of clinical entities and events. A Neo4j-powered agent was implemented using multiple LLM variants from the Gemini family, equipped with tools to inspect the graph schema and execute Cypher queries over the OMOP-based graph. System performance was evaluated using a curated control dataset of clinically grounded questions, assessing accuracy, completeness, factual consistency, and response latency. The results reveal clear performance differences across model tiers: higher-capacity models consistently achieved precision above 85 % and recall close to 84 %, producing more complete and clinically coherent answers, while lighter models offered faster responses at the expense of retrieval quality. Overall response latency ranged from tens of seconds per query, reflecting the trade-off between reasoning depth and execution time. These findings demonstrate the feasibility of combining knowledge graphs and LLM-based agents to provide structured, interpretable, and traceable access to large-scale clinical data, supporting the development of reliable decision-support tools for maternal care.
Download