With the rapid development of artificial intelligence (AI) and manufacturing, fault diagnosis (FD), as a critical technology for ensuring equipment safety and improving production efficiency, faces challenges such as increasing data complexity and higher requirements for diagnostic accuracy. Traditional FD methods primarily rely on expert experience and statistical learning, exhibiting limitations in handling multimodal data fusion, knowledge representation, and reasoning. Large language models (LLMs) have brought new opportunities to the field of FD. This review systematically examines the current applications, technical approaches, and development trends of LLMs in FD. Through comparative analysis of the advantages and disadvantages of different technical approaches, we summarize the main achievements and existing technical bottlenecks in current research. LLMs demonstrate clear advantages in FD, however, they also face challenges such as high data quality requirements, significant computational resource consumption, and limited domain adaptability. Future development trends are expected to focus on model lightweighting, domain specialization, multimodal fusion, and enhanced real-time diagnostic capabilities. This review offers systematic theoretical guidance and technical reference for researchers and engineering professionals.
This study proposes a multi-agent Large Language Model (LLM) framework for formulaic alpha generation and selection in quantitative finance. In this context, an alpha refers to a predictive trading signal or mathematical expression designed to capture excess returns beyond market benchmarks. While a single LLM can generate candidate alphas, relying on one model often results in redundancy, limited diversity, and suboptimal signal quality. To address this, we adopt a multi-agent approach where DeepSeek is responsible for generating a broad pool of formulaic alphas, and Gemini serves as an evaluator that selects the top ten most important alphas from this pool based on criteria such as predictive strength and diversity of signals. To assess the effectiveness of the framework, we compare two experimental setups: (1) an equal-weighted portfolio constructed from the top ten alphas selected by Gemini, and (2) an equal-weighted portfolio of all 50 alphas generated by DeepSeek. Empirical results demonstrate that the multi-agent framework can achieve better portfolio performance, delivering higher cumulative returns, improved risk-adjusted measures, and lower maximum drawdown relative to the full alpha pool. These findings highlight the potential of multi-agent LLM collaboration in enhancing alpha discovery and systematic trading strategies.
We present Cognitive SOC, a multi-agent system leveraging Large Language Models (LLMs) to transform raw security alerts into evidence-backed narratives with complete citation chains. The proposed architecture employs specialized Triage, Enrichment, and StoryWeaver agents that collaborate to analyze security incidents and gather contextual intelligence through integration with threat-intelligence frameworks, including MITRE ATT&CK, Sigma detection rules (via a curated subset), and CVE databases. Evaluated on 1,000 alerts from the UNSW-NB15 dataset, the system achieved 99.6% evidence linkage rate and 100% citation coverage while maintaining an exceptionally low hallucination rate of 0.1%. The Cognitive SOC demonstrated superior accountability compared to baseline approaches, with a mean time-to-recommendation of 1.860 seconds and comprehensive evidence tracking. The system generated an average of 3.8 evidence artifacts and 5.4 citations per alert, substantially outperforming single-agent RAG (3.0 evidence, 0 citations, 11.8% hallucination) and static runbook baselines (1.0 evidence, 0 citations). This work fundamentally reframes security operations as collaborative human-AI storytelling, shifting the analyst role from manual data correlation to narrative validation and strategic decision-making.
In a world of proliferating synthetic multimedia content, it is increasingly important to develop the ability to trace the origin of digital assets and verify their authenticity. A bold initiative is the Coalition for Content Provenance and Authenticity (C2PA), which proposes a data model for associating authenticated provenance information, known as content credentials, with multimedia assets. This paper situates C2PA within the field of Computational Archival Science (CAS), examining how cryptographic provenance and authenticity frameworks can operationalize archival trustworthiness in big data environments. With a focus on digital images (include AIgenerated), this paper explores the creation of a C2PA-based pipeline for digital asset preservation informed by archival science, computer science and cybersecurity. Using an analytical framework derived from archival diplomatics and computational provenance modeling, the study maps authenticity metadata for digital images. The role of C2PA in records management and archival preservation of AI-generated images is introduced. The use of emerging blockchain technology to support provenance of multimedia content is discussed in detail. Alternative solutions are discussed. C2PA security risks are reviewed. The findings demonstrate that C2PA represents a computational model of provenance - transforming authority-based trust into trust by design. This contribution addresses a key CAS research challenge of establishing provenance for digital assets across distributed systems.
Large Language Models (LLMs) are increasingly used in multi-agent systems, where autonomous agents collaborate on complex tasks through distributed reasoning and communication. While existing work focuses on privacy in single-agent or API-bound settings, inter-agent communication introduces new privacy leakage pathways that remain underexplored. This paper proposes a unified framework for quantifying and mitigating privacy leakage in multi-agent LLMs. It unifies Mutual Information (MI), Bayesian Inference Gain (BIG), and Renyi Differential Privacy (RDP) to jointly quantify empirical leakage, belief refinement, and worst-case privacy loss in multi-agent LLMs. MI captures statistical dependence between private inputs and observable messages, detecting unintended memorization or semantic correlation. BIG models adversarial belief refinement and complements MI by offering a posteriorcentric interpretation of leakage. RDP provides formal privacy guarantees across multi-turn dialogues, quantifying divergence under repeated queries and enabling DP-style defense via output perturbation. We simulate LLM agents with distinct prompts and introduce probing adversaries capable of membership inference, prompt injection, and extraction attacks. The empirical results show that the proposed combined defense reduces MI from 0.35 to 0.10 and the attack success rate (ASR) from 82% to 29%, with minimal Bilingual Evaluation Understudy (BLEU) drop. The combined defense achieves strong privacy with acceptable utility. These findings highlight the importance of privacy-aware agent design and protocol-level defense under black-box threats in multi-agent LLMs.
Understanding how urban morphology shapes human activity is crucial for promoting sustainable and equitable cities. However, traditional studies often rely on static or aggregated data, limiting insights into micro-scale behavioral patterns. This study integrates high-resolution GPS mobility data -GNSS-based sanitized mobility dataset (NDA restricted access to the data) with OpenStreetMap (OSM) data to analyze the spatial relationship between urban form and activity distribution at a 1km x 1km grid level. Using spatial autocorrelation, regression, and clustering methods, we identify key morphological features-such as building density, POI diversity, and attractive areas-that influence activity intensity and demographic composition. Results show that higher density and functional richness are associated with greater mobility diversity and more spatially concentrated walking activity, while attractive areas particularly affect older populations. The findings deepen understanding of the dynamic interaction between urban form and human behavior, offering evidence for data-driven urban planning.
Advances in remote sensing have enabled continuous access to global, multi-temporal imagery, creating a “big data” environment for large-scale deforestation mapping. Yet accurate detection of deforestation remains difficult due to heterogeneous disturbance patterns, complex spectral signatures, seasonal variability, and pronounced class imbalance. To address these challenges, we systematically evaluated recent deep learning methods in realistic, multi-region settings. We created a unified benchmark for forest change detection using a multi-year Sentinel-2 dataset (2017-2023) from 27 REDD+ projects across six tropical countries. Using a standardized pipeline to ensure fair comparison, we evaluated three representative deep learning families: ResUNet (a convolutional encoder-decoder), LWGANet (a lightweight attention architecture), and ChangeMamba (a state-space sequence model). Results revealed clear trade-offs in accuracy, efficiency, and robustness. On the global data split, LWGANet and ChangeMamba outperformed ResUNet, with ChangeMamba-Tiny achieving the best balance of accuracy and efficiency. In cross-site (out-of-sample) evaluations, the performance of CNN- and attention-based models degraded significantly, whereas ChangeMamba maintained stable performance, demonstrating resilience to geographic domain shifts. Qualitative analysis also showed that the state-space model yields cleaner boundaries and greater sensitivity to subtle forest degradation. These findings indicate that state-space models are a promising foundation for robust, efficient, and transferable large-scale deforestation mapping.
Root Cause Analysis (RCA) is essential for pinpointing the root causes of failures in microservice systems. Traditional data-driven RCA methods are typically limited to offline applications due to high computational demands, and existing online RCA methods handle only single-modal data, overlooking complex interactions in multi-modal systems. In this paper, we introduce OCEAN, a novel online multi-modal causal structure learning method for root cause localization. OCEAN introduces a long-term temporal causal learning module with two encoders: one captures stable causal dependencies from historical data, while the other models short-term variations in the current batch data. We further design a multi-factor attention mechanism to analyze and reassess the relationships among different metrics and log indicators/attributes for enhanced online causal graph learning. Additionally, a contrastive mutual information maximization-based graph fusion module is developed to effectively model the relationships across various modalities. Extensive experiments on three real-world datasets demonstrate the effectiveness and efficiency of our proposed method.
User-controllable privacy is essential in modern wearable and mobile devices that collect motion data from Inertial Measurement Unit (IMU) sensors, as privacy preferences can vary significantly from person to person and may evolve over time. While prior work has proposed privacy-preserving methods for sensor data, most rely on static, predefined privacy labels or require large quantities of private training data, limiting their adaptability and user controllability. In this work, we introduce PrivCLIP, a dynamic, user-controllable, few-shot privacy-preserving sensing framework that enables users to specify and modify their privacy preferences. Leveraging a multimodal contrastive learning approach, PrivCLIP aligns IMU sensor data with natural language activity descriptions in a shared embedding space, enabling few-shot detection of sensitive activities. PrivCLIP uses a language-guided activity sanitizer and a motion generation module (IMU-GPT) to transform the original data into a privacy-compliant version that semantically resembles a non-sensitive activity. We evaluate PrivCLIP on multiple human activity recognition datasets and demonstrate that it significantly outperforms baseline methods in terms of both privacy protection and data utility.
Transforming structured data into natural language sentences for pre-trained language models(PLMs) is an emerging paradigm to enhance downstream prediction tasks. While recent studies have primarily focused on designing sentence-generation templates for general domains, their application in the clinical domain remains limited. This limitation poses unique challenges for effectively representing structured preoperative data and predicting postoperative complications. Existing sparse and dense representations of preoperative data often suffer from limitations, including the loss of medical semantics, insufficient semantic associations among features, neglect of numerical unit information, and distortion caused by handling missing values. To address these limitations, this study proposes a sentenceization template tailored for preoperative data, converting structured records into natural language-like descriptions, which are subsequently encoded by PLMs to support complication risk prediction. Experimental validation on the prediction of major postoperative cardiovascular adverse events(MACE) demonstrates that, compared with comparison models such as LightGBM, the proposed sentenceization approach significantly improves both F 1 and recall, with the relative increase in recall reaching up to 40.9%.
Coherence is one of the desirable properties that determines the quality of AI-generated texts. Coherence ensures that the ideas are connected in a logical flow and the consistency in information and facts is maintained. Therefore, models for understanding and detecting incoherence are crucial to achieve high-quality AI text generation capabilities. In this paper, we focus on the task of detecting sentences that are inconsistent with the overall document context. Supervised models may not be viable for this task due to the lack of labeled data on text incoherence. Therefore, we propose a contrastive selfsupervised learning method that relies on both semantic content and neighborhood contexts of sentences for detecting incoherence. We propose a complex model that integrates graph neural networks with language models to contrast and compare two representations of each sentence-one based on its content and the other on its neighborhood context-for detecting incoherent sentences. Extensive experiments on several real-world datasets show the effectiveness of our proposed method.
Advanced medical devices increasingly rely on AIdriven frameworks to automate compliance processes, ensuring safety and efficacy while reducing regulatory burdens. In the US, software-based medical devices, including those utilizing AI/ML models, are regulated by the FDA's Center for Devices and Radiological Health (CDRH) under the Code of Federal Regulations (CFR) Title 21. These regulations are extensive, cross-referenced documents that require significant human effort to parse, leading to high compliance costs for manufacturers. We propose a novel, semantically rich framework that extracts regulatory knowledge from FDA documents and translates it into a machine-processable format. Our system encodes regulatory knowledge into an OWL/RDF based knowledge graph and uses the Mistral 7B Instruct model to dynamically generate SPARQL queries, perform compliance reasoning, and produce structured reports. This enables automated device classification (Class I, II, or III) and real time regulatory evaluation. Validated through real-world use cases, our framework significantly reduces manual review effort, enhances interpretability, and accelerates time-tomarket. The proposed approach integrates AI reasoning and semantic technologies to achieve scalable, transparent, and automated regulatory compliance.
Artificial Intelligence (AI)-powered features have rapidly proliferated across mobile apps in various domains, including productivity, education, entertainment, and creativity. However, how users perceive, evaluate, and critique these AI features remains largely unexplored, primarily due to the overwhelming volume of user feedback. In this work, we present the first large-scale, data-driven study of user feedback on AIpowered mobile apps, leveraging a curated dataset of 292 AIdriven apps across 14 categories with nearly one million AI-specific reviews from Google Play. We develop and validate a multi-stage analysis pipeline that begins with a human-labeled benchmark and systematically evaluates large language models (LLMs) and prompting strategies. Each stage, including review classification, aspect-sentiment extraction, and clustering, is validated for accuracy and consistency. Our pipeline enables scalable, highprecision analysis of user feedback, extracting over one million aspect-sentiment pairs clustered into 18 positive and 15 negative user topics. Our analysis reveals that users consistently focus on a narrow set of themes: positive comments emphasize productivity, reliability, and personalized assistance, while negative feedback highlights technical failures (e.g., scanning and recognition), pricing concerns, and limitations in language support. Our pipeline surfaces both satisfaction with one feature and frustration with another within the same review. These fine-grained, co-occurring sentiments are often missed by traditional approaches that treat positive and negative feedback in isolation or rely on coarsegrained analysis. To this end, our approach provides a more faithful reflection of the real-world user experiences with AIpowered apps. Category-aware analysis further uncovers both universal drivers of satisfaction and domain-specific frustrations. We expect our findings to advance understanding of user-centered development of AI-powered features and provide actionable guidance to software engineers and app developers who seek to align AI features of apps with user expectations.
With the increasing ubiquity of web systems across various domains, they have become primary targets of cyberattacks. Among these, SQL Injection (SQLi) remains one of the most critical and prevalent vulnerabilities, consistently listed in the OWASP Top 10. Early detection and mitigation of such flaws are crucial for ensuring information security. This paper presents the design and implementation of an automated SQLi penetration testing system based on a multi-agent architecture integrated with Large Language Models (LLMs) such as GPT-4, Gemini, and DeepSeek. The system enhances penetration testing by automating vulnerability detection, reducing reliance on manual effort, and supporting proactive defense strategies. The key contributions include: (i) flexible LLM integration, (ii) automated report generation with visual statistics, and (iii) comparative analysis of LLM performance against existing approaches. The results demonstrate the potential of combining multi-agent systems and LLMs to improve both the efficiency and intelligence of penetration testing, laying the foundation for future smart and proactive cybersecurity systems.
This paper presents a comprehensive framework for classifying normal and abnormal electrocardiogram (ECG) patterns by integrating unsupervised and supervised learning. The methodology begins with Fuzzy C-Means (FCM) clustering to generate pseudo-labels for ECG data, followed by a benchmarking of various classifiers, including Logistic Regression, Support Vector Machine, K-Nearest Neighbors, Decision Trees, Artificial Neural Networks, and Long Short-Term Memory (LSTM) networks. Evaluated on a large, aggregated dataset from PhysioBank database, the results demonstrate that the proposed approach effectively groups ECG signals. After basic filtering and clustering, the LSTM model effectively distinguishes between normal and abnormal heart activities. It consistently achieved superior performance, with an F1 score of up to 99.6%, highlighting the potential of deep learning to capture complex temporal dynamics in ECG data for enhanced cardiac diagnostics.
Clinical time series are often irregularly sampled, with varying sensor frequencies, missing observations, and misaligned timestamps. Prior approaches typically address these irregularities by interpolating data into regular sequences, thereby introducing bias, or by generating inconsistent and uninterpretable relationships across sensor measurements, complicating the accurate learning of both intra-series and inter-series dependencies. We introduce WaveGNN, a model that operates directly on irregular multivariate time series without interpolation or conversion to a regular representation. WaveGNN combines a decay-aware Transformer to capture intra-series dynamics with a sample-specific graph neural network that models both short-term and long-term inter-sensor relationships. Therefore, it generates a single, sparse, and interpretable graph per sample. Across multiple benchmark datasets (P12, P19, MIMIC-III, and PAM), WaveGNN delivers consistently strong performance, whereas other state-of-the-art baselines tend to perform well on some datasets or tasks but poorly on others. While WaveGNN does not necessarily surpass every method in every case, its consistency and robustness across diverse settings set it apart. Moreover, the learned graphs align well with known physiological structures, enhancing interpretability and supporting clinical decision-making.
Timeline Summarization (TLS) extends traditional multi-document summarization by generating summaries while explicitly modeling temporal information, enabling efficient understanding of event evolution. The method of using large language models (LLMs) can achieve state-of-the-art (SOTA) performance, but frequent interactions with LLMs greatly increase the processing time of the task. In this study, we propose a time-aware graph structural entropy minimization approach for unsupervised timeline summarization(GSE-TLS), which employs a graph clustering algorithm guided by structural entropy. Traditional graph construction methods for clustering rely solely on semantic similarity, rendering them insufficient for the TLS task. To address this, we design a novel graph construction strategy tailored to TLS. Additionally, when selecting representative nodes for event clusters, we compute node influence via structural entropy, overcoming the limitations of conventional degree- or frequency-based methods-notably their disregard for semantic relationships. Experiments on benchmark datasets demonstrate that GSE-TLS surpasses baselines in most metrics and achieves significantly faster clustering speeds than SOTA approaches. The data and source code are available at: https://github.com/pengfan8252/GSE-TLS/
Federated learning (FL) enables collaborative model training across decentralized clients without sharing raw data. However, its deployment in real-world settings faces severe challenges, including data heterogeneity, system heterogeneity with device capability variance, communication overhead, and privacy concerns. A common limitation of existing work is the lack of consideration for the joint impact of these issues, ultimately leading to degraded performance. These limitations are particularly evident in healthcare, where data distribution differs across hospitals and patient groups, which makes a single global model often inadequate. In this paper, we propose PFedCDP for medical imaging, a unified framework for privacy-preserving, heterogeneity-aware personalized federated learning (PFL), which addresses these issues. PFedCDP introduces a device capability-aware client clustering mechanism that assigns appropriately scaled neural architectures to devices based on computational capacity, memory, battery, and network metrics. Personalized training is then guided by a mechanism that preserves locally important parameters while still incorporating global knowledge, ensuring that clients benefit from both personalization and collaboration. To have a generalized server model, we add a post-hoc refinement stage that strengthens representation robustness and effectively transfers diverse cluster knowledge. Furthermore, the integration of differential privacy and model quantization ensures privacy and communication efficiency. Experimental evaluations on mammography datasets demonstrate that PFedCDP outperforms state-of-the-art baselines, achieving robust personalization and privacy preservation under heterogeneous and non-IID conditions in real-world healthcare scenarios. Our implementation repository is publicly available at https://github.com/shadhin39/PFedCDP.
With the widespread adoption of big data science, the retail industry has seen an increasing demand for data-driven decision-making. As a critical task within the sector, product sales forecasting faces significant challenges. Traditional statistical methods are limited by their reliance on assumptions of data stationarity, while DL and ML techniques improve accuracy but struggle to fully address complex features such as temporal dependencies, nonlinear trends, and external influences. No single method is universally effective in all scenarios. To address these challenges, this article introduces a Unified Sales Forecasting Framework (USFF) for the prediction of vending machine sales, which integrates statistical methods, machine learning, and deep learning techniques. The framework's core contribution is a data-driven architecture that systematically addresses data heterogeneity, significantly improving prediction accuracy by effectively leveraging the strengths of various models. Experiments conducted on more than 30 million real-world data points demonstrate that the proposed framework outperforms baseline methods across multiple evaluation metrics, confirming their effectiveness and reliability. The proposed framework and algorithm have been successfully implemented in a retail company in China, providing valuable business decision-making support for operational strategies.
Security analysts are overwhelmed by the volume of alerts and the low context provided by many detection systems. Early-stage investigations typically require manual correlation across multiple log sources, a task that is usually time-consuming. In this paper, we present an experimental, agentic workflow that leverages large language models (LLMs) augmented with predefined queries and constrained tool access (structured SQL over Suricata logs and grep-based text search) to automate the first stages of alert investigation. The proposed workflow integrates queries to provide an overview of the available data, and LLM components that selects which queries to use based on the overview results, extracts raw evidence from the query results, and delivers a final verdict of the alert. Our results demonstrate that the LLM-powered workflow can investigate log sources, plan an investigation, and produce a final verdict that has a significantly higher accuracy than a verdict produced by the same LLM without the proposed workflow. By recognizing the inherent limitations of directly applying LLMs to high-volume and unstructured data, we propose combining existing investigation practices of real-world analysts with a structured approach to leverage LLMs as virtual security analysts, thereby assisting and reducing the manual workload.
