Search Results (9,031)

Data annotation serves as a critical foundation for artificial intelligence (AI) and machine learning (ML). Recently, AI agents powered by large language models (LLMs) have emerged as effective solutions to longstanding challenges in data annotation, such as scalability, consistency, cost, and limitations in domain expertise. These agents facilitate intelligent automation and adaptive decision-making, thereby enhancing the efficiency and reliability of annotation workflows across various fields. Despite the growing interest in this area, a systematic understanding of the role and capabilities of AI agents in annotation is still underexplored. This paper seeks to fill that gap by providing a comprehensive review of how LLM-driven agents support advanced reasoning strategies, adaptive learning, and collaborative annotation efforts. We analyze agent architectures, integration patterns within workflows, and evaluation methods, along with real-world applications in sectors such as healthcare, finance, technology, and media. Furthermore, we evaluate current tools and platforms that support agent-based annotation, addressing key challenges such as quality assurance, bias mitigation, transparency, and scalability. Lastly, we outline future research directions, highlighting the importance of federated learning, cross-modal reasoning, and responsible system design to advance the development of next-generation annotation ecosystems. Full article

Social media platforms like Instagram significantly shape destination images and influence tourist behavior. Understanding how different cities are represented and perceived on these platforms is crucial for effective tourism marketing. This study provides a comparative analysis of Instagram content and engagement patterns in Seoul and Tokyo, two major Asian metropolises, to derive actionable marketing insights. We collected and analyzed 59,944 public Instagram posts geotagged or location-tagged within Seoul (n = 29,985) and Tokyo (n = 29,959). We employed a mixed-methods approach involving content categorization using a fine-tuned convolutional neural network (CNN) model, engagement metric analysis (likes, comments), Valence Aware Dictionary and sEntiment Reasoner (VADER) sentiment analysis and thematic classification of comments, geospatial analysis (Kernel Density Estimation [KDE], Moran’s I), and predictive modeling (Gradient Boosting with SHapley Additive exPlanations [SHAP] value analysis). A validation analysis using balanced samples (n = 2000 each) was conducted to address Tokyo’s lower geotagged data proportion. While both cities showed ‘Person’ as the dominant content category, notable differences emerged. Tokyo exhibited higher like-based engagement across categories, particularly for ‘Animal’ and ‘Food’ content, while Seoul generated slightly more comments, often expressing stronger sentiment. Qualitative comment analysis revealed Seoul comments focused more on emotional reactions, whereas Tokyo comments were often shorter, appreciative remarks. Geospatial analysis identified distinct hotspots. The validation analysis confirmed these spatial patterns despite Tokyo’s data limitations. Predictive modeling highlighted hashtag counts as the key engagement driver in Seoul and the presence of people in Tokyo. Seoul and Tokyo project distinct visual narratives and elicit different engagement patterns on Instagram. These findings offer practical implications for destination marketers, suggesting tailored content strategies and location-based campaigns targeting identified hotspots and specific content themes. This study underscores the value of integrating quantitative and qualitative analyses of social media data for nuanced destination marketing insights. Full article

Deep neural network (DNN) hardware (HW) accelerators have achieved great success in improving DNNs’ performance and efficiency. One key reason is the dataflow in executing a DNN layer, including on-chip data partitioning, computation parallelism, and scheduling policy, which have large impacts on latency and energy consumption. Unlike prior works that required considerable efforts from HW engineers to design suitable dataflows for different DNNs, this work proposes an efficient data-centric approach, named Dataflow Code Propagation (DCP), to automatically find the optimal dataflow for DNN layers in seconds without human effort. It has several attractive benefits that prior studies lack, including the following: (i) We translate the HW dataflow configuration into a code representation in a unified dataflow coding space, which can be optimized by back-propagating gradients given a DNN layer or network. (ii) DCP learns a neural predictor to efficiently update the dataflow codes towards the desired gradient directions to minimize various optimization objectives, e.g., latency and energy. (iii) It can be easily generalized to unseen HW configurations in a zero-shot or few-shot learning manner. For example, without using additional training data, Extensive experiments on several representative models such as MobileNet, ResNet, and ViT show that DCP outperforms its counterparts in various settings. Full article

Large language models (LLMs) often tend to hallucinate, especially in domain-specific tasks and tasks that require reasoning. Previously, we introduced SubGraph Retrieval Augmented Generation (SG-RAG) as a novel Graph RAG method for multi-hop question answering. SG-RAG leverages Cypher queries to search a given knowledge graph and retrieve the subgraph necessary to answer the question. The results from our previous work showed the higher performance of our method compared to the traditional Retrieval Augmented Generation (RAG). In this work, we further enhanced SG-RAG by proposing an additional step called Merging and Ordering Triplets (MOT). The new MOT step seeks to decrease the redundancy in the retrieved triplets by applying hierarchical merging to the retrieved subgraphs. Moreover, it provides an ordering among the triplets using the Breadth-First Search (BFS) traversal algorithm. We conducted experiments on the MetaQA benchmark, which was proposed for multi-hop question-answering in the movies domain. Our experiments showed that SG-RAG MOT provided more accurate answers than Chain-of-Thought and Graph Chain-of-Thought. We also found that merging (up to a certain point) highly overlapping subgraphs and defining an order among the triplets helped the LLM to generate more precise answers. Full article

Catalysis research is complex and interdisciplinary, involving diverse physical effects and challenging data practices. Research data often captures only selected aspects, such as specific reactants and products, limiting its utility for machine learning and the implementation of FAIR (Findable, Accessible, Interoperable, Reusable) workflows. To improve this, semantic structuring through ontologies is essential. This work extends the established ontologies by refining logical relations and integrating semantic tools such as the Web Ontology Language or the Shape Constraint Language. It incorporates application programming interfaces from chemical databases, such as the Kyoto Encyclopedia of Genes and Genomes and the National Institute of Health’s PubChem database, and builds upon established ontologies. A key innovation lies in automatically decomposing chemical substances through database entries and chemical identifier representations to identify functional groups, enabling more generalized reaction classification. Using new semantic functionality, functional groups are flexibly addressed, improving the classification of reactions such as saponification and ester cleavage with simultaneous oxidation. A graphical interface (GUI) supports user interaction with the knowledge graph, enabling ontological reasoning and querying. This approach demonstrates improved specificity of the newly established ontology over its predecessors and offers a more user-friendly interface for engaging with structured chemical knowledge. Future work will focus on expanding ontology coverage to support a wider range of reactions in catalysis research. Full article

Thermal management is an area of study in electronics focused on managing temperature to improve reliability and efficiency. When temperatures are too high, cooling systems are activated to prevent overheating, which can lead to reliability issues. To monitor the temperatures, sensors are often placed on-chip near hotspot locations. These sensors should be very small to allow them to be placed among compact, high-activity circuits. Often, they are connected to a central control circuit located far away from the hot spot locations where more area is available. This paper proposes sensing units for a novel temperature sensing architecture in the TSMC 180 nm process. This architecture functions by approximating the current through the sensing unit at a reference voltage, which is used to approximate the temperature in the digital back end using fitting functions. Sensing units are selected based on how well its temperature–current relationship can be modeled, sensing unit area, and power consumption. Many sensing units will be experimented with at different reference voltages. These temperature–current curves will be modeled with various fitting functions. The sensing unit selected is a diode-connected p-type MOSFET (Metal Oxide Semiconductor Field Effect Transistor) with a size of W = 400 nm, L = 180 nm. This sensing unit is exceptionally small compared to existing work because it does not rely on multiple devices at the sensing unit location to generate a PTAT or IPTAT signal like most work in this area. The temperature–current relationship of this device can also be modeled using a 2nd order polynomial, requiring a minimal number of trim temperatures. Its temperature error is small, and the power consumption is low. The range of currents for this sensing unit could be reasonably made on an IDAC. Full article

Background/Objectives: Neuromodulation is under investigation as a possibly effective add-on therapy in Alzheimer’s disease (AD). While transcranial pulse stimulation (TPS) has shown positive short-term effects, long-term effects have not yet been fully explored. This study aims to evaluate the long-term feasibility, safety, and potential cognitive benefits of TPS over one year in patients with Alzheimer’s disease, focusing on domains such as memory, speech, orientation, visuo-construction, and depressive symptoms. Methods: We analyzed preliminary data from the first ten out of thirty-five patients enrolled in a prospective TPS study who completed one year of follow-up and were included in a dedicated long-term database. The protocol consisted of six initial TPS sessions over two weeks, followed by monthly booster sessions delivering 6000 pulses each for twelve months. Patients underwent regular neuropsychological assessments using the Alzheimer Disease Assessment Scale (ADAS), Mini-Mental Status Examination (MMSE), Montreal Cognitive Assessment (MoCA), and Beck Depression Inventory (BDI-II). All adverse events (AEs) were documented and monitored throughout the study. Results: Adverse events occurred in less than 1% of stimulation sessions and mainly included mild focal pain or transient unpleasant sensations, as well as some systemic behavioral or vigilance changes, particularly in patients with underlying medical conditions, with some potentially related to the device’s stimulation as adverse device reactions (ADRs). Cognitive test results showed significant improvement after the initial stimulation cycle (ADAS total improved significantly after the first stimulation cycle (M_pre = 28.44, M_post = 18.56; p = 0.001, d = 0.80, 95% CI (0.36, 1.25)), with stable scores across all domains over one year. Improvements were most notable in memory, speech, and mood. Conclusions: TPS appears to be a generally safe and feasible add-on treatment for AD, although careful patient selection and monitoring are advised. While a considerable number of participants were lost to follow-up for various reasons, adverse events and lack of treatment effect were unlikely primary causes. A multifocal stimulation approach (F-TOP2) is proposed to enhance effects across more cognitive domains. Full article

(1) Background and objectives: Large language models (LLMs) such as GPT, Mistral, and LLaMA exhibit strong capabilities in text generation, yet assessing the quality of their reasoning—particularly in open-ended and argumentative contexts—remains a persistent challenge. This study introduces Dialectical Agent, an internally developed modular framework designed to evaluate reasoning through a structured three-stage process: opinion, counterargument, and synthesis. The framework enables transparent and comparative analysis of how different LLMs handle dialectical reasoning. (2) Methods: Each stage is executed by a single model, and final syntheses are scored via two independent LLM evaluators (LLaMA 3.1 and GPT-4o) based on a rubric with four dimensions: clarity, coherence, originality, and dialecticality. In parallel, a rule-based semantic analyzer detects rhetorical anomalies and ethical values. All outputs and metadata are stored in a Neo4j graph database for structured exploration. (3) Results: The system was applied to four open-weight models (Gemma 7B, Mistral 7B, Dolphin-Mistral, Zephyr 7B) across ten open-ended prompts on ethical, political, and technological topics. The results show consistent stylistic and semantic variation across models, with moderate inter-rater agreement. Semantic diagnostics revealed differences in value expression and rhetorical flaws not captured by rubric scores. (4) Originality: The framework is, to our knowledge, the first to integrate multi-stage reasoning, rubric-based and semantic evaluation, and graph-based storage into a single system. It enables replicable, interpretable, and multidimensional assessment of generative reasoning—supporting researchers, developers, and educators working with LLMs in high-stakes contexts. Full article

Carbon cap-and-trade is a practical policy in guiding manufacturers to produce economic and environmental production plans. However, previous studies on carbon cap-and-trade are from a macro level to guide manufacturers to make production plans, rather than from a perspective of specific production scheduling, which leads to a lack of theoretical guidance for manufacturers to develop reasonable production scheduling schemes for specific production orders. This article investigates a specific scheduling problem in a flexible job-shop environment that considers the carbon cap-and-trade policy, aiming to provide guidance for specific production scheduling (i.e., resource allocation). In the proposed problem, carbon emissions have an upper limit. A penalty will be generated if the emissions overpass the predetermined cap. To satisfy the carbon emission cap, the manufacturer can trade carbon credits or adopt outsourcing strategy, that is, outsourcing partial orders to partners at the expense of outsourcing costs. To solve the proposed model, a novel and efficient memetic algorithm (NEMA) is proposed. An initialization method and four local search operators are developed to enhance the search ability. Numerous experiments are conducted and the results validate that NEMA is a superior algorithm in both solution quality and efficiency. Full article

Detection of small vessels is of great significance for maritime safety assurance, abnormal vessel tracking, illegal fishing supervision, and combating smuggling. However, the radar reflection intensity of small vessels is low, making them difficult to detected with the radar’s constant false-alarm rate (CFAR) algorithm. To enhance the detection capability for small vessels, we propose an improved CFAR scheme. Specifically, we first compared traditional CFAR processing results of radar data with automatic identification system (AIS) data to identify some special targets. These special targets, which possessed AIS information, but remained undetected by radar, enabled an iTransformer model to generate more reasonable CFAR threshold adjustments. iTransformer adaptively lowered the threshold of the areas around these targets until they were detected by radar. This process made it easier to discover the small boats in the surrounding area. Experimental results showed that our method reduces the missed detection rate of small vessels by 73.4% and the false-alarm rate by 60.7% in simulated scenarios, significantly enhancing the CFAR detection capability. Overall, our study provides a new solution for ensuring maritime navigation safety and strengthening illegal supervision, while also offering new technical references for the field of radar detection. Full article

Medical education often emphasizes theoretical knowledge, limiting students’ opportunities to practice history taking, a structured interview that elicits relevant patient information before clinical decision making. Large language models (LLMs) offer novel solutions by generating simulated patient interviews. This study evaluated the educational potential of LLM-generated history-taking dialogues, focusing on clinical validity and diagnostic diversity. Chest pain was chosen as a representative case given its frequent presentation and importance for differential diagnosis. A fine-tuned Gemma-3-27B, specialized for medical interviews, was compared with GPT-4o-mini, a freely accessible LLM, in generating multi-branching history-taking dialogues, with Claude-3.5 Sonnet inferring diagnoses from these dialogues. The dialogues were assessed using a Chest Pain Checklist (CPC) and entropy-based metrics. Gemma-3-27B outperformed GPT-4o-mini, generating significantly more high-quality dialogues (90.7% vs. 76.5%). Gemma-3-27B produced diverse and focused diagnoses, whereas GPT-4o-mini generated broader but less specific patterns. For demographic information, such as age and sex, Gemma-3-27B showed significant shifts in dialogue patterns and diagnoses aligned with real-world epidemiological trends. These findings suggest that LLMs, particularly those fine-tuned for medical tasks, are promising educational tools for generating diverse, clinically valid interview scenarios that enhance clinical reasoning in history taking. Full article

Per- and polyfluoroalkyl substances (PFAS) have been found worldwide in water, soil, plants, and animals, including humans. A primary route of exposure for humans and animals to PFAS is through the diet and drinking water. Perfluorooctane sulfonate (PFOS), a long-chain PFAS with a relatively long half-life, has been associated with adverse health effects in humans and laboratory animals. There are few toxicokinetic studies on PFOS in domestic livestock raised for human food consumption, which are critical for assessing human food safety. This work aimed to develop a simple daily accumulation model (DAM) for predicting PFOS residues in edible beef cattle muscle. A one-compartment toxicokinetic model in a spreadsheet format was developed using simple calculations to account for daily PFAS into and out of the animal. The DAM was used to simulate two case studies to predict resultant PFOS residues in edible beef cattle tissues. The results demonstrated that the model can reasonably predict PFOS concentrations in beef cattle muscle in a real-world scenario. The DAM was then used to simulate dietary PFOS exposure in beef cattle throughout a typical lifespan in order to derive a generic bioaccumulation factor. The DAM is expected to work well for other PFAS in beef cattle, PFAS in other livestock species raised for meat, and other chemical contaminants with relatively long half-lives. Full article

Background/Objectives: Chemically or genetically detoxified pertussis toxin (PTx) is a crucial antigen component of the acellular pertussis vaccine. Chemical detoxification using glutaraldehyde generally causes significant structural changes to the toxin. However, how these structural changes in PTx affect its antigenic properties remains unclear. Additionally, there is limited knowledge regarding how many alterations in antigenic properties impact immunogenicity. Methods: To investigate the impact of structural changes on antigenic properties, we developed a sandwich ELISA to quantify the neutralizing epitopes on PTx. Subsequently, we analyzed different PTx toxoid (PTd) preparations with the assay. Additionally, we assessed the immunogenicity of various acellular pertussis vaccine candidates containing these PTd preparations. Finally, the assay was applied to evaluate the consistency of commercial batches of PTx and PTd intermediates. Results: The assay demonstrated reasonable specificity, accuracy, and precision, and it was sensitive enough to quantify variations in neutralizing epitopes among different PTd samples that shared the same protein concentration. Importantly, we found a positive correlation between the number of neutralizing epitopes in detoxified PTx and its immunogenicity, indicating that the amount of neutralizing epitopes present determines the immunogenicity of glutaraldehyde-inactivated PTx. Moreover, commercial batches of PTx and PTd intermediates exhibited minor variations in neutralizing epitopes. Conclusions: These findings have significant implications for developing acellular pertussis vaccines as they highlight the importance of preserving the neutralizing epitopes of PTx during detoxification to ensure the vaccine’s effectiveness. This assay is also valuable for the quality control of PTd as it more accurately represents the actual antigenic changes of PTx. Full article

Automated vehicles controlled by artificial intelligence are becoming capable of making moral decisions independently. This study investigates the differences in participants’ perceptions of the moral decision-maker’s permissibility when viewing scenarios (pre-test) and after witnessing the outcomes of moral decisions (post-test). It also investigates how permissibility, ten typical moral emotions, and perceived moral agency fluctuate when AI and the human driver make deontological or utilitarian decisions in a pedestrian-sacrificing dilemma (Experiment 1, N = 254) and a driver-sacrificing dilemma (Experiment 2, N = 269) from a third-person perspective. Moreover, by conducting binary logistic regression, this study examined whether these factors could predict the non-decrease in permissibility ratings. In both experiments, participants preferred to delegate decisions to human drivers rather than to AI, and they generally preferred utilitarianism over deontology. The results of perceived moral emotions and moral agency provide evidence. Moreover, Experiment 2 elicited greater variations in permissibility, moral emotions, and perceived moral agency compared to Experiment 1. Moreover, deontology and gratitude could positively predict the non-decrease in permissibility ratings in Experiment 1, while contempt had a negative influence. In Experiment 2, the human driver and disgust were significant negative predictor factors, while perceived moral agency had a positive influence. These findings deepen the comprehension of the dynamic processes of autonomous driving’s moral decision-making and facilitate understanding of people’s attitudes toward moral machines and their underlying reasons, providing a reference for developing more sophisticated moral machines. Full article

Adapting Large Language Models (LLMs) to specialized domains like geriatric care remains a significant challenge due to the limited availability of domain-specific data and the difficulty of achieving efficient yet effective fine-tuning. Current methods often fail to effectively harness domain-specific causal insights, which are crucial for understanding and solving complex problems in low-resource domains.To address these challenges, we propose Causality-Aware, Parameter-Efficient Learning (CPEL), a novel framework that leverages domain-specific causal relationships to guide a multi-layer, parameter-efficient fine-tuning process for more effective domain adaptation. By embedding causal reasoning into the model’s adaptation pipeline, CPEL enables efficient specialization in the target domain while maintaining strong task-specific performance. Specifically, the Causal Prompt Generator of CPEL extracts and applies domain-specific causal structures, generating adaptive prompts that effectively guide the model’s learning process. Complementing this, the MPEFT module employs a dual-adapter mechanism to balance domain-level adaptation with downstream task optimization. This cohesive design ensures that CPEL achieves resource efficiency while capturing domain knowledge in a structured and interpretable manner. Based on this framework, we delved into its application in the field of geriatric care and trained a specialized large language model (Geriatric Care LLaMA) tailored for the aged-care domain, leveraging its capacity to efficiently integrate domain expertise. Experimental results from question-answering tasks demonstrate that CPEL improves ROUGE scores by 9–14% compared to mainstream LLMs and outperforms frontier models by 1–2 points in auto-scoring tasks. In summary, CPEL demonstrates robust generalization and cross-domain adaptability, highlighting its scalability and effectiveness as a transformative solution for domain adaptation in specialized, resource-constrained fields. Full article