The pursuit of Artificial General Intelligence (AGI) requires learning frameworks that not only optimize task performance but also align with complex human values. Reinforcement Learning with Human Feedback (RLHF) has emerged as a promising approach to address this challenge; however, conventional RLHF pipelines face scalability issues, reward-model brittleness, and safety concerns. In this study, we propose a PPO-based RLHF framework enhanced with hybrid human–AI oversight and predictive reward evaluation metrics. The framework integrates human annotations with AI-generated critiques, improving data efficiency and robustness against reward hacking. Experimental evaluations on language alignment and control benchmarks demonstrate that the proposed approach achieves a preference win-rate of 78% (vs. 65% in standard RLHF and 54% in supervised fine-tuning), improves task accuracy to 83% (a 12% increase over Sparrow), and reduces safety violations by 31% compared to baseline RLHF. Furthermore, the hybrid oversight strategy enhanced sample efficiency by 1.5×, reducing overall training episodes and annotation costs. These results confirm that the proposed method significantly improves alignment, efficiency, and safety, positioning RLHF with hybrid oversight and predictive evaluation as a practical substrate for advancing safe and scalable AGI systems.

Earthquake damage prediction is essential to improve preparedness and resilience in seismically active regions such as Nepal. This study investigates the application of machine learning models to predict the grade of building damage during seismic events, using building attributes from the 2015 Gorkha earthquake dataset. We employed various algorithms, including Random Forest with both Random Oversampling and SMOTE, Multilayer Perceptron (MLP), Support Vector Machine (SVM), Gradient Boosting, and XGBoost, to tackle class imbalance in damage grades and enhance predictive accuracy. Key building attributes such as construction materials, number of floors, and building age were analyzed to forecast the severity of the damage, which was categorized into three grades. Our findings indicate that the Random Forest model, particularly with SMOTE oversampling, outperformed other techniques, achieving a macro F1-score of 0.67. Both Random Oversampling and SMOTE significantly improved performance over the imbalanced baseline, with Random Oversampling achieving a macro F1-score of 0.65. However, all models faced challenges in accurately predicting Grade 2 damage, underscoring the need for further refinement to address intermediate damage levels. The study highlights the strengths and limitations of various machine learning approaches and emphasizes the significance of feature selection and data balancing methods in improving earthquake damage prediction. It offers valuable insights for policymakers and engineers aiming to enhance earthquake preparedness and develop resilient infrastructure.

This paper examines current trends in Artificial Intelligence (AI), with a particular emphasis on the application of the Python programming language in educational settings. The existing literature reveals a gap in reviews regarding Python's role in AI applications within the education sector. This paper highlights and describes the most effective and popular Python libraries, particularly those relevant to AI in educational settings, that have emerged in recent years. Additionally, it presents basic and widely used AI applications in education that utilize Python. The research method employed is a state-of-the-art review, which involves searching the literature using specific keywords and applying inclusion and exclusion criteria for quality and reliability. In the findings, synthesis, and discussion section, results are presented in tables to facilitate use by researchers and practitioners, allowing for easier application. The tables include information on the most well-known and frequently used Python libraries and AI applications, the grouping categories of Python libraries, their tasks and functions, as well as how Python libraries can be applied in education. Moreover, the paper explores AI methods that enhance teaching and learning using Python tools and connects learning theories with these AI methods. During the discussion, the paper provides an analysis to assist new researchers and practitioners, especially those working with AI in education, on how to effectively leverage AI through Python. It also emphasizes the overall benefits that Python brings to the field of AI in education. The paper also offers various suggestions for addressing gaps in the existing literature.

Depression remains one of the most serious mental health challenges worldwide and is often underdiagnosed due to social stigma and limited access to medical services. With the proliferation of social media as a medium for self-expression, these platforms provide new opportunities for early detection of depressive symptoms through digital footprints. However, most prior research has primarily focused on linguistic features derived from text, overlooking behavioral dynamics that also reflect psychological states. To address this gap, we propose Depress-HybridNet, a hybrid deep learning framework that integrates linguistic embeddings with behavioral activity patterns. The architecture combines a BERT-BiLSTM encoder for linguistic feature extraction, a multi-layer perceptron for behavioral feature encoding, and an adaptive attention-based fusion mechanism to integrate multimodal signals optimally. Experiments conducted on the publicly available Kaggle Depression Dataset demonstrate that Depress-HybridNet consistently outperforms strong baselines, including fine-tuned BERT, achieving state-of-the-art results (F1 = 0.92, AUC = 0.93). A further ablation study highlights the critical role of behavioral features and the attention fusion layer in improving performance. In addition, clinical validation by licensed psychologists confirmed a high degree of consistency between the model’s predictions and expert judgment, underscoring its real-world applicability. These findings underscore the importance of modelling depression as a multifaceted phenomenon, rather than a purely linguistic task, and establish a reproducible benchmark for future research.

Accurate classification of dry bean varieties is essential for enhancing agricultural productivity, ensuring seed quality, and supporting market standardization. Traditional classification methods are time-consuming, labor-intensive, and prone to error, particularly when dealing with morphologically similar varieties. While recent data-driven approaches have shown promising performances, they often suffer from limited varietal representation, inadequate handling of high-resolution image data, and suboptimal dimensionality reduction techniques. To address these challenges, this study proposes a novel hybrid classification pipeline that integrates a fine-tuned InceptionV3 for low- and high-level feature extraction, Principal Component Analysis (PCA) for efficient dimensionality reduction, and a Support Vector Machine (SVM) for final classification. The model was evaluated on a comprehensive dataset combining 14 dry bean varieties from Dogan et al. with five locally sourced varieties, totaling 19 distinct classes. Extensive experimental results showed that the proposed model achieved an accuracy of 94.00% and 80.01% on benchmark and combined datasets, respectively, outperforming state-of-the-art approaches.

Face recognition (FR) is a widely used biometric technology. Nevertheless, achieving efficient and robust FR is still challenging due to variations in illumination, pose, and facial expression. A vital step in any FR system is to select the most informative features and eliminate the redundant ones. In this study, a hybrid approach combining Principal Component Analysis (PCA) and the Dolphin Swarm Algorithm (DSA) with Euclidean Distance as a lightweight classifier is proposed. Experiments were made by using the ORL dataset, which consists of 400 grayscale images. With a 98% recognition rate for this hybrid approach against a recognition rate of 90–92% that can be achieved by PCA only, the proposed PCA+DSA outperformed standalone PCA while still being computationally economical. The metrics of Recognition Rate, Receiver Operating Characteristic (ROC), Cumulative Match Curve (CMC), and Expected Performance Curve (EPC) provided numerous confirmations for this Hybrid model. Additionally, the convergence analysis corroborated DSA’s efficacy in feature selection as the fitness was nearly 92% after nine iterations. Without requiring sophisticated classifiers or deep learning models, our findings show that the identification rate can be improved by combining a bio-inspired optimization technique and the classical PCA method. However, the current study is limited to the ORL dataset in a controlled environment. Future research will focus on implementing and evaluating the system in real-time scenarios on larger and more diverse datasets to enhance its scalability and robustness in practical applications.

Lithium-ion battery demand continues to rise as consumer electronics, hybrid and electric vehicles, and other technologies evolve. This means that numerous lithium-ion batteries are likely to be disposed of, leading to serious disposal problems and negative impacts on the environment and energy conservation. The commonly used lithium-ion battery recycling methods, which are chemical and mechanical, pose challenges, such as some batteries exploding, thermal runaway, or fire. Classification of used lithium-ion battery waste is required to be efficient and reliable. The purpose of this study was to develop a deep learning-based classification model of lithium-ion battery components for automated recycling. A dataset containing images of end-of-life lithium-ion battery components was collected from selected recycling centres in Nairobi, Kenya, and Kaggle.com. The images from selected fields were annotated using the Labelimg annotation tool. The dataset was split into three sets: 70% as the training set, 15% as the validation set, and 15% testing set. A Yolov8n model was then trained using the training set to detect and classify end-of-life lithium-ion battery components. The performance of the model was evaluated using the validation set and test set. The final trained model attained 0.903 precision, 0.792 recall, 0.852 mAP@0.5, 0.724 mAP@0.5-0.95, and 0.844 f1-score. The results from this research study could pave the way for innovative battery recycling measures, guaranteeing that valuable resources are reclaimed and toxic battery waste is managed efficiently and safely.

Brain tumor detection using deep learning has become a focus of research due to its potential to improve patient diagnosis and management. This study presents a comparative analysis to evaluate how the fundamental design philosophies of convolutional neural network (CNN) architectures, namely sequential (VGG16), residual (ResNet50), and encoder-decoder (U-Net), influence performance on brain tumor classification tasks using a public MRI dataset. The data undergoes preprocessing, including normalization, data augmentation, and division into training and testing subsets. Evaluation is conducted using accuracy, precision, recall, and F1-Score metrics. The results show that ResNet50 consistently outperforms other architectures with an average accuracy of 95.35%, followed by VGG16 (93.93%). Conversely, U-Net, designed for segmentation, demonstrates the lowest performance (79%). The superiority of ResNet50 highlights the benefits of residual connections in addressing the vanishing gradient problem for classification tasks. The poor performance of U-Net confirms the hypothesis that architectural suitability for a specific task is more critical than complexity alone. This study highlights the strength of ResNet50 as a reliable approach for brain tumor classification. It underscores the importance of selecting an architecture specifically designed for the intended medical image analysis task. In addition, an analysis of model complexity and computational efficiency was conducted, showing that while ResNet50 achieved the highest accuracy, VGG16 provided a more favorable trade-off between performance and training time.

Credit card fraud detection (CCFD) remains a critical research domain due to the dynamic, adversarial, and highly imbalanced nature of fraudulent activities in financial systems. This study employs a systematic mapping review guided by the PRISMA 2020 guidelines. It analytically synthesizes 40 peer-reviewed and open-access studies, focusing on methodological trends, machine learning techniques, datasets, optimization strategies, and evaluation metrics. Supervised learning (SL) models, including Random Forest, Decision Trees, Support Vector Machine (SVM), and XGBoost, accounted for nearly half of the reviewed studies and consistently demonstrated strong performance. Deep learning (DL) frameworks, including Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and their variants, have demonstrated strong capabilities in capturing sequential and high-dimensional patterns of fraud. However, their effectiveness is constrained by class imbalance and dataset bias. Ensemble and hybrid models further enhanced predictive accuracy but introduced higher computational costs and lower interpretability. A key finding is the heavy reliance on the ECCT 2013 dataset (used in more than half of the reviewed studies), which supports reproducibility but limits generalizability to modern fraud contexts. Optimization strategies, such as the Synthetic Minority Oversampling Technique (SMOTE), hyperparameter tuning, and dimensionality reduction, have proven effective in improving recall and reducing false negatives; however, they have been inconsistently applied. Similarly, evaluation metrics were uneven, with accuracy dominating (reported in 75% of studies), while more informative measures such as recall, F1-score, Precision-Recall curves (AUPRC), and Matthews Correlation Coefficient (MCC) received less emphasis despite their relevance to imbalanced data. Overall, while many models achieve high accuracy in controlled environments, their scalability, adaptability, and trustworthiness in real-world deployment remain limited. Future research should prioritize cross-dataset evaluations, standardized metrics, and emerging paradigms such as federated learning, self-supervised approaches, and explainable AI to guide the development of robust and deployable fraud detection systems.

To meet the development requirements for the complex algorithms with a gradually clear mathematical model and technical routes, frequent modifications and updates for the workflows, and long exploratory cycles, this article conducted preliminary research on a new programming idea based on a visual logic structure - PING programming language. At this time, PING language is a proof-of-concept and is compatible with the C/C++ language. PING language took function modules as the basic programming units. It utilized various explicit execution orders and invoked relationships between function modules to construct the static software logic structure network, implemented a visual scheduling and execution mechanism for function modules, and employed variable symbol table technology to achieve shared variable value transfer. The PING programming language could explicitly represent the overall static logic structure of software during the software programming stage. During the software running stage, the function pointer was used to centrally invoke the function modules, which facilitated the dynamic tracking of the function modules' execution workflows and automatically detected redundant logic branches and untested paths within the software. It had great significance in improving the development, debugging, testing, and maintenance efficiency of interpretable intelligent software, and even in developing machine autonomous programming systems.