Articles

This review provides comparative analysis of GPT-4, BERT (bidirectional encoder representations from transformers), Gemini, and DeepSeek large language models (LLM), focusing architectures, training methodologies, and real-world applications. The primary research question is: How do these models differ in design, strengths, limitations, and potential areas for enhancement? By addressing this question, the study aims to provide insights into the trade-offs and future directions for optimizing LLM performance and deployment. The analysis reveals GPT-4 excels in natural language generation and complex reasoning, supporting up to 128K tokens with moderate latency and higher costs making it effective for conversational artificial intelligence (AI). BERT excels bidirectional contextual understanding with smaller computational overhead and broad open-source adoption, effective for text classification. Gemini demonstrates superior multimodal integration processing text, image, audio, and code with context lengths up to 1M tokens, enabling cross-domain adaptability. DeepSeek excels in specialized domains like finance and programming, is optimized for efficiency and supports extended context windows exceeding 200K tokens. However, all models face challenges related to computational cost, hallucinations, and ethical concerns, necessitating further improvements. Despite advancements, LLMs continue to grapple with issues such as data bias, model interpretability, and responsible AI deployment. Future research should focus on hybrid model approaches, domain-specific fine-tuning, and transparency to mitigate risks while maximizing the transformative potential of LLMs in real-world applications.

This study empirically evaluates generative fusion decoding (GFD) with open-weight large language models for Indonesian automatic speech recognition (ASR) under two controlled offline conditions: general read speech (FLEURS-ID) and medical dictation (Harkit). The training-free byte-level fusion integrates three frozen language-model backends—SEA-LION, Sahabat-AI, and Mistral—with frozen Whisper-large-v3 during decoding. A development partition was used to select the static fusion weight and prompting condition, and a held-out test partition was used for final evaluation. On FLEURS-ID, the best GFD configuration reduced corpus-level word error rate (WER) from 5.91% to 4.17%, corresponding to a 29.36% relative reduction. On Harkit, it reduced WER from 15.89% to 12.30%, a 22.59% relative reduction. Paired utterance-level tests yielded p < 0.01 for both datasets. Several non-prompted configurations already improved over the unfused Whisper-large-v3 baseline, suggesting that the external language-model prior contributed more consistently than explicit prompting. Prompting produced limited changes on FLEURS-ID and larger but backend-dependent changes on Harkit. The results support GFD as a training-free decoding strategy under the evaluated conditions, while broader validation is required for noisy, spontaneous, streaming, and clinical deployment settings.

Autonomous agents powered by large language models (LLMs) are deployed widely across finance, healthcare, legal review, and customer support. Each deployment assumes that an agent will remain within its assigned role—an assumption that adversarial users routinely challenge through identity drift attacks. This work asks whether placing enforcement logic outside the model, rather than embedding rules in the system prompt, is the primary factor governing drift suppression. We designed RMIC-Guard: a framework where a cryptographically sealed Role-Model Identity Contract (RMIC) is checked by a dual-pass middle ware engine on every planned tool call before execution proceeds. The engine runs a deterministic hard-rule pass followed by a semantic Intent Drift Score (IDS) computed from seven independent embedding-based signals. Experiments across four LLMs (claude-sonnet-4-6, claude-haiku-4-5, llama-3.3-70b-versatile, llama-3.1-8b-instant), four agent roles, and 5,212 API calls show that external enforcement reaches DSR 0.952–1.000, whereas the best prompt-only condition achieves only DSR 0.391. Ablation confirms the dual-pass architecture is necessary: neither pass alone achieves acceptable joint drift-suppression and false-positive performance. All code, contracts, and prompt data are released openly.

Autonomous agents transacting with each other need an escrow they can trust without a human in the loop. This study propose a context-aware micro-escrow framework for this setting: large language models (LLMs), the model context protocol (MCP), agent-to-agent (A2A) communication, and blockchain smart contracts together support automated, auditable micro-transactions among heterogeneous resource holders such as live application programming interfaces (APIs), internet of things (IoT) sensors, and service providers. Traditional escrow relies on rigid condition checks or a trusted intermediary. This study instead runs a deterministic, checklist-based semantic validation pipeline that emits signed binary verdicts and commits rationale hashes on-chain, so any verdict can be replayed off-chain while on-chain settlement stays minimal. Context travels in cryptographically stamped MCP envelopes; negotiation and evidence streaming run over authenticated A2A channels. An auditing layer computes risk signals from recent settlement metadata and applies a bounded fee surcharge that internalizes monitoring effort and performance drift. The framework is evaluated with a reproducible simulator under benign and adversarial mixes, reporting latency, decision quality (precision and recall under a binary-classifier view), and economic side effects (fee collection and slashing). The results show strong tolerance to malicious providers. Coordinated validator compromise, however, marks a clear trust boundary, motivating validator-side hardening as future work.

Startup growth and sustainability are heavily impacted by user comments. This research has done a comparative study between large language models (LLMs), small language models (SLMs), and traditional machine learning (ML) methods in classifying user comments on startups. The performances of bidirectional encoder representations from transformers (BERT), Indonesian bidirectional encoder representations from transformers (IndoBERT), distilled bidirectional encoder representations from transformers (DistilBERT), and ML models; logistic regression, support vector machine (SVM), random forest, and extreme gradient boosting (XGBoost) were evaluated with a diverse dataset and beyond extensive preprocessing. Out of all the models this research trained, IndoBERT outperformed with an accuracy and recall 90.99% and F1-score of 89.94%, due to its intermediate pretraining for Indonesian language task. DistilBERT, from the SLM family of models and only requiring a fraction of resources to make predictions performed significantly better (over 90% in both accuracy and recall). These outcomes indicate that LLMs, particularly those targeting specific languages, are significantly superior in performance for sentiment analysis. The results seem to imply that models like these can be utilized to study the phenomenon and further enhance user engagement with startups. Future work will involve the further development of more language specific BERT versions and expanding LLMs in other ways within startup spaces.

High-dimensional data can reduce interpretability and increase computational cost in machine learning (ML) pipelines. Large language models (LLMs) offer a promising approach for dimensionality reduction through feature selection (FS), but existing LLM-based methods lack structured reasoning and do not provide clear justifications for their choices. This study propose LLM-FS-Agent, a multi-agent architecture where specialized LLM roles collaborate through a deliberative process to evaluate and rank features. Each selection decision is accompanied by an explicit rationale. We tested our method on the CIC-DIAD 2024 internet of things (IoT) intrusion detection dataset and compared it against LLM-select and traditional techniques like principal component analysis (PCA) across multiple subset sizes. LLM-FS-Agent achieves comparable or better accuracy while reducing classifier training time by 46% on average (statistically significant, 0.094 s, p = 0.028 for extreme gradient boosting (XGBoost)). These results show that multi-agent deliberation improves transparency and computational efficiency in FS.

Structured query language injection (SQL injection/SQLi) remains a critical threat to web application security, particularly because it can bypass traditional detection systems through obfuscation. Existing approaches, including rule-based systems and classical machine learning (ML) models, often fail to capture the contextual semantics of com0plex SQL queries. This study proposes a lightweight and context-aware SQLi detection model by fine-tuning the large language model Meta artificial intelligence (LLaMA) 3.2–1B large language model using low-rank adaptation (LoRA). The model was trained using the SQL-Injection-Extend dataset obtained from Kaggle, consisting of 57,310 malicious SQLi samples and 52,208 benign user-input samples. The benign data comprise general user-generated text commonly submitted through web application interfaces. Experimental results demonstrate that the proposed model achieves an accuracy of 99.91%, precision of 99.96%, recall of 99.86%, and an F1-score of 99.91%, outperforming baseline models including convolutional neural network (CNN), long short-term memory (LSTM), and non-fine-tuned large language models (LLMs). These findings indicate that LoRA-based fine-tuning substantially enhances detection performance while maintaining computational efficiency, making the approach suitable for real-time deployment in modern web security systems. Furthermore, the integration of context-aware digital forensic reporting supports post-incident investigation and improves the interpretability of detection outcomes.

Over the past decades, Java-based applications have formed the backbone of enterprise, web, and mobile software systems, often containing complex business logic and domain-specific requirements embedded in older codebases (Java SE 1.2-8). Maintaining and migrating such systems is difficult due to unstructured code, missing documentation, and outdated design artifacts. To address these challenges, the paper introduces a computationally efficient static-backward slicing framework for Java programs. The framework restructures the source code, abstracts control and data flow information, computes functional dependencies using a data flow table (DFT), and extracts precise program slices. Extracted slices are further represented using object-Z notation to formalize software specifications. Furthermore, the extracted backward slices and object-Z specifications can be leveraged with artificial intelligence (AI) and large language models (LLMs) to automatically generate human-readable software requirements. This integration also ensures structured, accurate, and minimal requirements outputs, facilitating AI-assisted software modernization. Experimental evaluations on Java programs such as automated teller machine (ATM), student grading, and quadratic root calculators exhibit significant reductions in extracted code size while preserving functional dependencies. In comparison with traditional dynamic and machine learning (ML)-based slicing approaches, the proposed framework offers improved precision, computational efficiency, and scalability, making it suitable for modernizing legacy Java systems and supporting AI-driven software engineering tasks.

The dynamic evolution of industrial automation has created an imperative need to transition from inflexible, rule-based systems to flexible, intelligent agents that can facilitate human-robot collaboration. The purpose of this research was to integrate large language models (LLMs) with autonomous mobile robots (AMRs) to improve context-aware decision-making. The inflexibility of traditional systems has often hindered performance in dynamic environments, as systems often rely on predefined algorithms and sensor configurations. To improve this, a modular framework was created, consisting of a central processing unit and an LLM API to interpret natural language and process environmental information. Quantitative results have been clearly specified in the abstract, which states that the success rate in resolving navigation exceptions by the proposed framework was 89% with a 5% localization error rate. Moreover, substantial savings were observed in token usage and computational resources. This study provided an imperative framework for smarter industrial automation, filling the gap between mechanical precision and artificial intelligence. The practical experimentation of an AMR model gives an outcome of a successful navigation exception resolution rate of 89% by means of the proposed framework, with an error rate of 5% localized to each exception. Additionally, significant reductions in the number of tokens used and the time taken to process tokens provide a scalable means for developing contextually-based, robust, autonomous mobile platforms’ decisions.

Significant advances have been achieved in machine translation (MT) in recent times, particularly state of the art (SOTA) models for languages like English and Indian having distinct grammatical structures and limited monolingual training data. This paper analyses various recent state-of-the-art variants of large language models (LLMs) and neural machine translation (NMT) for Indian languages in comparison to statistical machine translation (SMT). It tackles key questions, such as idiomatic expressions, morphologically complex grammar or the scarceness of parallel corpora. Furthermore, it studies bytewise BPE, compares translation models in terms of BLEU scores using separate and shared-vocabulary representation with copy actions between the BPE translations, and analyses how multitask learning (Caruana (1997)) and attention mechanisms can contribute to the quality of translation. In summary, it provides directions for future work by suggesting new avenues of research including better curated datasets, more efficient approaches for lowresource languages and culturally aware translations.

The increasing prevalence of toxic content on social media, conveyed through text, images, and videos, poses significant challenges for automated content moderation systems. Although prior studies have reported promising results in unimodal and bimodal settings, they often fail to capture implicit and contextual toxicity emerging from interactions across multiple modalities, particularly in non-English environments. This paper proposes WVisionBERT-VL, an end-to-end multimodal framework for toxicity detection that integrates text, image, and video modalities within a unified architecture. The proposed model incorporates modality-specific encoders, bidirectional multi-head cross attention (BMHCA) for cross-modal synchronization, and an adaptive fusion gate to dynamically balance modality contributions. A balanced multimodal dataset is constructed from social media platforms, including X, Instagram, and TikTok, and refined using a model-based labeling strategy with limited human-in-the-loop validation. Experimental results on a custom Indonesian dataset demonstrate strong in-domain performance, achieving an accuracy of 94.12%, a macro-F1 of 0.9407, and a receiver operating characteristic - area under the curve (ROC-AUC) of 0.9721, with robustness further validated through five-fold cross-validation. Cross-dataset evaluation highlights challenges related to domain shift, underscoring the need for future research on robust and domain-adaptive multimodal toxicity detection.

The present study aims to investigate the impact of integrating artificial intelligence (AI) technologies into teaching Russian as a foreign language (RFL) from the perspective of educators. Employing a mixed-methods research design, the study utilized several methodologies, including a teacher survey, an analytical-descriptive approach to data interpretation, and the development and evaluation of AI-based interventions. The study sample comprised 120 RFL instructors from three public universities in Kazakhstan. Preliminary findings revealed a considerable awareness among teachers regarding various AI technologies, such as chatbots, voice assistants, the ChatGPT neural network, educational platforms, gaming applications, and task design tools. Nonetheless, the practical utilization of these technologies varied significantly, with only a subset of teachers incorporating them into their regular teaching practices. The study culminated in the development of a conceptual framework for AI-driven educational interventions, incorporating platforms such as Coursera, Moodle, Open EdX, and eFront; game-based applications including Duolingo, Talk2Russia, and Russian Verbs Pro; and task creation tools such as Kahoot! and Quizlet. Following the integration of these interventions into the curriculum, post-implementation evaluations indicated that teachers generally perceived the tools as effective, with the average effectiveness rating surpassing 4.0 out of 5.0 across all assessed categories. The findings of this study have practical applicability; they can be used to enhance professional development programs for teachers of the Russian language and to formulate strategies for the integration of AI technologies into language education within the Central Asian region.

New students often struggle when short articles clash with thick textbooks. Still, even though large language models offer some teaching support, standard online setups lack focused accuracy - sometimes making things up - and risk user data control. Here comes an idea: build a tightly tested, self- contained system that aligns learning materials automatically without needing the internet, keeping information private by design. One look at two setups shows how they handle local reasoning differently. Instead of using both encoder and decoder parts, one system skips the encoder entirely. That simpler design grabs full context through ChromaDB without shrinking the data first. Meanwhile, the older type crunches input down, losing meaning along the way. Even though it runs fast - just under a second - errors pop up often, four out of five responses drifting off course. On the flip side, the new method builds correct code nearly every time, adds clear explanations tied to lesson goals, yet takes more than fourteen seconds to reply. Slower? Yes. More accurate? Clearly. What stands out is how compressed models running locally can still catch up in understanding classroom content. Another key point emerges: building tutors powered by artificial intelligence (AI) becomes safer when data never leaves the device and outside services are not needed at all.

This study examines the reliability of an image-native multimodal AI system for automated scoring of handwritten responses to Advanced Placement (AP) Calculus free-response items without requiring optical character recognition (OCR) preprocessing. Using inter-rater agreement indices and test–retest reliability analyses, we found substantial to almost perfect agreement between artificial intelligence (AI)-generated scores and calibrated human ratings, as well as almost perfect stability across repeated scoring sessions. These results suggest that the observed reliability of the AI scoring system warrants further investigation of validity-related evidence and inferences. As a practical implication for assessment practice, we propose a human-in-the-loop teacher–AI collaborative (TAC) framework in which automated scoring operates under teacher oversight. Taken together, these findings provide initial evidence of reliability supporting the responsible use of AI-based scoring as a measurement instrument in high-stakes educational assessment.

Social media platforms have greatly accelerated the spread of news, but this rapid information flow also amplifies the risk of misinformation. Traditional automatic detection methods that rely solely on textual features often struggle with nuanced, emerging content. This paper present a novel pipeline that verifies the factuality of social news by clustering related articles into events and using large language models (LLMs) to extract and verify claims against trusted news sources. The approach groups social-media posts and mainstream reports on the same event, extracts atomic claims with a model like GPT-4, and checks each claim’s truthfulness by comparing it to the cluster’s reliable news. This pipeline was evaluated on a newly constructed Vietnamese news dataset of 1,765 articles (including 723 social-media items), manually annotating claims as true or false. The LLM-based method achieved high accuracy (≈88.9% F1-score on claim verification and 92.1% F1-score on overall news verification). These results demonstrate that carefully prompted LLMs, combined with event-level clustering of evidence, can outperform traditional methods (e.g., bidirectional encoder representations from transformers (BERT)-based classifiers) in verifying news. The paper discusses advantages of clustering over simple retrieval, scalability considerations for LLMs, and prospects for multilingual and knowledge-enhanced verification.