Abstract: The rapid adoption of generative AI in various sectors, particularly in finance, has introduced new challenges and opportunities for model risk management (MRM). This paper provides a comprehensive review of the current state of MRM in the context of generative AI, focusing on the risks, regulatory frameworks, and mitigation strategies. We explore the implications of generative AI on financial institutions, the evolving regulatory landscape, and the role of advanced MRM frameworks in ensuring compliance and mitigating risks. By synthesizing insights from 50+ recent articles, this paper aims to provide a roadmap for future research and practical applications of MRM in the generative AI era. It examines the key risks associated with these models, including bias, lack of transparency, and potential for misuse, and explores the regulatory frameworks and best practices being developed to mitigate these risks. We delve into the specific challenges faced by financial institutions in adapting their MRM strategies to encompass generative AI, and highlight the emerging tools and technologies that can support effective risk management. This paper also discusses **quantitative methods** for risk quantification, such as probabilistic frameworks, Monte Carlo simulations, and adversarial risk metrics, which are essential for assessing the reliability and robustness of generative AI models. Foundational metrics, including fairness measures like demographic parity and equalized odds, are explored to address bias and ensure ethical AI deployment. Additionally, the paper presents **pseudocode** for key algorithms, such as risk quantification and adversarial risk calculation, to provide a practical understanding of these methods. A detailed **gap analysis** identifies critical shortcomings in current MRM frameworks, such as the lack of standardized validation methods and inadequate handling of adversarial robustness. Based on these gaps, the paper proposes solutions, including the development of advanced validation frameworks, integration of fairness metrics, and alignment with regulatory standards. These findings and proposals aim to guide financial institutions in adopting generative AI responsibly while addressing the unique risks it poses. This paper serves as a valuable resource for professionals and researchers seeking to understand and navigate the complexities of MRM in the age of generative AI.

Abstract: We assume that the probability mass function Pr(Z)=(2Z)^-2 is at Newcomb-Benford Law's root and the origin of positional notation. Under its tail, we find that the harmonic (global) Q-NBL for bijective numeration is Pr(b,q)=(q Hb)^-1, where q is a quantum (1≤q≤b), Hn is the nth harmonic number, and b is the bijective base. Under its tail, the logarithmic (local) R-NBL for bijective numeration is Pr(r,d)=Log(r+1,1+1/d), where d≤r ≪ b, being d a digit of a local complex system’s bijective radix r. We generalize both lows to calculate the probability mass of the leading quantum/digit of a chain/numeral of a given length and the probability mass of a quantum/digit at a given position, verifying that the global and local NBL are length- and position-invariant in addition to scale-invariant. In the framework of bijective numeration, we also prove that the sums of Kempner’s series conform to the global Newcomb-Benford Law and suggest a natural resolution for the precision of a universal positional notation system

Abstract: We propose a novel framework that integrates a meta-learning adaptation phase with feature-level knowledge distillation to accelerate convergence and improve the generalization of lightweight neural networks. Our framework aligns intermediate features via a hybrid loss combining mean-squared error and cosine similarity, and refines the student model using a MAML-based meta-learning adaptation phase. Our theoretical analysis, under simplifying assumptions, demonstrates that this dual mechanism reduces generalization error by effectively lowering model complexity and enforcing an information bottleneck. Experimental results on CIFAR-10 confirm that teacher guidance accelerates early training, and our analysis suggests that dynamic scheduling of the distillation weight may further enhance performance.

Abstract: The rapid development of AI, particularly in the form of language models such as OpenAI’s ChatGPT, is fundamentally transforming the academic and educational landscape. AI tools like ChatGPT play a crucial role in accelerating text generation, enhancing writing processes, assisting in literature reviews, and facilitating personalized learning. The development and adoption of Generative Pre-trained Transformers (GPT), including ChatGPT, have sparked widespread interest due to their unprecedented capabilities in generating human-like text and automating communication. These advancements have a significant impact on various fields, including education and research, where AI-driven tools can optimize academic work and improve accessibility to knowledge. However, the integration of AI into academia comes with several serious challenges, such as algorithmic bias, data privacy concerns, ethical issues including plagiarism, and the potential decline in critical thinking skills due to excessive reliance on AI-generated content. While some educators and researchers view ChatGPT as a powerful tool for innovation, others express concerns about its impact on academic integrity and analytical reasoning. This review also compares ChatGPT with other advanced AI models, such as GPT-3 and BERT, highlighting their respective roles in science and education. An analysis of 159 independent literature sources, including 36 recently published papers on arXiv.org platform, indicates that interest in ChatGPT and its analogs in the field of science and education is rapidly growing. These studies emphasize the importance of AI in optimizing academic activities, improving knowledge accessibility, and developing new learning methods. The article concludes by discussing the future evolution of AI in academia, its integration with emerging technologies, and its potential for fostering interdisciplinary research in an era often described as the "new AI gold rush."

Abstract:

With the rise of ecommerce systems and web application usage, recommendation systems have become important to our daily tasks. They provide personalized suggestions to assist with any task under consideration. While various machine learning algorithms have been developed for recommendation tasks, existing systems still face limitations. This research focuses on advancing contextaware recommendation systems by leveraging the capabilities of Large Language Models (LLMs) in conjunction with realtime data. The research exploits the integration of existing realtime data APIs with LLMs to enhance the capabilities of the recommendation systems already integrated into smart societies. The experimental results demonstrate that the hybrid approach significantly improves the user experience and recommendation quality, ensuring more relevant and dynamic suggestions.

Abstract: Artificial intelligence (AI) models are widely adopted in various industries, yet their decision-making processes often exhibit biases that reflect societal inequalities. This review investigates how biases emerge in AI systems, the consequences of biased decision-making, and strategies to mitigate these effects. The paper follows a systematic review methodology, utilizing PRISMA guidelines to analyze existing literature. Key themes include data-driven biases, algorithmic influences, and ethical considerations in AI deployment. The review concludes with future research directions, emphasizing the need for fairness-aware AI models, robust governance, and interdisciplinary approaches to bias mitigation.

Abstract: Background/Objectives: In order to prevent the transmission rate of COVID-19, early diagnosis with high accuracy is essential. The aim of this study is to identify the disease using convolutional neural network (CNN) architectures that allow accurate and rapid diagnosis of COVID-19 pneumonia on computed tomography (CT) images and to evaluate the classification success of the architectures by comparing their classification success with various performance metrics. Methods: In the study dataset, a total of 15584 lung CT slices were obtained, 8395 slices from 361 positive cases and 7189 slices from 134 negative cases, selected in accordance with the study criteria, with RT-PCR test results and CT scans. The dataset were analysed using fine-tuned DenseNet169, MobileNetV2, ResNet50, InceptionResNetV2, InceptionV3, VGG16, VGG19, Xception, DenseNet121, DenseNet201 and ResNet101. Accuracy, sensitivity, specificity, precision, F1 score, ROC curve and AUC were used for performance evaluation. Results: In the SARS-CoV-2 CT scan dataset and the study dataset, the highest performance metrics were obtained from the fine-tuned DenseNet201 with accuracy of 99.19%, sensitivity of 98.65%, specificity of 99.64%, precision of 99.55%, F1 score of 99.10%, AUC of 99.14%, and accuracy of 99.13%, sensitivity of 99.86%, specificity of 98.52%, precision of 98.28%, F1 score of 99.07%, AUC of 99.19%, respectively. The other highest accuracies for the study dataset were 97.47% with fine-tuned DenseNet169 and 97.85% with fine-tuned DenseNet121. Conclusions: As a result, the fine-tuned DenseNet201 proposed in this study shows a promising performance on lung CT images and two different datasets.

Abstract: Jeffreys--Lindley paradox is a case where frequentist and Bayesian hypothesis testing methodologies contradict with each other. This has caused confusion among data analysts for selecting a methodology for their statistical inference tasks. Though the paradox goes back to mid 1930's so far there hasn't been a satisfactory resolution given for it. In this paper we show that it arises mainly due to the simple fact that, in the frequentist approach, the difference between the hypothesized parameter value and the observed estimate of the parameter is assessed in terms of the standard error of the estimate, no matter what the actual numerical difference is and how small the standard error is, whereas in the Bayesian methodology it has no effect due to the definition of the Bayes factor in the context, even though such an assessment is present. In fact, the paradox is an instance of conflict between statistical and practical significance and a result of using a sharp null hypothesis to approximate an acceptable small range of values for the parameter. Occurrence of type-I error that is allowed in frequentist methodology plays important role in the paradox. Therefore, the paradox is not a conflict between two inference methodologies but an instance of not agreeing their conclusions.

Abstract: Training deep neural networks is resource intensive. In this paper, we introduce a novel Rest-Wake training paradigm that alternates between active gradient updates (wake phase) and parameter consolidation (rest phase). Our method is enhanced with dynamic memory compression and an adaptive phase switching mechanism based on gradient variance. We also propose a new stability metric, the neural elasticity coefficient (γ), to measure parameter stability. Our experiments on CIFAR-10 demonstrate that the Baseline model achieves an average validation accuracy of 83.19% with an average GPU peak memory usage of 433.49 MB (standard deviation 52.34 MB). In comparison, the Rest-Wake (Original Architecture) model attains 82.85% accuracy with 465.12 MB (std. dev. 1.11 MB), and the Rest-Wake (Improved Architecture) reaches 83.01% accuracy with 485.96 MB (std. dev. 1.73 MB). A paired t-test yields a p-value of 0.0919, indicating that the performance differences are not statistically significant at the 0.05 level.

Abstract: This work focuses on investigating rough Marcinkiewicz integrals associated to specific surfaces. Whenever the kernel functions belong to Lq(Sm−1) space, the Lp boundedness of these Marcinkiewicz integrals is confirmed. This finding along with Yano’s extrapolation argument prove the Lp boundedness of the aforementioned integrals under weaker conditions on the kernels. The results in this work improve and generalize various previously known results on Marcinkiewicz integrals.

Abstract: The paper investigates the randomly stopped sums. Primary random variables are supposed to be nonnegative, independent, and identically distributed, whereas the stopping moment is supposed to be a~ nonnegative, integer-valued, and nondegenerate at zero random variable, independent of primary random variables. We find the conditions under which dominated variation or extended regularity of randomly stopped sum determines the stopping moment to belong to the class of dominatedly varying distributions. In the case of extended regularity, we derive the asymptotic inequalities for the ratio of tails of the distributions of randomly stopped sums and a stopping moment. The obtained results generalize analogous statements recently obtained for a narrower class of regularly varying distributions. Compared with the previous studies, we apply new methods to the proofs of the main statements whereas methods applied to regularly varying functions are not suitable for the broader class of generalized regularly varying distributions. At the end of the paper, we provide one example that illustrates the theoretical results.

Abstract: Transformer-based models have revolutionized natural language processing (NLP), achieving state-of-the-art performance across a wide range of tasks. However, their high computational cost and memory requirements pose significant challenges for real-world deployment, particularly in resource-constrained environments. Token pruning has emerged as a promising technique to improve efficiency by selectively removing less informative tokens during inference, thereby reducing FLOPs and latency while maintaining competitive performance. This survey provides a comprehensive overview of token pruning methods, categorizing them into static, dynamic, and hybrid approaches. We discuss key pruning strategies, including attention-based pruning, entropy-based pruning, reinforcement learning methods, and differentiable token selection. Furthermore, we examine empirical studies that evaluate the trade-offs between efficiency gains and accuracy retention, highlighting the effectiveness of token pruning in various NLP benchmarks. Beyond theoretical advancements, we explore real-world applications of token pruning, including mobile NLP, large-scale language models, streaming applications, and multimodal AI systems. We also outline open research challenges, such as preserving model generalization, optimizing pruning for hardware acceleration, ensuring fairness, and developing automated, adaptive pruning strategies.As deep learning models continue to scale, token pruning represents a crucial step toward making AI systems more efficient and practical for widespread adoption. We conclude by identifying future research directions that can further enhance the effectiveness and applicability of token pruning techniques in modern AI deployments.

Abstract: This article introduces a new fuzzy double integral transformation called the fuzzy double Yang transformation. We review some of the main properties of the transformation and find the conditions for its existence. We prove the theorems for partial derivatives and fuzzy unitary convolution. All new results are applied to find an exact solution to the fuzzy parabolic Volterra integro-differential equation with a memory kernel. In addition, a numerical example is provided to illustrate the accuracy and superiority of the proposed method with the help of symmetric triangular fuzzy numbers.

Abstract: This paper describes methods for optimal filtering of random signals that involve large matrices. We develop a procedure that allows us to significantly decrease the computational load needed to numerically realize the associated filter and increase the associated accuracy. The procedure is based on the reduction of a large covariance matrix to a collection of smaller matrices. It is done in such a way that the filter equation with large matrices is equivalently represented by a set of equations with smaller matrices. The filter Fp we develop is represented by Fp(v1,…vp)=∑j=1pMjvj and minimizes the associated error over all matrices M1,…,Mp. As a result, the proposed optimal filter has two degrees of freedom to increase the associated accuracy. They are associated, first, with the optimal determination of matrices M1,…,Mp and second, with the increase in the number p of components in the filter Fp. The error analysis and results of numerical simulations are provided.

Abstract: The concept of Artificial General Intelligence (AGI)—a machine capable of performing any intellectual task a human can—is both a central aspiration and a contested notion in AI research. Despite its prominence in scholarly and public discourse alike, AGI relies on unsettled definitions of intelligence and speculative assumptions about generalization. This paper critically examines AGI from multiple perspectives: conceptual theory, philosophy, psychometrics, and recent developments in large language models (LLMs). The foundations of AGI are undermined by the lack of consensus on what constitutes "intelligence'' in both human and artificial contexts. Furthermore, I explore how AGI systems may excel at benchmarks by optimizing for performance rather than demonstrating genuine understanding—akin to the "simulation without comprehension'' phenomenon described by Searle’s Chinese Room argument. I also investigate the emergent behaviors reported in advanced AI models, assess whether these indicate genuine steps toward general intelligence or illusory artifacts, and discuss how introspective features in LLMs might or might not constitute a move toward ``self-awareness.'' By integrating insights from multiple disciplines, I propose a framework for reevaluating AGI that prioritizes scientific rigor, conceptual clarity, and ethical considerations. This analysis underscores the urgent need to distinguish mere test-passing behavior from true intelligence and to develop robust, psychometrically grounded benchmarks for AGI evaluation.

Abstract: Since the discovery of the Differential Evolution algorithm, new and improved versions have continuously emerged. In this paper, we review selected algorithms based on Differential Evolution proposed in recent years. We examine the mechanisms integrated into them and compare the performances of algorithms. To compare their performances statistical comparisons were used as they enable us to draw reliable conclusions about algorithms performances. We use the Wilcoxon signed-rank test for pairwise comparisons and the Friedman test for multiple comparisons. Subsequently, the Mann-Whitney U test was added. We conducted not only a cumulative analysis of algorithms but we also focused on their performances regarding the function family (i.e., unimodal, multimodal, hybrid, and composition functions). Experimental results of algorithms were obtained on problems defined for the CEC’24 Special Session and Competition on Single Objective Real Parameter Numerical Optimization. Problem dimensions of 10, 30, 50, and 100 were analyzed. In this paper we highlight promising mechanisms for further development and improvements, based on the performed study of the selected algorithms.

Abstract: A continuous increase of Artificial Intelligence (AI) based functions can be expected for future aviation systems, posing significant challenges to traditional development processes. Established systems engineering frameworks, such as the V-model, are not adequately addressing the novel challenges associated with AI-based systems. Consequently, the European Union Aviation Safety Agency (EASA) introduced the W-shaped process as an advancement of the V-model to set a regulatory framework for the novel challenges of AI Engineering. In contrast, the agile Development Operations (DevOps) approach, widely adopted in software development, promotes a never-ending iterative development process. This article proposes a novel concept that integrates aspects of DevOps into the W-shaped process to create an AI Engineering framework suitable for aviation-specific applications. Furthermore, it builds upon proven ideas and methods using AI Engineering efforts from other domains. The proposed extension of the W-shaped process, compatible with ongoing standardizations from the G34/WG-114 Standardization Working Group, a joint effort between EUROCAE and SAE, addresses the need for a rigorous development process for AI-based systems while acknowledging its limitations and potential for future advancements. The proposed framework allows for a re-evaluation of the AI/ML constituent based on information from operations, enabling improvement of the system’s capabilities in each iteration.

Abstract: In the modern financial sector, the exponential growth of data has made efficient and accurate financial data analysis increasingly crucial. Traditional methods, such as statistical analysis and rule-based systems, often struggle to process and derive meaningful insights from complex financial information effectively. These conventional approaches face inherent limitations in handling unstructured data, capturing intricate market patterns, and adapting to rapidly evolving financial contexts, resulting in reduced accuracy and delayed decision-making processes. To address these challenges, this paper presents an intelligent financial data analysis system that integrates Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG) technology. Our system incorporates three key components: a specialized preprocessing module for financial data standardization, an efficient vector-based storage and retrieval system, and a RAG-enhanced query processing module. Using the NASDAQ financial fundamentals dataset from 2010 to 2023, we conducted comprehensive experiments to evaluate system performance. Results demonstrate significant improvements across multiple metrics: the fully optimized configuration (gpt-3.5-turbo-1106+RAG) achieved 78.6% accuracy and 89.2% recall, surpassing the baseline model by 23 percentage points in accuracy while reducing response time by 34.8%. The system also showed enhanced efficiency in handling complex financial queries, though with a moderate increase in memory utilization. Our findings validate the effectiveness of integrating RAG technology with LLMs for financial analysis tasks and provide valuable insights for future developments in intelligent financial data processing systems.

Abstract: As AI technologies advance at an exponential rate, traditional governance frameworks and ethical guidelines are struggling to keep pace. The existing discourse largely focuses on isolated issues such as algorithmic fairness, privacy, and accountability, often overlooking the complex and interconnected nature of AI systems. This article proposes a comprehensive governance framework that integrates complex systems theory, post-capitalist governance models, and global justice perspectives to provide an innovative approach to AI governance. By addressing critical gaps in current AI governance scholarship, this framework offers pioneering insights into how AI can be ethically regulated and governed in the 21st century.

Abstract: In the field of natural language processing, Biaffine model is a typical neural network structure based on double affine transformation, which helps to understand sentence structure and the relationships between words, which can be used for tasks such as text classification and relation extraction. However, this model faces challenges in entity pair relation classification tasks, such as imbalanced relation types and unclear entity pair feature information. Therefore, this paper proposes an improved relation classification model named Bert-CL-Biaffine, which is based on bidirectional entity and contrastive learning, combining a global pointer network with contrastive learning to enhance the Biaffine model for relation classification tasks. By training the model to identify the start and end positions of entities in sentences, it performs better in classifying overlapping entity pairs in complex scenarios. Experimental results show that on the NYT and WebNLG datasets, the F1 score of Bert-CL-Biaffine model improves by 1% and 1.2%, respectively, compared to baseline models, indicating that the improved relation classification model effectively enhances performance in complex scenarios.