Artificial Intelligence Review最新文献

The tremendous advancements in artificial intelligence (AI) techniques, particularly those pertinent to computer vision and image recognition, are revolutionizing the automotive industry towards the development of intelligent transportation systems for smart cities. Integrating AI techniques into connected autonomous vehicles (CAVs) and unmanned aerial vehicles (UAVs) and their data fusion, enables a new paradigm that allows for unparalleled real-time awareness of the surrounding environment. The potential of emerging wireless technologies can be fully exploited by establishing communication and cooperation among AI-augmented CAVs and UAVs. However, configuring appropriate deep learning (DL) models for connected vehicles is a complex task. Any errors can result in severe consequences, including loss of vehicles, infrastructure, and human lives. These systems are also susceptible to cyber attacks, necessitating a thorough and timely threat analysis and countermeasures to prevent catastrophic events. Our findings highlight the effectiveness of AI-driven data fusion in enhancing cooperative perception between CAVs and UAVs, identify security vulnerabilities in DL-based systems, and demonstrate how V2X-enabled UAVs can significantly improve situational awareness in corner cases.

Infrared–visible image fusion (IVIF) integrates complementary thermal and photometric cues for surveillance, remote sensing, and autonomous perception. Existing surveys, while comprehensive, provide limited guidance for design-to-deployment and seldom relate fusion quality to task outcomes or device constraints. This work provides a unified perspective that organizes IVIF methods along an interface-attention-alignment coordinate system covering classical spatial/transform pipelines and contemporary deep paradigms (generative, discriminative, multi-task, hybrid/Transformer, dynamic). Building on literature through 2025, we synthesize fidelity-robustness-efficiency trade-offs and introduce a comparison-to-deployment protocol that couples fusion metrics with task accuracy (AP/mIoU), latency, memory footprint, and condition-performance characterization (misregistration, noise, illumination/weather). We consolidate Transformer/hybrid coverage with practical recipes and focused guidance on temporal consistency, robustness auditing, and physics-grounded interpretability. Compared with previous reviews, our survey concurrently addresses four under-covered dimensions-video temporal consistency, robustness auditing, task-aware evaluation, and deployment reporting-and distills a practical checklist linking architectural choices to operating conditions and hardware budgets, enabling reproducible, task-relevant IVIF practice.

The exponential growth of Large Language Models (LLMs) continues to highlight the need for efficient strategies to meet ever-expanding computational and data demands. This survey provides a comprehensive analysis of two complementary paradigms: Knowledge Distillation (KD) and Dataset Distillation (DD), both aimed at compressing LLMs while preserving their advanced reasoning capabilities and linguistic diversity. We first examine key methodologies in KD, such as task-specific alignment, rationale-based training, and multi-teacher frameworks, alongside DD techniques that synthesize compact, high-impact datasets through optimization-based gradient matching, latent space regularization, and generative synthesis. Building on these foundations, we explore how integrating KD and DD can produce more effective and scalable compression strategies. Together, these approaches address persistent challenges in model scalability, architectural heterogeneity, and the preservation of emergent LLM abilities. We further highlight applications across domains such as healthcare and education, where distillation enables efficient deployment without sacrificing performance. Despite substantial progress, open challenges remain in preserving emergent reasoning and linguistic diversity, enabling efficient adaptation to continually evolving teacher models and datasets, and establishing comprehensive evaluation protocols. By synthesizing methodological innovations, theoretical foundations, and practical insights, our survey charts a path toward sustainable, resource-efficient LLMs through the tighter integration of KD and DD principles.

Advances in generative artificial intelligence (AI), such as recent developments in text, audio, and video production, have amplified societal concerns, with threat probabilities estimated between 5 and 50%. This manuscript undertakes a comprehensive study to understand the factors influencing AI development, focusing on the interplay between AI research, cinematographic representations, and regulatory policies. The study reveals a strong interaction between scientific advances and cultural representations, indicating shared concerns and themes across both domains. It also highlights broad support for ethical and responsible AI development, with temporal analyses showing the significant influence of films on public perception and slower growth in policy implementation relative to cultural diffusion. The findings discuss the presence of a Pygmalion effect, where cultural representations shape perceptions of AI, and a potential Golem effect, where increased regulation may limit the dangerous development of AI and its societal impact. The study underscores the importance of balanced and ethical AI development, requiring continued monitoring and careful management of the relationship between research, cultural representations, and regulatory frameworks.

The analysis and treatment of brain tumors are among the most difficult medical conditions. Brain tumors must be detected accurately and promptly to improve patient outcomes and plan effective treatments. Recently used advanced technologies such as artificial intelligence (AI) and machine learning (ML) have increased interest in applying AI to detect brain tumors. However, concerns have emerged regarding the reliability and transparency of AI models in medical settings, as their decision-making processes are often opaque and difficult to interpret. This research is unique in its focus on explainability in AI-based brain tumor detection, prioritizing confidence, safety, and clinical adoption over mere accuracy. It gives a thorough overview of XAI methodologies, problems, and uses, linking scientific advances to the needs of real-world healthcare. XAI is a sub-section of artificial intelligence that seeks to solve this problem by offering understandable and straightforward and providing explanations for the choices made by AI representations. Applications such as healthcare, where the interpretability of AI models is essential for guaranteeing patient safety and fostering confidence between medical professionals and AI systems, have seen the introduction of XAI-based procedures. This paper reviews recent advancements in XAI-based brain tumor detection, focusing on methods that provide justifications for AI model predictions. The study highlights the advantages of XAI in improving patient outcomes and supporting medical decision-making. The findings reveal that ResNet 18 performed better, with 94% training accuracy, 96.86% testing accuracy, low loss (0.012), and a rapid time ((sim 6text {s})). ResNet 50 was a little slower ((sim 13text {s})) but stable, with 92.86% test accuracy. DenseNet121 (Adam W) achieved the highest accuracy at 97.71%, but it was not consistent across all optimizers. ViT-GRU also got 97% accuracy with very little loss (0.008), although it took a long time to compute (around 49 s). On the other hand, VGG models (around 94% test accuracy) and MobileNetV2 (loss up to 6.024) were less reliable, even though they trained faster. Additionally, it explores various opportunities, challenges, and clinical applications. Based on these findings, this research offers a comprehensive analysis of XAI-based brain tumor detection and encourages further investigation in specific areas.

Recent advances in Large Language Models (LLMs) have revitalized interest in Agent-Based Models (ABMs) by enabling “generative” simulations, with agents that can plan, reason, and interact through natural language. These developments promise greater realism and expressive power, but also revive long-standing concerns over empirical grounding, calibration, and validation—issues that have historically limited the uptake of ABMs in the social sciences. This paper systematically reviews the emerging literature on generative ABMs to assess how these long-standing challenges are being addressed. We map domains of application, categorize reported validation practices, and assess their alignment with the stated modeling goals. Our review suggests that the use of LLMs may exacerbate rather than alleviate the challenge of validating ABMs, given their black-box structure, cultural biases, and stochastic outputs. While the need for validation is increasingly acknowledged, studies often rely on face-validity or outcome measures that are only loosely tied to underlying mechanisms. Generative ABMs thus occupy an ambiguous methodological space—lacking both the parsimony of formal models and the empirical validity of data-driven approaches—and their contribution to cumulative social-scientific knowledge hinges on resolving this tension.

Over the past few decades, credit scoring has become an important tool in the financial sector. It enables banks and financial institutions to assess the creditworthiness of individuals and reduce the risk of default. As a result of significant advances in artificial intelligence techniques. Machine learning (ML) has made it possible to improve credit scoring by distinguishing between people with good creditworthiness and those with poorer creditworthiness. In this article, we propose a systematic literature review of ML-based financial credit scoring methods published between 2018 and 2024. A total of 330 research papers were extracted from four different online databases and digital libraries. After the study selection procedure, 63 research papers were selected for this systematic review. This paper aims to identify the major ML methods used in credit scoring, assess their strengths and limitations, and highlight notable trends and advancements. In addition, the review addresses the critical challenges faced in the adoption of ML models for credit scoring. This study not only contributes to the understanding of effective ML techniques used for credit scoring but also guides future research by highlighting the promising avenues in ML-based credit scoring efforts.

In recent years, the intensity and frequency of fires have increased significantly, resulting in considerable damage to properties and the environment through wildfires, oil pipeline fires, hazardous gas emissions, and building fires. Effective fire management systems are essential for early detection, rapid response, and mitigation of fire impacts. To address this challenge, unmanned aerial vehicles (UAVs) integrated with advanced state-of-the-art deep learning techniques offer a transformative solution for real-time fire detection, monitoring, and response. As UAVs play an essential role in the detection, classification and segmentation of fire-affected regions, enhancing vision-based fire management through advanced computer vision and deep learning technologies. This comprehensive survey critically examines recent advancements in vision-based fire management systems enabled by autonomous UAVs. It explores how baseline deep learning models, including convolutional neural networks, attention mechanisms, YOLO variants, generative adversarial networks and transformers, enhance UAV capabilities for fire-related tasks. Unlike previous reviews that focus on conventional machine learning and general AI approaches, this survey emphasizes the unique advantages and applications of deep learning-driven UAV platforms in fire scenarios. It provides detailed insights into various architectures, performance and applications used in UAV-based fire management. Additionally, the paper provides detailed insights into the available fire datasets along with their download links and outlines critical challenges, including data imbalance, privacy concerns, and real-time processing limitations. Finally, the survey identifies promising future directions, including multimodal sensor fusion, lightweight neural network architectures optimized for UAV deployment, and vision-language models. By synthesizing current research and identifying future directions, this survey aims to support the development of robust, intelligent UAV-based solutions for next-generation fire management. Researchers and professionals can access the GitHub repository.

Agentic AI represents a transformative shift in artificial intelligence, but its rapid advancement has led to a fragmented understanding, often conflating modern neural systems with outdated symbolic models—a practice known as conceptual retrofitting. This survey cuts through this confusion by introducing a novel dual-paradigm framework that categorizes agentic systems into two distinct lineages: the symbolic/classical (relying on algorithmic planning and persistent state) and the neural/generative (leveraging stochastic generation and prompt-driven orchestration). Through a systematic PRISMA-based review of 90 studies (2018–2025), we provide a comprehensive analysis structured around this framework across three dimensions: (1) the theoretical foundations and architectural principles defining each paradigm; (2) domain-specific implementations in healthcare, finance, and robotics, demonstrating how application constraints dictate paradigm selection; and (3) paradigm-specific ethical and governance challenges, revealing divergent risks and mitigation strategies. Our analysis reveals that the choice of paradigm is strategic: symbolic systems dominate safety-critical domains (e.g., healthcare), while neural systems prevail in adaptive, data-rich environments (e.g., finance). Furthermore, we identify critical research gaps, including a significant deficit in governance models for symbolic systems and a pressing need for hybrid neuro-symbolic architectures. The findings culminate in a strategic roadmap arguing that the future of Agentic AI lies not in the dominance of one paradigm, but in their intentional integration to create systems that are both adaptable and reliable. This work provides the essential conceptual toolkit to guide future research, development, and policy toward robust and trustworthy hybrid intelligent systems.

The broad application of passive acoustic monitoring provides a critical data foundation for studying soundscape ecology, necessitating automated analysis methods to accurately extract ecological information from vast soundscape data. This review comprehensively and cohesively examines two predominant approaches in soundscape analysis: soundscape component recognition and acoustic indices methods. Focusing on machine learning (ML)-based analysis methods for bird diversity assessment over the past five years, this review surveys representative research within each category, outlining their respective strengths and limitations. This not only addresses the growing interest in this field but also identifies research gaps and poses key questions for future studies. The insights from this review are anticipated to significantly enhance the understanding of ML applications in soundscape analysis, guiding subsequent investigative efforts in this rapidly evolving discipline, and thereby better supporting long-term biodiversity monitoring and conservation initiatives.