Artificial Intelligence Review最新文献

Artificial intelligence (AI) has undergone transformative advances since 2022, particularly through generative AI, large language models (LLMs), and diffusion models, fundamentally reshaping the creative industries. However, existing reviews have not comprehensively addressed these recent breakthroughs and their integrated impact across the creative production pipeline. This paper addresses this gap by providing a systematic review of AI technologies that have emerged or matured since our 2022 review, examining their applications across content creation, information analysis, post-production enhancement, compression, and quality assessment. We document how transformers, LLMs, diffusion models, and implicit neural representations have established new capabilities in text-to-image/video generation, real-time 3D reconstruction, and unified multi-task frameworks–shifting AI from support tool to core creative technology. Beyond technological advances, we analyze the trend toward unified AI frameworks that integrate multiple creative tasks, replacing task-specific solutions. We critically examine the evolving role of human-AI collaboration, where human oversight remains essential for creative direction and mitigating AI hallucinations. Finally, we identify emerging challenges including copyright concerns, bias mitigation, computational demands, and the need for robust regulatory frameworks. This review provides researchers and practitioners with a comprehensive understanding of current AI capabilities, limitations, and future trajectories in creative applications.

Cancer diagnosis has entered an era where precision depends not only on image quality but on the intelligence that interprets it. While deep learning has revolutionized medical imaging, its reliance on centralized data limits collaboration due to privacy constraints and fragmented data ownership. Federated Learning (FL) offers a breakthrough enabling multiple institutions to co-train robust diagnostic models without sharing sensitive patient data. This survey provides a comprehensive cancer-specific synthesis of state-of-the-art FL applications in medical imaging, spanning five critical domains: lung, breast, brain, skin, and colorectal cancers. Beyond summarizing prior work, we uncover patterns in architecture choice (U-Net variants, Convolutional Neural Network (CNN)–Recurrent Neural Network(RNN) hybrids, dataset reuse, and state-of-the-art privacy frameworks such as homomorphic encryption and blockchain-backed consensus. We expose performance bottlenecks, heterogeneity risks, and critical absences of clinical deployment benchmarks. What emerges is not just a landscape of what has been done but a roadmap for what is needed to build secure, distributed, and clinically validated Artificial Intelligence (AI) systems. This work offers a foundation for researchers and healthcare technologists aiming to close the gap between research silos and real-world deployment.

Artificial Intelligence-Generated Content (AIGC) represents a paradigm shift in biomedical research and healthcare delivery, offering unprecedented capabilities for content creation, medical data analysis, and patient care optimization. This review examines the evolution of AIGC technologies from rule-based systems to advanced multimodal large models, with specific focus on their applications in healthcare settings. We analyze the three core capabilities of AIGC: intelligent digital content twinning, editing, and creation, and their transformative potential in medical imaging, clinical documentation, drug discovery, and personalized medicine. This paper discusses key challenges including algorithmic transparency, data privacy, and regulatory compliance, particularly in light of World Health Organization (WHO) guidelines for AI in health. Our findings indicate that while AIGC technologies show remarkable promise in enhancing diagnostic accuracy, streamlining clinical workflows, and democratizing healthcare access, careful consideration of ethical implications and regulatory frameworks is essential for safe and effective implementation.

Traffic congestion has a major impact on urban mobility, resulting in travel delays, fuel use, and emissions. Traffic Signal Control (TSC) is one of the main strategies to alleviate these issues. This systematic review using PRISMA combines 50 peer-reviewed articles from 2015 to 2025, addressing Evolutionary Algorithms (EAs) and Swarm Intelligence (SI) methods applied to TSC optimization. Our review compares algorithmic performance, application contexts, and parameter adjustment strategies. Findings confirm that hybrid methods (e.g., Genetic Algorithms (GA) + Particle Swarm Optimization (PSO)) perform better than single algorithms, with up to a 28.9% decrease in average vehicle delay. PSO shows higher resilience for real-time usage, while GA provides robustness for offline, multi-objective planning. Parameter tuning plays an important role in improving performance, with the best GA mutation rates (0.01–0.1) and PSO inertia coefficients (~ 0.7) delivering the optimal results. The present review synthesizes current evidence into practical recommendations for researchers, transportation planners, and policymakers seeking to promote traffic management effectiveness and environmental sustainability.

Deep generative models, such as generative adversarial networks (GANs), variational autoencoders (VAEs), and diffusion models, have demonstrated remarkable success in static medical imaging tasks, including image synthesis, domain translation, and data augmentation. However, their application to dynamic vascular modalities such as digital subtraction angiography (DSA) and X-ray angiography (XA) has received comparatively limited attention. This systematic review synthesizes 20 peer-reviewed studies published between January 2018 and July 2025, offering a comprehensive analysis of generative model usage in angiographic imaging. We categorize methods by model type, learning paradigm, and application focus, covering tasks such as image generation, vessel segmentation, and stenosis detection. GANs dominate the field, while diffusion models show emerging promise. Key limitations include the scarcity of public datasets, inconsistent evaluation metrics, and limited code availability, which hinder reproducibility and benchmarking. We conclude by outlining methodological trends and highlighting future directions, including the development of anatomically conditioned generative models and the need for large-scale, open-access datasets to support robust evaluation and clinical translation.

Demand for mobility is growing, and traffic on roads has increased substantially, leading to suboptimal traffic flow and congestion. Road pricing can encourage vehicles to change their behavior by charging for road use. Because traffic is not static, dynamic road pricing can help dynamically control traffic. Reinforcement Learning is an effective approach to optimizing the performance of a system. It has already been applied to control traffic signals and has recently found an application in dynamic road pricing for traffic optimization. We survey recent solutions and find that the methods proposed demonstrate the usefulness of reinforcement learning for road pricing. We compared how common challenges in reinforcement learning were approached in the works. Challenges which remain little explored are generalizability and scalability of solution approaches. Approaches to partial observability, credit assignment and non-stationarity are not in all cases taking full account of existing solutions for these common challenges. We further note the need for standardized benchmarks to allow comparisons between the performance of the provided solutions.

The rapid proliferation of multimodal generative models has sparked critical discussions on their reliability, fairness and potential for misuse. While text-to-image models excel at producing high-fidelity, user-guided content, they often exhibit unpredictable behaviors and vulnerabilities that can be exploited to manipulate class or concept representations. To address this, we propose an evaluation framework to assess model reliability by analyzing responses to global and local perturbations in the embedding space, enabling the identification of inputs that trigger unreliable or biased behavior. Beyond social implications, fairness and diversity are fundamental to defining robust and trustworthy model behavior. Our approach offers deeper insights into these essential aspects by evaluating: (i) generative diversity, measuring the breadth of visual representations for learned concepts, and (ii) generative fairness, which examines the impact that removing concepts from input prompts has on control, under a low guidance setup. Beyond these evaluations, our method lays the groundwork for detecting unreliable, bias-injected models and tracing the provenance of embedded biases. Our code is publicly available at https://github.com/JJ-Vice/T2I_Fairness_Diversity_Reliability.

Large pre-trained models have demonstrated remarkable capabilities across domains, but their comparative effectiveness in time series forecasting, especially against smaller, efficient models, remains underexplored. This work empirically examines whether pre-trained large-scale time series models (LSTSMs) trained on diverse datasets can outperform traditional non-pretrained small-scale transformers in forecasting tasks. We specifically compare large models trained from scratch against those benefiting from pretraining to measure the direct impact of transfer learning on forecasting performance. We analyze state-of-the-art (SOTA) pre-trained universal time series models (e.g., Moirai, GPT4TS, Timer, CALF, LLM4TS) alongside conventional small-scale transformers, evaluating accuracy and computational efficiency across multiple benchmarks. We further conduct an extensive ablation study across varying fine-tuning data sizes (10%, 25%, and 75%) to assess few-shot, moderate, and near full-data adaptation capabilities. Additionally, explainability of large time series models is examined using comprehensiveness via feature ablation, occlusion, integrated gradients and gradient shap methods. Besides that, interpretability of pretraining and finetuning strategies is also examined using spectral metrics via WeightWatcher to quantify layer-wise generalization and representation quality, while theoretical and quantitative computational complexity analyses, including parameter counts, training time, model sizes, and inference latency, highlight the trade-offs between predictive performance and resource efficiency. Our findings reveal the strengths and limitations of pre-trained large-scale models, providing insights into their suitability for time series tasks compared to task-specific small-scale architectures. The results highlight scenarios where pretraining offers advantages and where simpler models remain competitive.

Text-to-image (T2I) models are rapidly advancing into creative practice and increasingly support generating illustrated storybooks, i.e., sequential and image-based narratives conditioned on written text. Previous surveys have examined challenges in video coherence or single-image fidelity. To our best knowledge, there is no comprehensive review that addresses the unique requirements of storybook illustration. This survey fills this gap by grounding the study of AI-illustrated storybook generation in a narratology framework. Specifically, this survey introduces a six-dimensional consistency model encompassing time, space, character, event and plot, style, and theme. For each dimension, we include consolidate definitions, representative methods, datasets, and evaluation metrics, thereby mapping the current landscape of the field. Building on this analysis, we further identify cross-dimensional failure modes and limitations of current approaches. Finally, we propose potential future research directions, including the development of book-scale integrated evaluation systems tailored for illustrated storybooks, more robust and controllable generation pipelines, enhanced multimodal semantic–visual alignment mechanisms, and the establishment of reader-oriented safety and educational guidelines.

This research proposes a holistic agentic Artificial Intelligence framework that seeks to solve the palpable requirement of autonomy, versatility, and world-scale generative Artificial Intelligence systems. When framing the concept of an agentic Artificial Intelligence as a paradigm shift where Large Language Models are relatively reactive, tool-augmented to a proactive, modular agent that pursues goals in the long term, the research discusses how large language models are becoming an architectural mandate. The proposed research uses a skillful literature synthesis and develops a concept and technical framework that integrates perception, memory, planning, execution, and communication modules. In contrast to current frameworks like AutoGPT and ReAct, where the deep memory, explainability, and certain control of scalability are often lacking, the proposed framework offers a solution with persistent memory layers, semantic routing, and modular orchestration pipelines when it comes to cloud-native deployments. Experimental verification offers a greater degree of autonomy, coordination, and resilience in a diverse range of activities such as enterprise automation and robotics. It is also an edge deployment with the help of lightweight microservices framework. In practice, this method supports scaled, comprehensible agents adapted to long-loop thinking and human-in-the-loop management and adaptive value chain serving. The novelty of the work is attributed to its reusable architecture, as it is not only capable of modifying agentic behavior alone, but it can also connect the theoretical principles and industry feasibility. Future efforts will be with the integration of ethical governance, uniform benchmarking, and multimodal memory remodeling in next generation of real-world autonomous systems.