Penetration testing (PT) using reinforcement learning (RL) is challenged by large action spaces, sparse rewards, and slow convergence. This study presents modular RL framework integrating Double Deep Q-Network (DDQN) and Soft Actor–Critic (SAC) with architecture decoupling (D) and Two-Stage Action Selection (2SAS) to enable structured decision-making in high-dimensional environments. The framework separates high-level attack planning from low-level execution, improving scalability and learning efficiency. To enhance exploration and sample efficiency, the approach combines Intrinsic Curiosity Module (ICM) and Random Network Distillation (RND) with Prioritised Experience Replay (PER) and Hindsight-inspired relabeling (HER′). Experiments on NASim medium scenario demonstrate that our RIPH-D-DDQN significantly outperforms NDSPI-D, achieving up to + 60.3
Drug-drug interactions (DDIs) occur when two or more medications interact within the body, producing effects that range from reduced therapeutic efficacy to harmful adverse reactions such as elevated blood pressure or excessive sedation. Yet existing machine learning (ML) based computational approaches predominantly focus on binary edge prediction or relation classification. Hence, DDI prediction without providing mechanistic explanations remains a critical challenge for patient health. In this research work, we propose Causal Mechanism-Disentangled Relational Variational Autoencoder (CMR-VAE), a novel permutation-invariant and graph-regularized generative framework that jointly models DDI edges and their natural-language mechanism descriptions. Unlike prior methods, CMR-VAE disentangles latent factors into (1) drug-agnostic interaction mechanisms and (2) pair-specific interaction strength, enforced through β -TCVAE regularization, weakly supervised mechanism cues, and graph contrastive learning. Our model leverages a Deep Sets encoder to ensure order invariance in drug pairs, incorporates graph Laplacian smoothing for embedding coherence, and employs counterfactual text masking to prevent drug identity leakage. Comprehensive experiments on a large-scale DDI dataset (191k interactions, 1.7k drugs) demonstrate that CMR-VAE achieves state-of-the-art predictive performance (AUROC ≈ 0.999 in both pair-inductive and drug-inductive settings, validated with 5-seed mean ± std and external benchmarking), while yielding interpretable mechanism labels (micro-F1 ≈ 0.98) and generating concise mechanistic rationales. Beyond accurate DDI detection, CMR-VAE supports mechanism annotation, explanation generation, and cold-start generalization for rare drugs, offering a unified, mechanism-aware foundation for safer pharmacological vigilance and automated drug interaction discovery.
Support Vector Machines (SVMs) can be sensitive to corrupted training labels, which may significantly degrade classification performance. This study uses robust optimization as the main research method to model and protect the classifier against label uncertainty. Robust optimization provides a natural framework to address this issue by protecting the classifier against label perturbations. However, existing robust SVM formulations under label uncertainty typically result in large mixed-integer optimization models. In this study, we propose a compact mixed-integer reformulation for SVM classification under label uncertainty. The proposed model reduces the number of binary variables, while preserving the robustness properties of the original formulation. Computational experiments on synthetic datasets show that the compact formulation significantly decreases solving time. Additional experiments on benchmark datasets from the UCI Machine Learning Repository, including SUSY, Skin Segmentation, Covertype, Census-Income, Adult, CNAE-9, and Banknote Authentication, further demonstrate the computational advantage of the proposed formulation on real-world data. Simulation results under random label perturbations further demonstrate improved robustness as the uncertainty budget increases.
Traffic management in dense urban environments demands accurate, real-time, and interpretable forecasting systems capable of reasoning over complex spatial and temporal relationships. This study presents DTKGC and STaRNS (Spatio-Temporal Augmented Relational Neuro-Symbolic framework), a two-phase integrated pipeline for intelligent traffic management applied to the Philippine road network. Phase 1 automates structured extraction of traffic events from geotagged tweets using a fine-tuned GPT-4o model augmented with schema-guided prompts, Retrieval-Augmented Generation (RAG), and Chain-of-Verification (CoVe), achieving an entity extraction F1 of 95.8
The rapid advancement of generative artificial intelligence has made fake news more persuasive, scalable, and difficult to detect. This creates an urgent need for detection models that remain reliable under dynamic news environments. However, most existing approaches rely on supervised learning with static datasets, which limits their ability to generalize to unseen writing styles, emerging topics, and shifting source characteristics. To address this limitation, this study proposes a reinforcement learning-based framework for fake news detection that treats the task as a multi-objective decision problem. The proposed method extends pretrained detection models with a multi-component reward function that incorporates prediction accuracy, source reliability, simulated user feedback, and uncertainty-aware penalties. This design enables the model to optimize multiple reliability-related criteria beyond fixed supervised objectives. In addition, a reward-aware attention mechanism is introduced to guide the model toward contextually important cues that contribute to reliable veracity judgments. Experimental results on benchmark datasets show that the proposed framework improves F1 score, cross-domain performance, and confidence calibration compared with strong supervised baselines. These findings suggest that multi-objective optimization provides a more effective and flexible approach for improving both the adaptability and reliability of fake news detection systems in evolving online environments.
Biodiversity loss is one of the most pressing sustainability challenges of our time. Many wildlife species are threatened, and effective conservation depends on our ability to monitor them accurately and at scale, including the (re-)identification of individual animals (animal re-identification). The Eurasian badger is one such species facing notable pressures. Despite this, the species currently lacks a dedicated dataset to support automated re-identification. In this study, we introduce a novel dataset designed to address this gap. We also evaluate, for the first time, on Eurasian badgers (a cryptic species lacking conspicuous visual markings), a state-of-the-art AI foundation model for animal re-identification, MegaDescriptor. Our dataset comprises both daytime and nighttime camera-trap images, capturing multiple individuals across different encounters. Furthermore, results show that model fine-tuning can lead to substantial improvements in re-identification performance, and that these gains seem to hold when fine-tuning is done using greyscale imagery, including results for nighttime re-identification. We publicly release the dataset and the fine-tuned model to support future research on Eurasian badger re-identification, aiming to advance the development of AI-enabled passive monitoring tools for this species.
Understanding how intrinsic motivation (IM) and test anxiety (TA) vary among students is critical for supporting self-regulated learning in constructivist Computer Science (CS) education; however, there is limited empirical evidence on how these factors differ across demographic, academic, and lifestyle contexts in Leaving Certificate Computer Science (LCCS). This study examines patterns of association between IM and TA, and a range of student-related factors, in an Irish upper-second-level CS context. Data were collected across 94 national outreach sessions (2020–2022), yielding 734 valid responses, with selected items drawn from the Motivated Strategies for Learning Questionnaire (MSLQ). Descriptive analysis showed high IM (M = 3.93, SD = 0.58) and moderate, variable TA (M = 3.33, SD = 0.89). Statistical analyses identified significant associations between IM and TA scores and factors including gender, prior CS experience, mathematics level and achievement, science subject choice, social media use, and teacher gender, with small-to-moderate effect sizes. Additional interaction analyses demonstrated that IM and TA vary across combinations of factors, particularly mathematics level, school gender, and science pathways, as well as lifestyle patterns involving gaming and social media. These findings suggest that IM and TA are associated with a range of intersecting demographic, academic, and lifestyle characteristics. However, as the analysis is correlational, the results indicate associations rather than cause-and-effect relationships. The results contribute to CS education research by highlighting patterns that may inform the design of targeted supports to sustain motivation and reduce anxiety in constructivist learning environments such as Leaving Certificate Computer Science.
Large Language Models (LLMs) are widely used to automate complex workflows. While such LLMs are meant to provide efficiency improvements, a key factor is trust and perceived control over these systems. Using such systems, end-to-end, without human intervention may not always yield the desired output, especially when compliance and financial decisions are involved in e-commerce or the onboarding of fintech platforms. One question that arises is when should artificial intelligence act on its own and when it should wait for human input. We report a controlled experiment between different subjects, where (N = 69) participants were subjected to different autonomy governance variables in an LLM-assisted enterprise onboarding workflow and its effects on perceived workload, task completion time, usability, and trust ratings. The participants had to complete a task that simulated joining a fintech platform under one of three conditions: (A) manual interaction; (B) confidence-displayed agentic automation, wherein the system gave structured workflow inputs while displaying the prediction certainty; and (C) confidence-governed automation, in which agent actions were selectively enacted only when certainty exceeded a predefined threshold. Results show a statistically significant effect of autonomy strategy on user trust ratings. Confidence-displayed automation condition contributed in significantly lower trust ratings than both manual interaction and confidence-governed automation, while no statistically significant difference was measured between manual and confidence-governed conditions. Perceived workload and usability metrics, measured using NASA-TLX and UMUX-Lite, did not differ between the test groups. These findings establish autonomy governance as a mechanism for shaping trust in agentic workflows.
In today’s digital economy, knowledge is one of the most valuable assets, and its effective sharing is essential for fostering innovation and economic growth. In Saudi Arabia, knowledge sharing plays an important role in achieving Vision 2030 by promoting academic excellence and supporting the transition to a knowledge-based economy. However, many existing Learning Management Systems lack the integration, scalability, and cross-institutional collaboration required to meet evolving higher education needs. Limited collaboration tools, inefficient course management, and weak support for research networking continue to hinder digital transformation initiatives. To address these challenges, this study proposes the Saudi Arabia Higher Educational Institutes Knowledge Sharing Platform (SA-HEIKSP), a cloud-based platform designed to support collaboration among students, faculty members, researchers, and administrators across Saudi universities. The platform integrates course management, interactive discussion forums, and research collaboration features within a unified environment. A preliminary mixed-method evaluation involving participants from Saudi higher education institutions indicated positive support for the platform and its potential to improve collaboration, knowledge sharing, and academic engagement. SA-HEIKSP provides a scalable model for strengthening inter-university cooperation and supporting Saudi Arabia’s transition towards a knowledge-driven economy.
Many machine learning studies on autism spectrum disorder (ASD) screening report near-perfect accuracy on public benchmarks, but accuracy alone is a weak signal of clinical readiness. We present REX-ASD, an evaluation protocol covering robustness, explainability, fairness, and statistical rigour. A feature leakage audit sits at the core of the findings. Point-biserial correlation and permutation importance flagged five features—most prominently the composite screening score—that carry direct information about the label. Once those features are removed, the area under the receiver operating characteristic curve falls from roughly 1.00 to 0.91–0.94, and the harmonic mean score from near 1.00 to 0.75–0.86. Near-ceiling scores appear to reflect dataset artefacts rather than genuine learning. Six classifiers were evaluated on the cleaned University of California, Irvine (UCI) adult and child ASD datasets. The protocol covers cross-dataset transfer (Adult ↔ Child), Gaussian noise perturbation at seven intensities from 0
Generative AI (GenAI) systems are predominantly accessed through detached, prompt-first interfaces that operate outside users’ primary file environment. While this model lowers entry barriers by retaining a familiar text-input pattern that resembles search, it structurally separates AI invocation from established desktop interaction workflows and introduces interaction overhead before any model response is generated. This study investigates whether AI invocation can be made more efficient when embedded directly at the point of file interaction instead of behind a detached prompt-first workflow. We present an adaptive operating-system-level interface that activates when a file is dragged toward the macOS notch, exposing contextual smart labels (e.g., summarize, translate, rephrase) for direct manipulation of document content. A within-subjects study (n = 30) compared participants’ habitual ChatGPT workflow with the prototype across three standardised document-based tasks: file attachment, summarisation, and follow-up rephrase/translate. Model response latency was excluded, so the analysis isolates pre-response interaction overhead. Across all three tasks, paired-samples Wilcoxon signed-rank tests showed that every participant completed the prototype condition faster (Wilcoxon signed-rank, all p < 0.001; the negative-rank sum was zero in every task because no participant was slower in the prototype), with mean reductions of Δ = 17.69 s, 32.55 s, and 41.52 s respectively, large effect sizes (Cohen’s d_z = 1.18–2.12), and lower variability than baseline. Qualitative feedback was broadly positive but also surfaced critical observations regarding predefined prompts and missing recovery affordances. Findings are limited to short document-based tasks with a single one-page PDF and do not evaluate model output quality or longitudinal adoption. The results suggest that interaction placement is a meaningful design dimension for Human–AI Interaction in document-bound workflows.
Focusing on the main research challenge of supporting Online Analytical Mining (OLAM)-based Big Data Analytics over Big Data Warehouses, this study introduces and experimentally assesses ℳℛℰ-𝒦𝒟𝒟^+ , an innovative Multi-Resolution Ensemble-based Knowledge Discovery from Data. ℳℛℰ-𝒦𝒟𝒟^+ incorporates several innovations, including a multi-resolution OLAP cubing function and an ensemble-based knowledge fusion model. A comprehensive experimental assessment and analysis of the framework performance is presented and discussed, demonstrating its scalability and efficiency against state-of-the-art systems such as WEKA and SQL Server Analysis Services (SSAS), thereby significantly integrating conceptual and practical contributions of our research.
Multimodal interaction in transactional systems raises a mixed-initiative design question: should users or AI systems decide whether users click, speak, or use chat at each step of the workflow? A controlled experiment was conducted using a multi-step gifting checkout flow to examine the effects of AI governing the input modality in multimodal interfaces on participants’ performance, usability, and cognitive experience. Participants completed identical gifting workflows under three conditions: (A) traditional on-screen interaction (B) user-selected input modality (choosing between traditional, speech, or chat input per step), and (C) AI-selected modality (system-assigned input modality for each step). When the LLM governed modality selection, participants took significantly longer to complete tasks ( η ^2 = 0.416) and reported lower ratings of usability, satisfaction, trust, and perceived control than in the user-selectable condition. When participants governed their own modality choice, task completion performance was comparable to that of traditional on-screen interaction; trust ratings and perceived control remained high; and perceived workload did not differ across conditions. Notably, despite the availability of multimodal options, participants relied exclusively on standard on-screen input, without adopting voice or chat modalities. These results demonstrate a gap between modality availability and actual usage. Finally, AI-governed interaction is not inherently beneficial when decision authority over input modality transfers without constraint.
This research examines whether an AI-based academic advising chatbot can improve service efficiency and student experience compared with traditional advising channels. The purpose of the study is to determine the effectiveness of an AI-powered academic advising chatbot in enhancing the efficiency of service provision and student experience compared to traditional advising practices. An experimental design was created and assessed to develop and test a prototype chatbot (AdvisingBot): 30 students participated in the experiment, with response latency, resolution rate, and System Usability Scale (SUS) scores as the performance indicators. A functional prototype, AdvisingBot, was developed incorporating Natural Language Understanding (NLU), TF-IDF retrieval, rule-based dialogue management, and analytics logging. The system was evaluated through an experimental design comprising real-time interaction testing and a usability survey administered to an independent sample of 30 students using the System Usability Scale (SUS). System metrics revealed low response latency and a moderate resolution rate, demonstrating operational viability for handling routine advising enquiries. Inferential analysis found no statistically significant difference in usability scores between the chatbot and baseline advising conditions. The findings show that the chatbot recorded low response latency and moderate rates of resolution, but there was no statistically significant difference in usability between the chatbot and traditional advising techniques. These results offer initial indications of the practicability of AI-based advising systems and validate the possibility of human-AI models of services in higher education. Descriptively, however, participants reported broadly positive perceptions of the chatbot’s responsiveness and convenience. This research contributes to the growing literature on AI-enabled student services by examining the feasibility, limitations, and institutional considerations for adopting chatbot-based advising and by providing recommendations for future development.
The majority of artificial intelligence (AI) modelling of cranial implant reconstruction is effective with in-domain raw datasets but can fail when considering new or altered conditions, such as augmented data. This study presents a comprehensive end-to-end evaluation of the most recent developments in algorithms for cranial implant reconstruction, with an emphasis on their adaptability, robustness, and capacity for generalization after training and testing on expanded and cross-domain datasets. A synthetic, control dataset of cranial defects has been generated using a range of augmentation strategies: intensity-based, geometric, and morphological to simulate real variability in patients’ response to treatment. Two different model architectures were evaluated, each included attention-enhanced 3D U-Nets and encoder-decoder networks, and each was structured to contain systematic benchmarking to allow analysis of how both model architecture and data augmentation influenced fidelity of reconstruction while preserving symmetry. Experimental results show that the 3D U-Net significantly outperformed the VAE, achieving a Dice score of 0.919 compared to 0.340, a Boundary Dice of 0.942 versus 0.296, and a substantially lower HD95 of 1.246 compared to 15.393. After the 3D reconstruction stage, the study extended the evaluation to include exploring techniques for structural optimization by which to improve the anatomical fit and material efficiency, such as the strength-to-weight ratio, of implants. The optimized designs were then manufactured via 3D printing, and their real-world viability was assessed and validated using morphological and mechanical analyses. Such an integrated investigation was able to provide new insights into the interactions between the architecture of the model, data augmentation, and optimization of design, assisting the integrated development of intelligent, topology-aware fabrication of cranial implants.
Cyber-Physical Systems (CPS) demand rigorous modeling and verification to ensure their trustworthiness. This study presents a framework for the formal modeling and automated verification of CPS using the hierarchical, component-based language MedTiny. We demonstrate MedTiny’s expressiveness in capturing the intertwined cyber and physical behaviors through a detailed case study of a Debounced Toggle mechanism, complete with Linear Temporal Logic (LTL) property specifications. To enable formal verification, we introduce the MedTiny2SMT compiler, a key contribution that automatically transforms MedTiny models into constraints for the SMT solver Z3, reducing property verification to an SMT satisfiability problem. Furthermore, our framework supports practical downstream workflows, including executable code generation and integration with proof assistants. The results establish MedTiny as a practical language for designing verifiable CPS, bridging high-level modeling with powerful automated reasoning tools.
Online learning algorithms offer a scalable, more biologically plausible alternative to Back-Propagation-Through-Time (BPTT) for training recurrent Spiking Neural Networks (SNNs). Of these proposed online learning methods, e-prop has emerged as an interesting candidate, not least because of its efficiency compared to some other online algorithms, but also its compatibility with neuromorphic processors and conventional Artificial Neural Networks (ANNs). However, despite these advantages, it has not been studied with regard to its sensitivity to different arithmetic formats, especially under reduced precision. This study presents a systematic evaluation of floating-point, and posit arithmetic formats at 16-bit and 8-bit precision for e-prop training on the Spiking Heidelberg Dataset. Our results show that 16-bit posit formats can match or exceed 32-bit floating-point (FP32) accuracy for this dataset. Bfloat16 provides limited performance and IEEE 16-bit floating-point (FP16) fails to support training. All 8-bit formats are largely ineffective. Analysis reveals that both dynamic range and precision are critical for stable online learning. These findings highlight the importance of arithmetic choice and position posits as a strong candidate for efficient neuromorphic implementations in the future.
Context-aware mobile apps must sense, interpret, and react to changing conditions (location, connectivity, device state) while remaining portable and maintainable. We present CAMS+GenAI, a model-driven pipeline that couples a Generative DSL (DSL-G) with LLM-assisted synthesis to specify context models, rules, and connectors, and to generate Flutter code and infrastructure-as-code for a hybrid edge–cloud runtime. The pipeline includes promptable templates, static checks for rules, scenario-based simulation for offline conditions, and multicloud deployment targets (Firebase/AWS/Azure) behind feature flags. Building on the CAMS family, our approach preserves traceability between models, prompts, and generated artifacts, and introduces guardrails for security and observability. We illustrate CAMS+GenAI with a low-connectivity logistics case that executes rules on-device and reconciles events upon reconnection. We outline an evaluation plan comparing a CAMS-F baseline to CAMS+GenAI in terms of modeling effort, rule coverage, defect density, latency on edge, and provider switch MTTR. We also provide a replication package with scripts, synthetic datasets, and a runnable prototype. Contributions include: (1) a generative DSL and G-MDD pipeline for context-aware apps; (2) a verifiable, offline-first edge–cloud runtime with multicloud connectors; and (3) a replicable case study and evaluation plan.
Event planning involves complex constraint satisfaction across budgets, guest capacity, vendor availability, and user preferences. This study proposes a conversational multi-agent architecture that integrates large language models with specialized reasoning agents. Instead of relying on a monolithic model, the system decomposes planning into cooperative agents responsible for intent detection, slot extraction, budget validation, retrieval, ranking, and natural language generation. A centralized orchestration policy selects agents based on conversation state. A hybrid retrieval-augmented generation pipeline integrates semantic embeddings, lexical similarity, and machine-learned ranking to guarantee constraint-aware recommendations as part of an intelligent conversational system. Experimental results demonstrate slot extraction accuracy exceeding 91
