Articles

Improving user experience in accessing information on organizational websites remains a challenge. Users often face complex navigation and multi step searches that slow information retrieval. This study introduces the real time intelligent virtual assistant (RIVA), which integrates large language models (LLMs) with the retrieval-augmented generation (RAG) framework to support real-time interaction with website content. The system was implemented on the Universitas Pendidikan Ganesha (Undiksha) website using a WordPress content management system (CMS) and developed following the design science research (DSR) approach, which includes six stages: problem identification, solution objectives, design and development, demonstration, evaluation, and communication. The retrieval-augmented generation assessment (RAGAS) evaluation indicated that the combined model of text-embedding-ada-002 and semantic chunking yielded the best results, with context precision=0.83, context recall=0.90, response relevancy=0.91, faithfulness=0.83, and answer correctness=0.85. User experience questionnaire (UEQ) testing performed well, particularly in the novelty and stimulation dimensions. These results demonstrate that RIVA can provide users with access to relevant and engaging information. As a result, future research will focus on improving retrieval and developing adaptive semantic chunking for structured and complex data.

Large language models (LLMs) have paved a way for geographic information system (GIS) that can solve spatial problems with minimal human intervention. However, current commercial LLM-based GIS solutions pose many limitations for researchers, such as proprietary APIs, high operational costs, and internet connectivity requirements, making them inaccessible in resource-constrained environments. To overcome this, this paper introduced the LLM-Geo framework with the DS-GeoAI platform, integrating the DeepSeek-Coder model (the open-source, lightweight version deepseek-coder-1.3b-base) running directly on Google Colab. This approach eliminates API dependence, thus reducing deployment costs, and ensures data independence and sovereignty. Despite having only 1.3 billion parameters, DeepSeek-Coder proved to be highly effective: generating accurate Python code for complex spatial analysis, achieving a success rate comparable to commercial solutions. After an automated debugging step, the system achieved 90% accuracy across three case studies. With its strong error- handling capabilities and intelligent sample data generation, DS-GeoAI proves highly adaptable to real-world challenges. Quantitative results showed a cost reduction of up to 99% compared to API-based solutions, while expanding access to advanced geo-AI technology for organizations with limited resources.

The rise of hate and offensive content on social media platforms, such as Facebook and Twitter, has emerged as an escalating concern, especially in Vietnam. Consequently, detecting hate and offensive spans in Vietnamese text is an essential area of research. This study introduces ViHateOff, an advanced framework that combines a hated speech dictionary (HSD) automatically constructed from the Vietnamese hate and offensive spans (ViHOS) dataset with the pre-trained language models for Vietnamese (PhoBERT)-large language model to enhance the detection of offensive expressions. The framework functions through two primary modules. First, it constructs an HSD from the ViHOS dataset, which serves as a reference for identifying hate and offensive language in Vietnamese text. Second, the framework integrates the PhoBERT-large language model with HSD, enhancing the detection of harmful words in the input text. Experimental results demonstrate that the proposed framework significantly outperforms existing state-of-the-art (SOTA), achieving an F1-score of 0.8693 on the all spans subset and 0.8709 on the multiple-spans subset representing relative improvements of over 10% compared to the strongest baseline.

Transcribing natural language to visualization (NL2VIS) has been investigated for years but still suffer from several fundamental limitations (e.g., feature selection). Although large language models (LLMs) are good candidates but they incur computation cost and hard to trace their made decisions. To alleviate this problem, we introduced an alternative information-theoretic framework that utilized mutual information (MI) to quantify the statistical relationship between utterances and database features. In our approach, kernel density estimation (KDE) and neural estimation techniques were utilized to estimate MI, and to optimize a diversity-promoting objective balancing feature relevance and redundancy. We also introduced the information coverage ratio (ICR) to quantify the amount of information content preserved in feature selection decisions. In our experiments, we found that the proposed approach improved information-theoretic metrics, with F1-score of 0.863 and an ICR of 0.891. We observed that these improvements did not come at the cost of traditional benchmarks: validity reached 88.9%, legality 85.2%, and chart-type accuracy 87.6%. Moreover, significance tests (p < 0.001) and large effect sizes (Cohen’s d > 0.8) further supported that these improvements were meaningful for feature selection. Thus, this study provides a mathematical framework for applications requiring analytical validity that extends beyond NL2VIS to other machine learning contexts.

In the evolving field of Spirulina cultivation, the integration of the internet of things (IoT) has facilitated the optimization of spirulina growth and significantly enhanced biomass yield in the culture medium. This study outlines a control open-pond system for Spirulina cultivation that employs generative artificial intelligence (AI) and edge computing within an IoT framework. This transformative approach maintains optimal conditions and automates tasks traditionally managed through labor-intensive manual processes. The system is designed to detect, acquire, and monitor basin data via electronic devices, which is then analyzed by a large language model (LLM) to generate precise, context-aware recommendations based on domain-specific knowledge. The final output comprises SMS notifications sent to the farm manager, containing the generated recommendations, which keep them informed and enable timely intervention when necessary. To ensure continued autonomous operation in case of connectivity loss, pre-trained TinyML models were integrated into the Raspberry Pi. These models display alarm signals to alert the farm owner to any irregularities, thereby maintaining system stability and performance. This system has substantially improved the growth rate, biomass yield, and nutrient content of Spirulina. The results highlight the potential of this system to transform Spirulina cultivation by offering an adaptable, autonomous solution.

This research paper explores the use of large language model Meta AI 3 (LLaMA 3) for automating the resume screening process. Traditional resume screening methods that rely on keyword searching and human review can be inefficient, biased, and fail to identify qualified candidates. LLaMA 3, trained on large-scale text datasets, has the potential to accurately analyze resumes by understanding context and semantic details beyond simple keyword matching.The study presents a system that converts resume PDFs to text, inputs the text along with the job description into the LLaMA 3 model, and generates a ranked list of candidates with reasoning for their job fit. This discusses the data preparation, model setup, and performance evaluation of this system. Results show LLaMA 3 can rapidly process batches of resumes while reducing human bias in the screening process. The system aims to streamline hiring by automating the initial resume screening stage to surface top candidates for further in-depth evaluation. Key benefits include improved accuracy in identifying relevant skills, reduced bias compared to human screeners, and significant time savings for recruiters. The paper also examines ethical considerations around using AI for hiring decisions. Overall, this work demonstrates the promising application of large language models (LLMs) like LLaMA 3 to transform and enhance resume screening practices.

Large language models (LLMs) have demonstrated impressive capabilities in code generation; however, their high computational demands, privacy limitations, and challenges in edge deployment restrict their practical use in domain-specific applications. This study explores the effectiveness of parameter efficient fine-tuning for small language models (SLMs) with fewer than 3 billion parameters. We adopt a hybrid approach that combines low-rank adaptation (LoRA) and 4-bit quantization (QLoRA) to reduce fine-tuning costs while preserving semantic consistency. Experiments on the CodeAlpaca-20k dataset reveal that SLMs fine-tuned with this method outperform larger baseline models, including Phi-3 Mini 4K base, in ROUGE-L. Notably, applying our approach to the LLaMA 3 3B and Qwen2.5 3B models yielded performance improvements of 54% and 55%, respectively, over untuned counterparts. We evaluate models developed by major artificial intelligence (AI) providers Google (Gemma 2B), Meta (LLaMA 3 1B/3B), and Alibaba (Qwen2.5 1.5B/3B) and show that parameter-efficient fine-tuning enables them to serve as cost-effective, high-performing alternatives to larger LLMs. These findings highlight the potential of SLMs as scalable solutions for domain-specific software engineering tasks, supporting broader adoption and democratization of neural code synthesis.

In this study, we present an end-to-end, big-data–driven framework for continuously enriching and fine-tuning large language models (LLMs) with the latest professional and scientific medical knowledge. Streaming updates from premier sources such as The New England Journal of Medicine (NEJM) are ingested via an Apache Kafka cluster for low-latency delivery and durably archived in a three-node Apache Hadoop (Hadoop distributed file system (HDFS)) system. Each new article is preprocessed into high dimensional embeddings and indexed in a Milvus vector database to enable sub-second semantic retrieval over millions of records. At query or batch time, our retrieval-augmented generation (RAG) module retrieves the top-k relevant embeddings from Milvus and injects them into prompts for DeepSeek-R1, GPT-4o-mini, and Llama 3, models which are hosted, fine tuned, and served via Ollama on an NVIDIA GeForce RTX 3050 Ti GPU for efficient inference and continual learning. The enriched outputs are seamlessly delivered to end users through a Telegram bot programmed in Python using the Telebot library, linking the RAG-enhanced LLMs to an intuitive chat interface. Our Kafka, HDFS, Milvus, RAG, LLM, or Telegram bot pipeline demonstrably improves factual accuracy and topical currency of AI-generated medical insights across clinical decision support, patient engagement and education, drug discovery and development, virtual health assistants, and mental health support, laying the groundwork for truly intelligent, responsive, and data-driven healthcare solutions.

Automatic short answer grading (ASAG) systems offer a promising solution for improving the efficiency of reading literacy assessments. While promising, current Indonesian artificial intelligence (AI) grading systems still have room for improvement, especially when dealing with different domains. This study explores the effectiveness of large language models, specifically bidirectional encoder representations from transformers (BERT) variants, in conjunction with traditional hand-engineered features, to improve ASAG accuracy. We conducted experiments using various BERT models, hand-engineered features, text pre-processing techniques, and dimensionality reduction. Our findings show that BERT models consistently outperform traditional methods like term frequency-inverse document frequency (TF-IDF). IndoBERTLite-Base-P2 achieved the highest quadratic weighted kappa (QWK) score among the BERT variants. Integrating handengineered features with BERT resulted in a substantial enhancement of the QWK score. Utilizing comprehensive text pre-processing is a critical factor in achieving optimal performance. In addition, dimensionality reduction should be carefully used because it potentially removes semantic information.

Automatic code repair is an important task in software development to reduce bugs efficiently. This research focuses on developing and evaluating a chain-of-thought (CoT) prompting approach to improve the ability of large language models (LLMs) in automated program repair (APR) tasks. CoT prompting is a technique that guides LLM to generate step-by-step explanations before providing the final answer, so it is expected to improve the accuracy and quality of code repair. This research uses the QuixBugs dataset to evaluate the performance of several LLM models, including DeepSeek-V3 and GPT-4o, with two prompting methods, namely standard and CoT prompting. The evaluation is based on the average number of plausible patches generated as well as the estimated token usage cost. The results show that CoT prompting improves performance in most models compared with the standard. DeepSeek-V3 recorded the highest performance with an average of 36.6 plausible patches and the lowest cost of $0.006. GPT-4o also showed competitive results with an average of 35.8 plausible patches and a cost of $0.226. These results confirm that CoT prompting is an effective technique to improve LLM reasoning ability in APR tasks.

Large language models (LLMs) are widely deployed in settings where both reliability and efficiency matter. We present a calibrated, seed‑robust empirical comparison of an encoder fine‑tuned model (bidirectional encoder representations from transformers (BERT)‑base) and a decoder in‑context model (generative pre-trained transformer (GPT)‑2 small) across Stanford question answering dataset v2.0 (SQuAD v2.0) and general language understanding evaluation (GLUE)-multi-genre natural language inference (MNLI), Stanford sentiment treebank 2 (SST‑2). Beyond accuracy, we assess reliability (expected calibration error with reliability diagrams and confidence–coverage analysis) and efficiency (latency, memory, throughput) under matched conditions and three fixed seeds. BERT‑base yields higher accuracy and lower calibration error, while GPT‑2 narrows gaps under few‑shot prompting but remains more sensitive to prompt design and context length. Efficiency benchmarks show that decoder‑only prompting incurs near‑linear latency/memory growth with k‑shot exemplars, whereas fine‑tuned encoders maintain stable per‑example cost. These findings offer practical guidance on when to prefer fine‑tuning versus prompting and demonstrate that reliability must be evaluated alongside accuracy for risk‑aware deployment.

Monitoring mental health via social media often utilizes unimodal approaches, such as sentiment analysis on text or single-staged image categorization, or executes early feature fusion. However, in real-world contexts where emotions are conveyed via text, emojis, and images, unimodal approach leads to obscured decision-making pathways and overall diminished performance. To overcome these limitations, we propose EmoVibe, a hybrid multimodal AI framework for emotive analysis. EmoVibe uses attention-based late fusion strategy, where text embeddings are generated from bidirectional encoder representations from transformers (BERT) and visual features are extracted by vision transformer. Subsequently, emoticon vectors linked to avatars are processed independently. Later, these independent data features are integrated at higher levels, enhancing interpretability and performance. In contrast to early fusion methods and integrated multimodal large language models (LLMs) like CLIP, Flamingo, GPT-4V, MentaLLaMA, and domain-adapted models like EmoBERTa, EmoVibe preserves modality-specific contexts without premature fusion. This architecture saves processing cost, allowing for clearer, unambiguous rationalization and explanations. EmoVibe outperforms unimodal baselines and early fusion models, obtaining 89.7% accuracy on GoEmotions, FER, and AffectNet, compared to BERT's 87.4% and ResNet-50's 84.2%, respectively. Furthermore, a customizable, real time, privacy-aware dashboard is created, supporting physicians and end users. This technology enables scalable and proactive intervention options and fosters user self-awareness of mental health.

This document describes the implementation and evaluation of a retrieval-augmented generation (RAG) system to improve access to and understanding of Moroccan law, particularly the family code in Arabic. The research addresses the drawbacks of the widely used linguistic model applied to complex legal terminology in Arabic and aims to help citizens access crucial legal data. We built a new custom dataset with 2.5 k question-answer pairs while preprocessing and using the BGE-m3 embedding model in this experiment. Performance metrics, such as mean reciprocal rank (MRR), Recall@k, and F1-score, indicate that the RAG approach is effective compared to the use of standalone large language models (LLMs). Moreover, an evaluation on metrics such as the blue score, fidelity, response relevance, and contextual relevance indicated that the matching of meanings and context were well captured, which signifies a very good semantic understanding. The research highlights the need for language-specific model specialization in Arabic and presents its main challenges, such as dialectal variations and appropriate evaluation measures. The results indicate that well-developed RAG systems offer a promising approach to improving access to legal information in Arabic-speaking practice communities and to guiding future research and development in this field.

This paper presents the development of a low-cost learning assistant embedded in an NVIDIA Jetson Xavier board that uses speech and gesture recognition, together with a long language model for offline work. Using the large language model (LLM) Phi-3 Mini (3.8B) model and the Whisper (model base) model for automatic speech recognition, a learning assistant is obtained under a compact and efficient design based on extensive language model architectures that give a general answer set of a topic. Average processing times of 0.108 seconds per character, a speech transcription efficiency of 94.75%, an average accuracy of 9.5/10 and 8.5/10 in the consistency of the responses generated by the learning assistant, a full recognition of the hand raising gesture when done for at least 2 seconds, even without fully extending the fingers, were obtained. The prototype is based on the design of a graphical interface capable of responding to voice commands and generating dynamic interactions in response to the user's gesture detection, representing a significant advance towards the creation of comprehensive and accessible human-machine interface solutions.

Right in the context of digital marketing, this paper presents a comprehensive integrated system that combines the latest artificial intelligence advancement – large language models and diffusion models – to generate marketing email subjects and content that result in higher engagement. The system uses finetuned large language models for compelling email subject generation and finetuned Stable Diffusion model for visually appealing and convincing email content images creation. For the latter, both knowledge graphs and vector embeddings have been incorporated to improve contextual relevance. Experimental results demonstrated significant improvement in all engagement metrics that marketers rely on, including 46% growth in open rates, 56% higher click-through rates, and an 51% boost in conversion rates, all compared to human generated content. The unified approach presented by this paper outperforms standalone models and human-generated content in terms of engagement, as the comparative analysis shows. We also discuss the ethical considerations related to content bias and personalization boundaries, alongside challenges faced in this type of projects, such as computation demands and probable solutions. Finally, this paper proposes future directions to be taken, including expansion to other digital marketing channels, the use of other advanced artificial intelligence techniques, and the development of real-time content adaptation mechanisms based on user feedback.