Earticle

기존 비디오 게임은 사전 정의된 스크립트와 정적인 캐릭터 관계 모델로 인해 상호작용 과 몰입감에 한계가 있다. 최근 대규모 언어 모델(LLM)이 대안으로 떠올랐으나, 생성된 서 사와 게임 내부 데이터가 단절되는 문제가 여전히 존재한다. 이에 본 연구는 Unity 엔진, LLM, 게임 서버 플랫폼을 통합하여 캐릭터 데이터가 서사를 생성하고, 그 결과가 다시 데 이터베이스에 반영되는 '순환형 동적 스토리텔링 시스템'을 제안한다. 실험 결과, 캐릭터 성 향에 따른 차별화된 서사 생성 및 관계도의 실시간 DB 동기화를 확인하였다. 본 시스템은 기존 스크립트의 경직성을 극복하고 무한히 확장 가능한 개인화된 게임 경험을 제공한다.

Traditional video games suffer from limited interactions due to predefined scripts and static relationship models. While Large Language Models (LLMs) offer generative freedom, a disconnect remains between AI-generated narratives and internal game data. To address this, we propose a cyclical dynamic storytelling system integrating Unity, an LLM, and a game server. In this system, character attributes drive narrative generation, and the outcomes are continuously synchronized back to the game's database. Experiments confirmed that differentiated narratives were generated based on character alignments, with real-time updates to their affinity data. This approach overcomes the rigidity of traditional scripts, offering a highly personalized gaming experience.

초거대 언어 모델(LLM)이 발전하면서 사람들은 많은 곳에서 LLM을 사용하여 생산성을 증대시키고 있다. 이 는 사람과의 자연스러운 대화 진행, 추론 등을 할 수 있는 LLM의 능력을 미루어 볼 때 매우 자연스러운 일이라 고 할 수 있다. LLM은 회사, 일상생활뿐만 아니라 교육 현장에서도 사용되면서 많은 변화를 이끌어내고 있는 데, 학습자들이 정보 수집, 문제 해결 등의 다양한 분야에서 LLM을 활용하여 많은 효과를 보고 있는 것도 사실 이다. 그러나, 학습자가 수행해야 하는 활동을 대신하는 등의 무분별한 LLM의 사용은 학습자의 역량을 배양해 야 하는 교육 목표의 달성에 장애 요인으로 작용할 수 있는 가능성이 있다. 특히, 학습자의 자기 주도적인 활동 을 핵심으로 하고 있는 자기 조절 학습 모델에서 LLM의 올바른 사용은 매우 중요하다고 할 수 있다. 따라서, 본 연구에서는 자기 조절 학습 모델에서의 LLM의 역할에 대하여 탐구하고 활동의 예를 제안하였다.

With the advancement of Large Language Models (LLMs), people are increasingly using them to enhance their productivity. This is understandable, given LLMs' ability to facilitate natural conversations and reasoning. LLMs are driving significant change not only in the workplace and everyday life, but also in education. Learners are reaping significant benefits from their use in diverse areas, including information gathering and problem solving. However, indiscriminate use of LLMs, such as replacing required activities, can hinder the achievement of educational goals, which aim to cultivate learner competencies. In particular, the appropriate use of LLMs is crucial in self-regulated learning models, which prioritize learner self-directed activity. Therefore, this study explores the role of LLMs in self-regulated learning models and proposes examples of activities.

This study proposes an automated system for medical image modality classification that combines large language models (LLMs) and a custom deep learning architecture. First, the Llama 3.2 model was used to label the textual descriptions from the radiology objects in context (ROCO) dataset, which contained 81,000 medical images. The extracted labels and their corresponding images were then used to train a specialized image classification model consisting of residual and squeeze-and-excitation (SE) blocks with attention mechanisms. To ensure fair comparison and consistency, all samples were preprocessed, the data were augmented, and the class balance was maintained. The experiments revealed that the SEBwithAttention model obtained better results, with 88.10±0.145% accuracy and an F1-score of 83.55±0.295, which was higher than those of DenseNet, ResNet50, VGG16, AlexNet and vision transformers. In conclusion, the study findings demonstrated that LLM-based modality extraction could effectively classify medical images using their captions, enabling the training of a model that accurately identified image modalities.

Semantic Interoperability (SI) enables cross-domain data integration by allowing diverse systems to share and process information effectively. While existing reviews focus on general AI-driven interoperability, this systematic literature review (SLR) is the first to exclusively analyze the integration of Large Language Models (LLMs) with SI. This SLR uniquely evaluates LLMs' role in schema alignment, knowledge integration, and security risks. It also introduces a novel taxonomy and identifies challenges like bias propagation and computational costs, providing a new research framework for adversarial robustness, ethical AI, and real-world SI optimization.

Autonomous mobile robots (AMRs) must operate in dynamic, unstructured environments where traditional control and reinforcement learning (RL) face adaptability and reward design limitations. This study proposes a hybrid framework combining proximal policy optimization (PPO) with a large language model (LLM) as an adaptive reward designer. Using GPT-4o-mini, the LLM dynamically shapes rewards based on performance logs, improving exploration and stability. Experiments in complex indoor navigation show the LLM-PPO model reduces collisions by 38%, shortens completion time by 21%, and increases rewards by 8% over PPO. Results demonstrate LLM-RL integration enhances safety, efficiency, and consistency, offering a promising paradigm for AMR control.

The work presented in the paper challenges the common view of hallucination as a flaw in Large Language Models (LLMs), proposing instead that it is an inherent feature of their architecture. It explores the relationship between “creativity,” “generalization,” and “hallucination,” showing that they result from the transformer model’s high-dimensional vector blending. Using Gemini 2.0 and 2.5 models, the authors demonstrate a “Self-Correction Blind Spot,” where LLMs can recognize errors but lack the reasoning pathways to correct them. The paper introduces the “Generalization-Hallucination Trade-off,” suggesting that minimizing hallucination would impede generalization. The research advocates for managing hallucination with external verification rather than eliminating it.

Hate speech ought to be identified and minimized to diminish the undesirable influx of toxic content on the Internet, particularly in the fields of sensitive matters, such as the climate change debate. This article analyzes hate speech detection using climate change-related information and focuses on tweet classification and the identification of the hate target via prompt engineering with large language models (LLMs). Specifically, we compare three instruction-tuned LLMs, TinyLlama, Flan-T5, and Gemma, and evaluate them in contrast to a standard transformer baseline, bidirectional encoder representations from transformers (BERT). The aspects of the experimental framework include prompt design in a systematic way, reproducible implementation environments, and performance measures such as accuracy, precision, recall, and F1-score. The results suggest that prompt-based LLMs for zero-shot inference of hate speech do not require task-specific training. However, the finetuning BERT baseline is more precise and achieves a high F1-score, suggesting the importance of classification reliability. The findings indicate the pros and cons of prompt engineering, which is feasible but highlights the need for prompt design and model choice to achieve reliable hate speech detectors in climate-related speech.

This study presents an AI-augmented self-interview method that utilizes a large language model (LLM) as both the interviewer and an initial analytic assistant. This approach addresses long-standing challenges in autoethnographic self-study, particularly issues of subjectivity and weak audit trails. The primary contribution is a standardized and reproducible workflow that outlines interviewer prompts, turn-taking rules, audit-trail artifacts, and a human adjudication stage. This structure helps restore organization and reflective distance in self-research. In a proof-of-concept case, the workflow generated stable, quote-anchored themes and an explicit codebook with traceable interpretive moves. While we do not claim that this method is superior to human-led interviews, we provide evidence for procedural objectivity-defined by transparency and traceability-as well as reliability indicators such as short-interval test-retest stability and alignment between LLM-generated codes and human adjudication. Additionally, we propose a pre-registered design for a controlled comparison between human and LLM interviews, and we provide prompts and templates for others to reuse. Overall, this work positions LLMs as methodological supports rather than replacements, clarifying what is innovative, what is currently achievable, and what requires further validation.

This study proposes artificial intelligence (AI) technology for the automatic classification of Korean scientific and technical papers, aiming to achieve high accuracy even with a small amount of labeled data. Unlike existing BERT-based Korean document classification models that perform supervised learning based on a large amount of accurately labeled data, this study proposes a structure that utilize large language models (LLMs) and retrieval-augmented generation (RAG) technology. The proposed method experimentally demonstrates that it can achieve higher accuracy than existing technologies across all cases using various amounts of labeled data. Furthermore, a qualitative comparison between manually-generated labels, and recognized as correct answers and those produced by LLM responses confirmed that the LLM responses were more accurate. The findings of this study, while limited to Korean scientific documents, provide evidence that a system utilizing LLM and RAG for document classification can easily be extended to other domains with diverse document datasets, owing to its effectiveness even with limited labels.

As the scope of healthcare data expands beyond hospital-generated electronic medical records (EMRs) to include personal health records (PHRs), there is a growing need for automated methods to efficiently summarize large volumes of patient-specific information. In this study, we propose a summarization approach that leverages large language models (LLMs) and standardized data formats to improve accessibility and usability of patient data. Specifically, we developed a prototype system that summarizes patient Bundles formatted in accordance with the Fast Healthcare Interoperability Resources (FHIR) standard. Using the ChatGPT API and document processing techniques, we generated summaries and evaluated their accuracy using a checklist based on clinical criteria. The summarization model achieved an accuracy of 81.6%, suggesting its potential for real-world application. Our findings indicate that healthcare professionals can more quickly and effectively review a patient's primary conditions using summarized PHR data, particularly as FHIR adoption increases. However, the results also highlight certain limitations, including the generalization of summaries and the absence of domain-specific fine-tuning. These findings underscore the importance of future research involving multidisciplinary clinical evaluations, targeted fine-tuning strategies, and question-driven summarization to enhance accuracy and clinical relevance. Overall, this study demonstrates the feasibility of integrating LLM-based summarization into healthcare workflows, contributing to improved interoperability and decision-making in clinical settings.

The enactment of Republic Act 11106 establishes Filipino Sign Language (FSL) as the primary mode of communication within the deaf-mute community in the Philippines. However, this legal recognition has highlighted a significant communication gap between the deaf-mute and non-deaf-mute populations, as the latter typically does not understand FSL. This study introduces a mobile application designed to bridge this gap by translating FSL gestures into textual sentences. The application leverages a CNN-BiLSTM deep learning architecture integrated with Mistral 7B, a state-of-the-art Large Language Model (LLM), to recognize continuous multi-sign gestures and translate them into coherent text. To evaluate the system’s effectiveness, two gesture recognition models were compared based on Word Error Rate (WER), calculated using the Levenshtein distance to measure word-level discrepancies. The 1080p30 model with a stride of 5 and a window size of 30 frames achieved a WER of 27.02%, while the 720p60 model achieved, with a stride of 5 and a window size of 60 frames, a WER of 43.37%. The superior performance of the 1080p30 model is attributed to its higher spatial resolution. This research addresses the critical need for accessible communication tools, offering a solution that enhances inclusivity for the Filipino deaf community.

대형언어모델(LLM)은 AI 기술의 일종으로, 인간의 언어를 이해하고 생성하는 능력을 갖추고 있으며, 이를 통해 학습자에게 맞춤형 학습 지원을 제공할 수 있다. 본 연구에서는 경력전환교육 분야에서 LLM의 정의를 명확히 하고, 이를 활용한 교육적 가능성과 효과를 체계적으로 분석하고자 한다. 이를 위해 PRISMA(Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 체계적 문헌고찰 방법론을 활용하여 2015년부터 2024년까지 Google Scholar, Scopus, JSTOR, ProQuest, DOAJ와 같은 주요 데이터베이스에서 수집된 영 어권 문헌과 국내 연구를 검토하였다. 연구는 LLM의 교육적 활용 가능성과 경력전환교육 분야에서의 역할 및 실증적 효과를 중심으로 분석하였다. 연구 결과는 다음 세 가지로 요약된다. 첫째, LLM은 개인화된 학습 환경 설계와 실시간 피드백 제공을 통해 학습 동기를 촉진하고, 학습자 중심의 맞춤형 학습 지원을 가능하게 하는 등 교육적 활용 가능성을 보 여준다. 둘째, LLM의 교육적 활용에 대한 실증적 연구를 통해, 창업 아이디어 개발, 시장 분석, 복잡한 문제 해결 등에서 효과적인 학습 결과가 도출되었으며, 경력전환교육에 적용 할 가능성을 확인하였다. 셋째, 경력전환교육 분야에서 LLM은 창업 지원, 팀워크 훈련, 갈 등 관리와 같은 직무 역량 개발과 학습자 중심의 교육을 지원하는 효과적인 도구로 활용 될 수 있음을 보여준다. 본 연구는 LLM의 교육적 활용 가능성과 실질적 효과를 탐구하여, 경력전환교육의 학습 환경을 혁신하고 창업 및 직업 역량 개발을 체계적으로 지원하는 데 기여한다. 이는 디지털 전환 시대의 요구에 부응하는 경력전환교육 패러다임 구축의 기반 을 마련할 것으로 기대된다.

This study systematically analyzes how Large Language Models(LLMs) can be utilized in career transition education using the PRISMA(Preferred Reporting Items for Systematic Reviews and Meta-Analyses) methodology. The research examines English-language literature collected from major databases such as Google Scholar, Scopus, JSTOR, ProQuest, and DOAJ, as well as Korean literature from domestic databases like RISS, spanning the years 2015 to 2024. The findings of the study are summarized as follows: First, LLMs demonstrate potential for educational applications by fostering learning motivation and enabling personalized, learner-centered education through the design of tailored learning environments and real-time feedback. Second, in the field of career transition education, LLMs can serve as critical tools for practical learning and the development of job competencies, including entrepreneurship support, teamwork training, and conflict management. Third, an empirical analysis of the effects of LLMs reveals positive learning outcomes in areas such as entrepreneurial idea development, market exploration, and complex problem-solving. By concurrently analyzing global and domestic literature, this study validates the educational potential and practical impact of LLMs, contributing to the establishment of a new paradigm for career transition education in the era of digital transformation.

Large language models (LLMs) have demonstrated remarkable capabilities across a range of tasks, from complex reasoning to creative writing. Yet, these systems remain prone to hallucination, rely entirely on static training data, and risk homogenizing scientific thought through mode collapse. Retrieval-augmented generation (RAG) addresses these limitations by conditioning model outputs on dynamically retrieved external knowledge, tethering the immense generative capacity of LLMs to verifiable, human-curated corpora. This survey traces the development of RAG from its foundational formulations as a simple hallucination- mitigating pipeline to contemporary, highly adaptive architectures. We examine the core components—retrieval, re-ranking, and grounded generation—and survey advanced techniques including active retrieval, hierarchical memory, and multimodal extensions. Crucially, we highlight how LLM is evolving beyond a mere fact-checking mechanism into a collaborative, creative thinking partner. By feeding pluralistic evidence into the generation loop rather than relying on monolithic statistical patterns, multi-layered agentic RAG frameworks actively mitigate systemic biases, safeguard intellectual diversity, and amplify human reasoning. Ultimately, RAG represents an essential epistemological bridge for human-AI symbiosis, empowering users to make fairer and balanced decisions and accelerating the realization of co-prosperity in future society.

We investigate whether Large Language Model (LLM)s can learn strategic reasoning and social deception abilities through Reinforcement Learning (RL) finetuning via a multi-agent “Mafia Game” simulation environment. We finetune a baseline 7B model using Proximal Policy Optimization (PPO) with sparse binary rewards based on game outcomes. Training samples are collected through an opponent pool consisting of different versions of the finetuned model. Our experiment results show that the finetuned model outperforms the baseline model by a significant margin and suggest that strategic capabilities unseen in baseline models emerge.

This study investigates how the Korean modality expression -ul-swu-iss with -ess attached is interpreted by 54 Korean speakers and three variants of the large language model [LLM] ChatGPT: GPT3.5-turbo, GPT4, and GPT4o. The results show a discrepancy between the way in which humans and LLMs interpret this expression. Human participants demonstrate a clear preference for the uncertain interpretation for -ess-ul-swu-iss and the unactualized interpretation for -ul-swu-iss-ess with the desirability of the event modulating the preference, but none of the models capture this nuanced distinction in meaning. The human results are discussed based on Condoravdi(2002), such that the past perspective in -ul-swu-iss-ess derives the counterfactual implicature from the unactualized result whereas the present perspective in -ess-ul-swu-iss derives the epistemic reading from the speaker’s uncertainty about the result. The results from these LLMs indicate that they are not yet able to interpret these expressions as humans do.

대규모 언어모델(LLM)은 높은 문맥 이해 능력과 추론, 데이터 생성 능력을 바탕으로 보안 데이터 분석에 유용한 도구가 될 수 있으나, 블랙박스 특성과 환각 문제, 높은 연산 비용 등으로 인해 정확성과 신뢰성이 중요한 보안 과제에 이를 적용하는 것이 쉽지 않다. 본 연구는 중형 로컬 LLM에 검색 증강 생성(RAG) 기반 퓨샷 학습을 적용하여 CVE 설명으로부터 CVSS 3.1 벡터를 자동 분류하는 사례를 소개하여 보안 데이터 분류에 대규모 언어 모델을 활용하는 실제적 가이드를 제공한다. 실험을 통해 RAG 기반 퓨샷 학습만으로도 미세조정된 대형 LLM이나 BERT 모델과 유사한 성능을 비용 효율적으로 달성할 수 있음을 보였으며. 상세 분석을 바탕으로 보안 데이터 분류에 대규모 언어 모델을 활용하는 방향을 제시하였다.

Large Language Models (LLMs) can be valuable tools for security data analysis due to their strong capabilities in context understanding, reasoning, and data generation. However, applying them to security tasks where accuracy and reliability are critical is challenging because of issues like their black-box nature, hallucination problems, and high computational costs. This study presents a case demonstrating the automatic generation of CVSS 3.1 vectors from CVE descriptions by applyingRetrieval-Augmented Generation (RAG)-based few-shot learning to a mid-sized local LLM, providing a practical guide for utilizing large language models in security data classification. Experiments showed that RAG-based few-shot learning alone can achieve performance similar to fine-tuned large LLMs or BERT models in a cost-effective manner. Based on detailedanalysis, we propose directions for leveraging large language models in security data classification.

대규모 언어 모델(large language model, LLM)은 비정형 문서 해석과 복합 분석 절차 구성 능력을바탕으로 지리정보시스템(geographic information system, GIS) 기반 재생에너지 모델링의 새로운방법론으로 주목받고 있다. 그러나 기존 연구는 GIS 또는 에너지 시스템 분야에 개별적으로 초점을 두어, GIS 기반 재생에너지 모델링 관점에서 LLM의 역할을 종합적으로 검토한 연구는 제한적이다. 본 연구에서는 GIS 기반 재생에너지 모델링에서 LLM의 활용을 프롬프팅, 검색증강생성, 도메인 특화 학습 및 모델 적응, 도구 활용 및 LLM 기반 에이전트의 네 범주로 구분하고, 각 접근의적용 방식, 기존 방법 대비 개선점, 한계 및 향후 연구 방향을 검토하였다. 이를 통해 LLM이 GIS 기반 재생에너지 분석의 단계를 해석·연계·통합하는 지능형 분석 계층으로서 갖는 가능성과 과제를 정리한다.

Large language models (LLMs), with their ability to interpret unstructured documents and construct complex analytical procedures, are gaining attention as a new methodology for geographic informa tion system (GIS)-based renewable energy modeling. However, existing studies have focused on either GIS or the energy systems domain, and comprehensive reviews on the role of LLMs from the perspective of GIS-based renewable energy modeling remain limited. This review classifies the use of LLMs in GIS-based renewable energy modeling into four categories?prompting, retrieval- augmented generation, domain-specific learning and model adaptation, and tool-using and LLM- based agents?and examines, for each approach, its mode of application, improvements over conven tional methods, limitations, and future research directions. This review highlights both the potential and challenges of LLMs as an analytical and interpretive layer that integrates the stages of GIS-based renewable energy analysis.

ChatGPT의 공개 이후 연구자들이 LLM을 적극적으로 활용하면서, 생성형 AI가 학술 글쓰기 과정에서 수행할 수 있는 역할에 대한 관심이 높아지고 있다. 이에 본 연구는 논문의 핵심 내용을 요약하는 초록 작성 단계에서 AI 활용 가능성과 한계를 살펴보고자 하였다. 이를 위해 2022년부터 2024년까지 '한국문헌정보학회지'에 출판된 논문 204편을 대상으로 하여, ChatGPT가 생성한 초록과 저자가 작성한 원초록을 비교·분석하였다. 원초록과 생성초록, 본문, 그리고 프롬프트 유형별 유사도를 정량적으로 비교하고 전문가 인식 조사를 병행하였다. 분석 결과, 원초록과 생성초록은 BERT와 TF-IDF 기준에서 의미적으로 유사했지만, 표현과 어휘 사용에서 차이를 보였다. 또한 한국어 프롬프트로 생성된 초록이 원초록 및 본문과의 유사도가 가장 높게 나타나, 프롬프트 언어가 생성초록의 표현과 내용 반영 정도에 영향을 미친 것으로 나타났다. 전문가들은 LLM을 문체 개선이나 표현 보완에 유용한 보조적 도구로 인식하였다. 본 연구는 이러한 결과를 토대로 인간과 AI가 상호 보완적으로 참여하는 협업적 초록 작성의 가능성을 제시하였다.

Since the release of ChatGPT, researchers have shown growing interest in using large language models (LLMs) for academic writing. This study explored the possibilities and limitations of AI use in summarizing the core content of research papers into abstracts. We analyzed 204 papers published in the Journal of the Korean Society for Library and Information Science (2022-2024) by comparing author-written abstracts with those generated by ChatGPT. Quantitative analyses examined similarities among the originals, AI-generated abstracts, full texts, and different prompt types, supplemented by expert perception. Results showed that AI-generated abstracts were semantically close to the originals based on BERT and TF-IDF scores but differed in word choice and expression. Abstracts generated with Korean prompts showed the highest similarity to both the originals and the full texts, indicating that the prompt language affected style and content representation. Experts viewed LLMs as helpful tools for improving clarity and fluency in writing. Overall, the findings suggest the potential of LLMs as collaborative partners in abstract writing.

최근 거대 언어 모델을 기반으로 한 다양한 인공지능 챗봇이 출시되고 있다. 챗봇은 대화형 프롬프트를 통해 사용자에게 빠르고 간편하게 정보를 제공할 수 있다는 이점을 가지고 있어서 질의응답, 글쓰기, 프로그래밍 등 다양한 분야에서 활용되고 있다. 그러나 최근에는 챗봇의 취약점을 악용하는 '프롬프트 주입 공격'이 제안되었는데, 이는 챗봇이 기입력된 지시사항을 위반하도록 하는 공격이다. 이와 같은 공격은 거대 언어 모델 내부의 기밀 정보를 유출하거나 또 다른 악성 행위를 유발할 수 있어서 치명적이다. 반면 이들에 대한 취약점 여부가 한국어 프롬프트를 대상으로는 충분히 검증되지 않았다. 따라서 본 논문에서는 널리 사용되는 챗봇인 ChatGPT를 대상으로 악성 한국어 프롬프트를 생성하여 공격을 수행해보고, 이들에 대한 실행 가능성을 분석하고자 한다. 이를 위해 기존에 제안된 프롬프트 주입 공격 기법을 분석하여 악의적인 한국어 프롬프트를 자동으로 생성하는 시스템을 제안하고자 한다. 특히 유해 표현을 유도하는 악성 프롬프트를 중점적으로 생성하였고 이들이 실제 유효함을 보이도록 한다.

Recently, various AI chatbots based on large language models have been released. Chatbots have the advantage of providing users with quick and easy information through interactive prompts, making them useful in various fields such as question answering, writing, and programming. However, a vulnerability in chatbots called "prompt injection attacks" has been proposed. This attack involves injecting instructions into the chatbot to violate predefined guidelines. Such attacks can be critical as they may lead to the leakage of confidential information within large language models or trigger other malicious activities. However, the vulnerability of Korean prompts has not been adequately validated. Therefore, in this paper, we aim to generate malicious Korean prompts and perform attacks on the popular chatbot to analyze their feasibility. To achieve this, we propose a system that automatically generates malicious Korean prompts by analyzing existing prompt injection attacks. Specifically, we focus on generating malicious prompts that induce harmful expressions from large language models and validate their effectiveness in practice.