Recent advances in large language models (LLMs) have enabled impressive performance across reasoning tasks, yet their capacity for autonomous moral judgment remains underexplored. Prior research has not systematically benchmarked LLMs against expert standards of postconventional moral reasoning. This study aims to evaluate the moral reasoning capabilities of twelve state-of-the-art LLMs by comparing their responses to complex socio-moral dilemmas with those of a moral psychology expert, using Kohlberg’s framework as a referential standard. Models were prompted with four dilemmas (passive euthanasia, incurable diseases decisions, population control, and capital punishment) across three zero-shot tasks: moral recognition, option selection, and justification. Responses were coded using Kohlberg’s levels of moral development and statistically compared to expert judgments. Quantitative metrics included weighted kappa for concordance and logistic regression for explanatory depth. Qualitative analysis examined the structure and coherence of LLM justifications. LLMs predominantly exhibited conventional reasoning (47.7%), diverging sharply from the expert’s consistently postconventional judgment. Concordance was weak to negative, with kappa coefficients ranging from –0.333 to 0.286, indicating systemic misalignment. Binary logistic regression analysis showed that model classification significantly predicted the capacity for moral articulation; the odds ratio was 7.00 (p = 0.005), accounting for 23.4% of the variance according to Nagelkerke’s R² = 0.234. The Kruskal–Wallis test yielded H = 20.34 with p < .05, confirming significant differences in moral reasoning levels between the models and the expert. Finally, a qualitative analysis revealed that although 84.3% of justifications aligned with the selected options, they often lacked principled depth, replicating normative patterns and rejecting postconventional stances through lower-order reasoning. In conclusion, contemporary LLMs reproduce social conventions rather than engaging in autonomous ethical deliberation. The observed limitations have substantive implications for the ethical design and deployment of AI systems in morally sensitive domains. To address this, the study proposes six complementary strategies for ethical alignment, aiming to foster principled moral reasoning in AI systems: staged fine-tuning, symbolic ontology integration, expert supervision, reinforcement learning with philosophical rewards, adversarial training, and continual calibration.
Flood disasters pose severe threats to human lives and infrastructure worldwide. Accurate and timely flood detection using satellite imagery is essential for emergency response and disaster management. This paper introduces MUST-Former, a family of five deep learning models designed for robust global flood mapping using multi-modal satellite data. The proposed architectures leverage Sentinel-1 radar and Sentinel-2 optical imagery from the Sen1Floods11 dataset, which covers 11 flood-affected regions across 6 continents. Three distinct fusion strategies are presented: Projection-based fusion, Cross-Attention fusion, and the flagship Principal Component Analysis-inspired fusion. The models address critical challenges in flood detection, including severe class imbalance, geographic bias, and sensor-specific limitations. The MUST-Former-PCA variant achieves a mean Intersection over Union of 90.35% on the test set and 89.32% on the unseen Bolivia region, demonstrating superior generalization. Compared to existing baselines, MUST-Former-SAR shows a 160% improvement in mean IoU over the original Sen1Floods11 baseline, while MUST-Former-Optical achieves competitive performance against foundation models using 3.6 times fewer parameters and superior geographic generalization on the unseen Bolivia split. The fusion models successfully mitigate single-modality failures across diverse geographic regions. Statistical reproducibility is confirmed across three independent random seeds. These results prove the effectiveness of multi-modal approaches for operational flood mapping systems.
Classification of medical images for precise detection and treatment of diseases is very vital. This study focuses mainly on classifying MRI images of brain tumors for neurological disorders. This study introduces a Systematic deep learning framework based on an enhanced ResNet50 architecture to improve brain tumor classification. The proposed approach incorporates dropout regularization, Global Average PoolingGAP), additional dense layers, and leaky ReLU activation to boost feature extraction and reduce overfitting. This work uses a varied set of MRI images. The dataset has gone through a different enhancement and preprocessing pipeline that includes cropping, resizing, contrast enhancement, and smoothing. The model evaluation process incorporated stratified 5-fold cross-validation. Different performance Indicators were measured through accuracy, precision, recall, and F1-score. The proposed architecture attained an average testing accuracy of 99.62% (95% Confidence Interval (CI)). These measures surpass benchmark models such as VGG16, MobileNet, InceptionV3, and Xception, in addition to previously documented methodologies in the existing literature. The result shows the stability across the folds and was statistically corroborated by ablation. Its ablation studies revealed that although the architectural alterations yielded only marginal enhancements over the baseline.
The dynamic and ever-evolving cyber threat landscape necessitates proactive and sophisticated defensive strategies. Cyber threat hunting, a critical practice, moves beyond reactive security measures. It actively searches for undetected malicious activities. Traditionally reliant on human expertise and structured data analysis, threat hunting is increasingly augmented and transformed by the capabilities of artificial intelligence (AI) and machine learning (ML). Among the latest advancements, large language models (LLMs) have emerged as powerful tools with significant potential to revolutionize how cybersecurity professionals detect, analyze, and respond to threats. This paper presents a comprehensive and systematic review of the state-of-the-art LLM-based approaches for cyber threat hunting, focusing significantly on research published in 2025 and grounding these advances in foundational work published before 2023. We explore the diverse applications of LLMs throughout the threat-hunting lifecycle, including advanced cyber threat intelligence (CTI) analysis, real-time anomaly detection, incident response enhancement, and vulnerability prediction. Furthermore, we examine the foundational AI and LLM concepts driving these innovations. We delve into emergent datasets and evaluation methodologies crucial for progress and critically assess the substantial challenges, limitations, and ethical considerations such as robustness, explainability, data privacy, and adversarial vulnerabilities that are inherent in their deployment. Our review provides a novel synthesis of recent Large Language Model (LLM) applications in threat hunting and identifies critical research gaps. Additionally, it presents a refined perspective on the practical implementation and future trajectory of these technologies.
This Systematic Literature Review examines prompt engineering in automatic code generation using large language models (LLMs). A methodological protocol identified 26 relevant primary studies to characterize the status, trends, challenges, and opportunities of prompt engineering in software development. The results show that prompt engineering has been established as a key discipline for optimizing interaction with LLMs and improve the accuracy, robustness, and applicability of the generated code. The findings were grouped into six recurring thematic categories: structured methodologies, pedagogical strategies, accuracy and robustness, code improvement techniques, security through prompts, and the relevance of prompt engineering in the interaction with LLMs. It also highlights that the accelerated growth of publications between 2021 and 2025, as well as the sustained academic interest, reflect the strategic value of prompt engineering in the software lifecycle. This review provides a critical knowledge base for researchers and developers seeking to integrate prompt engineering into their processes and practices effectively.
Rare diseases are medical conditions affecting a small percentage of the population, often leading to diagnostic challenges due to limited clinical data and symptom overlap with common diseases. Due to the lack of knowledge about these disorders and the fact that many of their symptoms are similar to those of common disorders, doctors find it difficult to diagnose these conditions. This paper leverages DeepSeek Plus Knowledge Graph for high-precision symptom extraction in rare diseases, integrating natural language processing (NLP) and graph-based reasoning to enhance disease recognition. This study presents a novel model called “Rarepheno” that extracts symptoms related to rare diseases from clinical text using DeepSeek Large Language Model (LLM) and knowledge graphs (KG). The proposed model integrates DeepSeek AI, N-GRAM based term extraction, and the Human Phenotype Ontology (HPO) knowledge graph to enhance symptom identification. To improve contextual precision and reduce false positives, the system leverages semantic similarity through BioClinicalBERT embeddings. The proposed model outperforms the state-of-the-art rare disease phenotyping algorithms, with a precision score of 0.83, compared to PhenoGPT (GPT-J) (0.809), PhenoTagger (0.720), and ClinPhen (0.590). Compared to rule-based systems like MetaMap (0.707) and NCBO (0.777), Rarepheno demonstrated superior accuracy in extracting standardized symptom terms. This performance gain reflects its ability to reduce false positives and improve symptom-disease association, particularly in heterogeneous and unstructured clinical texts. This study demonstrates how the new-age technologies, such as GenAI and Natura Language Processing (NLP) can aid in collecting valuable clinical data, improving patient safety and rare disease identification.
In India's mountainous areas, maternal mortality is still a serious public health concern, especially in Uttarakhand, where access to healthcare is hampered by geographical obstacles. UttaraRisk-Next, a multi-task ensemble learning framework for thorough maternal health risk assessment, is presented in this paper. Three crucial outcomes are simultaneously predicted by the model: the probability of abortion, the continuous risk percentage (0–100%), and the risk of maternal mortality. We created 78 clinical features in accordance with WHO guidelines using a synthetic but epidemiologically representative dataset of 2,500 pregnancies from 13 districts in Uttarakhand. These features included blood pressure classifications, hemoglobin categories, and socioeconomic vulnerability indicators. UttaraRisk-Next employs an ensemble architecture combining gradient boosting and random forest models with isotonic calibration for probability refinement. On validation data (n=500), the model achieved: risk prediction MAE 5.557% with R^2=0.708 and 97.6% interval coverage; abortion classification ROC-AUC 0.558 with excellent calibration (ECE=0.020); mortality prediction ECE=0.001 despite rare event frequency (0.6%). Comprehensive fairness analysis across rural-urban, age, and socioeconomic dimensions demonstrated equitable performance (ECE differences <0.025). The model identifies 22.4% of pregnancies as high-risk, enabling targeted resource allocation. With 2.1ms inference time and 45MB memory footprint, UttaraRisk-Next is deployable in resource-constrained settings, directly supporting SDG-3.1 (maternal mortality reduction) and SDG-5 (gender equality) objectives in the Indian Himalayan region
The performance of Large Language Models (LLMs) on multiple-choice university-level exam benchmarks such as MMLU is often reported as highly competitive; however, such results raise persistent concerns regarding contamination of public datasets, English-centric bias, and over-reliance on aggregate accuracy as the primary evaluation signal. In particular, the widespread public availability of evaluation data makes it difficult to disentangle genuine generalization from memorization of seen content, while offering limited insight into models’ abilities on culturally grounded assessments beyond English. To address these issues, we introduce lunes (Leakage-controlled Undergraduate National Exams of Spain), a new benchmark of 11,881 multiple-choice questions drawn from official final-year undergraduate exams in Spanish, covering 104 courses across 22 degree programs. The dataset has been rigorously verified to exhibit minimal public web exposure through a combination of automated web search and manual inspection, which enables evaluation under minimal contamination conditions in a non-English, country-specific academic setting. Our results show that (i) LLMs retain strong performance on general knowledge and factual questions, even in the absence of web-accessible training data, suggesting that contamination alone does not explain their success on public benchmarks; (ii) however, their performance degrades substantially on culturally grounded and country-specific content, particularly in domains such as Spanish law, economy, and social structure. Remarkably, models consistently perform better on Anglo-centric content than on Spain-specific material even when answering in Spanish, suggesting that the bottleneck lies in culturally grounded knowledge rather than in language skills per se. A question-level error analysis further reveals that these failures reflect systematic gaps in local institutional, legal, and geographical knowledge, even for high-resource languages such as Spanish, that aggregate metrics systematically obscure.
We investigate whether Large Language Models (LLMs) can refine the given codebase of an optimization algorithm without requiring specialized user expertise. This is in contrast to works that study optimization algorithm code generation from scratch. To this end, 10 baseline algorithms covering metaheuristics, reinforcement learning, and exact methods are applied to the Traveling Salesman Problem. The results demonstrate that our simple methodology leads to improved algorithm variants in 9 out of the 10 cases analyzed. Notably, the LLMs autonomously incorporated advanced techniques---such as heuristic initializations in exact methods---leading to significant runtime reductions. Furthermore, this performance enhancement did not come at the cost of software quality; the generated code preserved a high maintainability index (averaging 53.40), and for certain models, coincided with simplified structures, observing a reduction in average cyclomatic complexity of up to 19.4%, all without requiring specialized optimization knowledge from the user.
Automatic separation of text from digital images holds significant importance in various domains, including document processing and content-based image retrieval. This paper presents a statistical model-based approach for automatic text component extraction from digital images. The methodology comprises two primary tasks. The first step involves color image segmentation by means of mixture and neural models. This process helps to identify distinct components within the image, where some of the components contain text. In the second step, the task is to separate text components from the non-text components. This task requires a learned model for text features. In this context, we utilize ground truth text components provided by the ``Born-Digital Images'' dataset. From these text components, we extract text-representing features. Later, a D-vine multivariate distribution is fitted to these features, which serves as a model for text features. This trained model is used to discriminate text and non-text components obtained after segmentation. For this purpose, a statistical hypothesis testing method is employed on the log-likelihood statistic. The experimental performance of the D-vine based model is compared to the multivariate Gaussian copula-based model, and the former generally outperforms the latter in terms of recall percentages. Moreover, the segmentation algorithms are evaluated based on recall and precision percentages. The novelty of this research lies in the utilization of D-vine modeling. A D-vine model is capable to capture various feature distributions and associations, significantly enhancing the approximation of the joint distribution of features. This, in turn, boosts the method's ability to discriminate between text and non-text features effectively.
The occurrence of HAB has a direct impact on shellfish farming, leading to economic losses due to the contamination of shellfish with toxins harmful to human health. Predicting these blooms accurately is therefore crucial for minimizing their negative effects on the industry. The DoME machine learning model is particularly notable for its high interpretability, as the trained model is expressed as a mathematical equation, allowing for transparent analysis and a better understanding of the factors driving the predictions. This characteristic distinguishes DoME from other black-box models, making it a valuable tool for stakeholders seeking not only accurate predictions but also insights into the dynamics behind HAB events. In this study, we evaluated the novel DoME (Development of Mathematical Expressions) algorithm for the prediction of Harmful Algal Blooms (HAB) associated with Diarrhoeic Shellfish Poisoning (DSP), a significant concern for the shellfish industry. Our testing involved analysing the model's performance in various environmental conditions, demonstrating its robustness and adaptability. DoME achieved a F1-score of 97.80%, which corresponds to an improvement of around 8% over previous studies. This superior performance, combined with its explainability, underscores the model's potential as a practical and reliable solution for early warning systems in the shellfish industry, helping to protect both public health and economic stability.
This study aims to automate source code summarization by introducing a novel machine learning architecture that integrates multiple feature perspectives. Specifically, it combines lexical, syntactic, and semantic representations of code and employs a transformer-based decoder to generate natural language summaries, benchmarking performance against established baselines. Experiments were conducted on the CanonCode Corpus, a high-quality dataset of 8,542 validated C programs. Individual feature extractors-Convolutional Neural Networks (CNN) for lexical features, Tree-LSTM for syntactic features and Graph Neural Networks (GNN) for semantic features were evaluated and compared with the proposed Hybrid Feature Fusion Network (HFFN). The fused feature vector from HFFN was decoded using a transformer to generate summaries. Performance was measured using ROUGE, BLEU, CodeBLEU, BERTScore, and Exact Match metrics. The HFFN model consistently outperformed all baselines across standard natural language generation metrics, achieving a ROUGE-L score of 0.94 and a BERTScore of 0.93. An ablation study confirmed the complementary contributions of each feature type, with syntactic features providing the greatest individual impact. The improvement over the strongest baseline (CodeBERT) was statistically significant (p < 0.001). The proposed HFFN framework demonstrates the value of combining diverse code representations for summarization. It offers a robust and interpretable architecture that advances multi-view representation learning in software engineering and provides a foundation for future research in automated documentation.
Depression and anxiety disorders affect millions of individuals globally and are commonly addressed through psychological interventions. A growing technological approach to support such treatments involves the use of embodied conversational agents that employ motivational interviewing, a method that promotes behavioral change through empathic engagement. Despite its critical role in therapeutic efficacy, empathy remains a significant challenge for virtual agents to emulate. Emotion Recognition (ER) technologies offer a potential solution by enabling agents to perceive and respond appropriately to users' emotional states. Given the inherently multimo dal nature of human emotion, unimo dal ER approaches often fall short in accurately interpreting affective cues. In this work, we propose a multimo dal emotion recognition model that integrates verbal and non-verbal signals (text and video) using a Cross-Modal Attention fusion strategy. Trained and evaluated on the IEMOCAP dataset, our approach leverages Ekman's taxonomy of basic emotions and demonstrates superior performance over unimo dal baselines across key metrics such as accuracy and F1-score. By prioritizing text as the main modality and dynamically incorporating complementary visual cues, the model proves effective in complex emotion classification tasks. The proposed model is designed for integration into an existing conversational agent aimed at supporting individuals experiencing emotional and psychological distress. Future work will involve embedding the model in the conversational agent platform for emotionally distressed users, aiming to assess its real-world impact on engagement, user experience, and perceived empathy.
This study employs a multilayer network approach to analyze the spatial and temporal patterns of suicide and depression across Mexican municipalities from 2015 to 2020. Using a panel dataset of mental health cases, substance use, and healthcare infrastructure, we constructed a multilayer graph based on cosine similarity. The Infomap clustering algorithm was then applied to identify communities of municipalities with similar mental health profiles. Our results reveal five distinct clusters with significant variations in the levels and temporal dynamics of the analyzed indicators. Notably, two clusters consistently exhibited higher rates of substance use and adverse mental health outcomes. These findings demonstrate the efficacy of network-based methods for identifying at-riskmunicipal groupings, thereby informing targeted public health interventions.
Plant diseases are the main factor in plant mortality and destruction, especially in trees. Early discovery, however, can assist to manage and treat this issue efficiently. To increase output, crop and plant lesions are detected and stopped as soon as feasible. Because it relies solely on visual observation, manual inspection of plant leaf diseases is time-consuming and expensive. The authors offer methods for identifying and categorizing plant leaf diseases using computer vision. Pre-processing original images to visualize contaminated areas, feature extraction from unprocessed or segmented images, feature fusion, feature selection, and classification are a few examples of computer vision approaches. The fusion technique is used to combine the target's numerical data features, which go beyond the picture, with the extracted image features to increase the target's feature representation. The following are the principal issues that researchers found in the literature: Low-contrast infected regions. Extract redundant and irrelevant information, which degrades classification accuracy; Redundant and irrelevant information may lengthen computation times and the targeted models performance will suffer as a result. This study proposed a framework for classifying plant leaf diseases based on the best feature selection and a deep learning fusion model. In the suggested approach, contrast is first enhanced using a pre-processing model, and then the issue of an unbalanced dataset is resolved via data augmentation. The proposed Deep Fusion Learning Model (DFLM) shows an accuracy of 98.8% in comparison with other models.
Anxiety during pregnancy constitutes a relevant factor that can significantly influence labor development. This study presents a novel approach based on explainable artificial intelligence to predict both the type and duration of labor using psychological indicators of anxiety prior to delivery. Employing data from 235 full-term pregnant women from two Spanish hospitals, we developed a multilayer perceptron model to classify eutocic and dystocic deliveries, achieving a capacity to identify 88\% of dystocic deliveries. Additionally, we implemented a regression model that predicts labor time with a mean error of 2 hours, correctly predicting 86% of cases with an error margin of less than 3 hours. The application of explainability techniques to the developed models allows for understanding the specific influence of each anxiety factor on labor development. These results demonstrate the potential of AI models to improve obstetric care and optimize healthcare resource allocation.
Depression and anxiety disorders affect millions of individuals globally and are commonly addressed through psychological interventions. A growing technological approach to support such treatments involves the use of embodied conversational agents that employ motivational interviewing, a method that promotes behavioral change through empathic engagement. Despite its critical role in therapeutic efficacy, empathy remains a significant challenge for virtual agents to emulate. Emotion Recognition (ER) technologies offer a potential solution by enabling agents to perceive and respond appropriately to users' emotional states. Given the inherently multimodal nature of human emotion, unimodal ER approaches often fall short in accurately interpreting affective cues. In this work, we propose a multimodal emotion recognition model that integrates verbal and non-verbal signals (text and video) using a Cross-Modal Attention fusion strategy. Trained and evaluated on the IEMOCAP dataset, our approach leverages Ekman's taxonomy of basic emotions and demonstrates superior performance over unimodal baselines across key metrics such as accuracy and F1-score. By prioritizing text as the main modality and dynamically incorporating complementary visual cues, the model proves effective in complex emotion classification tasks. The proposed model is designed for integration into an existing conversational agent aimed at supporting individuals experiencing emotional and psychological distress. Future work will involve embedding the model in the conversational agent platform for emotionally distressed users, aiming to assess its real-world impact on engagement, user experience, and perceived empathy.
Aspect-based sentiment analysis is a process aimed to understanding the sentiment expressed in opinions or reviews about specific features of an entity. The automatic extraction of aspects is the most challenging task, as it requires the ability to understand the context and to recognize the relevant and characteristic elements of an entity about which you have an opinion. To increase the quality results in the solution to this problem is still a challenge in Spanish reviews, because very few papers have been reported and the reported efficacy rates need to be improved. The use of deep learning models has proven an advantage for aspect extraction task, but the combination of several models for obtaining a final prediction has not yet been exploited. In this work, an aspect extraction method in which several Transformer models are combined through an ensemble learning approach using the Average Voting technique is presented. The proposed solution was evaluated using the SemEval2016 dataset and the results obtained were compared to those reported by other state-of-the-art solutions. The evaluation process not only provides a starting point to have a broader perception of the performance of the Transformers in this context, but also highlights the improvement of the quality results of the aspect extraction with the Transformers-Based Ensemble
In the field of AI, word embedding models have proven to be one of the most effective methods for capturing semantic and syntactic relationships between words, enabling significant advancements in natural language processing. However, producing word embeddings for low-resource indigenous languages-such as Yucatec Maya-often suffers from poor reliability due to limited data availability and unsuitable evaluation benchmarks. In this work, we propose a novel methodology for constructing reliable word embeddings by adapting the Swadesh List for semantic similarity evaluation. Our approach involves translating the Swadesh List from a high-resource pivot language into the target language, applying linguistic and cultural filtering, and correlating similarity scores between pivot-language embeddings from large language models and target-language embeddings. Our results demonstrate that this method produces reliable and interpretable embeddings for Yucatec Maya. Furthermore, our analysis provides compelling evidence that the choice of evaluation benchmark has a far greater impact on reported performance than hyperparameter optimization. This approach establishes a robust new framework with the potential to be adapted for improving word embedding generation in other low-resource languages.
In recent years, the rise in diabetic retinopathy cases has posed significant challenges to existing computer-aided diagnosis (CAD) solutions. These systems often focus on detecting specific abnormalities, such as microaneurysms, exudates, or hemorrhages, rather than providing a comprehensive diagnostic approach. Moreover, state-of-the-art deep learning-based methods face critical limitations, including a lack of contextual understanding, gradient vanishing/explosion issues, and failure to address class imbalance at the instance level, which impacts multi-class classification accuracy. To overcome these challenges, a novel diabetic retinopathy prediction model is proposed, leveraging firefly heuristic segmentation and residual deep spatio-textural feature learning. Instead of processing entire fundus images, the model applies Firefly heuristic-driven Fuzzy C-Means (FFCM) clustering to segment regions of interest (ROIs) corresponding to microaneurysms, exudates, and hemorrhages. Residual deepspatio-textural features are then extracted using Gray-Level Co-occurrence Matrix (GLCM), ResNet50, and AlexNet. These complementary features enhance diversity and heterogeneity, which are further processed using random forest learning. The proposed model achieves outstanding performance, with an average accuracy of 99.77%, precision of 99.88%, recall of 99.64%, F-measure of 99.75%, sensitivity of 99.64%, and specificity of 99.86%, surpassing existing approaches. FFCM mitigates the class imbalance problem, ResNet50 addresses gradient challenges, and AlexNet contributes high-dimensional features, ensuring robust and scalable diagnostics. This innovative solution demonstrates exceptional generalizability and runtime efficiency, offering a cost-effective, comprehensive CAD tool for diabetic retinopathy detection.
