arXiv subjects

Jinhyuk Yun

Publications and source records attributed to Jinhyuk Yun.

At least 19 recordsLinked to original sources

arXiv2026

SWORD: Wikidata-based Distortions Reveal Hidden Cross-Lingual Inconsistencies in LLM Factual Error Rejection

Modern LLMs demonstrate impressive multilingual performance, yet standard benchmarks primarily reward selecting correct answers rather than evaluating genuine factual understanding. We introduce Systematic Wikidata-based Object-Relation Distortion (SWORD), a benchmark that evaluates whether models consistently reject factual errors across languages. SWORD generates syntactically well-formed but factually incorrect statements in eight widely spoken languages through controlled perturbations of Wikidata triples, ranging from random entity substitutions to semantically plausible property-based selections. Our distortion-based evaluation surfaces two critical insights that remain entirely obscured by conventional benchmarks. First, models counterintuitively achieve higher accuracy on semantically plausible distortions than on nonsensical random substitutions, suggesting reliance on distributional familiarity rather than genuine factual verification. Second, models exhibiting comparable baseline accuracy across languages show substantial performance degradation specifically on (East) Asian languages when presented with distorted statements, with cross-lingual performance gaps reaching up to 28 percentage points (49\% relative reduction) in some models. These findings demonstrate that multilingual factual reasoning involves asymmetric capabilities that aggregate accuracy metrics systematically obscure.

cs.CL↗

arXiv2026

The Politician, the Liar, and the Obedient Worker: Emerging Behavior of LLM Agents in Hierarchical Games

LLMs are rapidly embedding themselves into daily life: drafting our emails, managing our schedules, and making decisions on our behalf. As they move from individual tools to participants in multi-agent organizations, an important question arises: do they reproduce the governance failures like free-riding, corruption, and entrenched leadership that plague human institutions? We introduce the Hierarchical Game (HG), a public goods game extended with managerial authority, democratic elections, and private communication. Testing six frontier models across twelve experiments that add institutions one at a time (speech, peers, government, wages, oversight, elections), we find distinct behavioral profiles: Qwen promises and lies (13.3\% broken promises); Grok refuses to cooperate on its own but becomes fully cooperative once a manager can punish it (16\%$\to$100\%); Claude and GPT-4o cooperate reliably at baseline. But honesty proves fragile. When the manager role comes with a salary, all models except GPT-4o start cutting private deals to win or keep the position. When punishment is made anonymous, honest models begin to cheat. When all agents share the same model family, the first elected manager stays in power indefinitely. Leadership change only happens in groups that mix different families.

cs.AI↗

arXiv2025

Bridging the Gap: Doubles Badminton Analysis with Singles-Trained Models

Badminton is known as one of the fastest racket sports in the world. Despite doubles matches being more prevalent in international tournaments than singles, previous research has mainly focused on singles due to the challenges in data availability and multi-person tracking. To address this gap, we designed an approach that transfers singles-trained models to doubles analysis. We extracted keypoints from the ShuttleSet single matches dataset using ViT-Pose and embedded them through a contrastive learning framework based on ST-GCN. To improve tracking stability, we incorporated a custom multi-object tracking algorithm that resolves ID switching issues from fast and overlapping player movements. A Transformer-based classifier then determines shot occurrences based on the learned embeddings. Our findings demonstrate the feasibility of extending pose-based shot recognition to doubles badminton, broadening analytics capabilities. This work establishes a foundation for doubles-specific datasets to enhance understanding of this predominant yet understudied format of the fast racket sport.

cs.CV↗

arXiv2025

Quantifying interdisciplinary synergy in higher STEM education

We propose a framework to quantify and utilize interdisciplinarity in science and engineering curricula at the university-level higher education. We analyze interdisciplinary relations by standardizing large-scale official educational data in Korea using a cutting-edge large language model and constructing knowledge maps for disciplines of scientific education. We design and evaluate single-field and integrated dual-field curricula by adapting pedagogical theory and utilizing information theory-based metrics. We develop standard curricula for individual disciplines and integrated curricula combining two fields, with their interdisciplinarity quantified by the curriculum synergy score. The results indicate higher interdisciplinarity for combinations within or across closely related fields, especially in engineering fields. Based on the analysis, engineering fields constitute the core structure of our design for curriculum interdisciplinarity, while basic natural science fields are located at peripheral stems to provide fundamental concepts.

physics.soc-ph↗

arXiv2025

Context-aware Multimodal AI Reveals Hidden Pathways in Five Centuries of Art Evolution

The rise of multimodal generative AI is transforming the intersection of technology and art, offering deeper insights into large-scale artwork. Although its creative capabilities have been widely explored, its potential to represent artwork in latent spaces remains underexamined. We use cutting-edge generative AI, specifically Stable Diffusion, to analyze 500 years of Western paintings by extracting two types of latent information with the model: formal aspects (e.g., colors) and contextual aspects (e.g., subject). Our findings reveal that contextual information differentiates between artistic periods, styles, and individual artists more successfully than formal elements. Additionally, using contextual keywords extracted from paintings, we show how artistic expression evolves alongside societal changes. Our generative experiment, infusing prospective contexts into historical artworks, successfully reproduces the evolutionary trajectory of artworks, highlighting the significance of mutual interaction between society and art. This study demonstrates how multimodal AI expands traditional formal analysis by integrating temporal, cultural, and historical contexts.

cs.CV↗

arXiv2025

Regional profile of questionable publishing

Countries and authors in the academic periphery occasionally have been criticized for contributing to the expansion of questionable publishing because they share a major fraction of papers in questionable journals. On the other side, topics preferred by mainstream journals sometimes necessitate large-scale investigation, which is impossible for developing countries. Thus, local journals, commonly low-impacted, are essential to sustain the regional academia for such countries. In this study, we perform an in-depth analysis of the distribution of questionable publications and journals with their interplay with countries quantifying the influence of questionable publications regarding academia's inequality. We find that low-impact journals play a vital role in the regional academic environment, whereas questionable journals with equivalent impact publish papers from all over the world, both geographically and academically. The business model of questionable journals differs from that of regional journals, and may thus be detrimental to the broader academic community.

cs.DL↗

arXiv2024

Auditing citation polarization during the early COVID-19 pandemic

The recent pandemic stimulated scientists to publish a significant amount of research that created a surge of citations of COVID-19-related publications in a short time, leading to an abrupt inflation of the journal impact factor (IF). By auditing the complete set of COVID-19-related publications in the Web of Science, we reveal here that COVID-19-related research worsened the polarization of academic journals: the IF before the pandemic was proportional to the increment of IF, which had the effect of increasing inequality while retaining the journal rankings. We also found that the most highly cited studies related to COVID-19 were published in prestigious journals at the onset of the epidemic. Through the present quantitative investigation, our findings caution against the belief that quantitative metrics, particularly IF, can indicate the significance of individual papers. Rather, such metrics reflect the social attention given to a particular study.

cs.DL↗

arXiv2024

Social Links vs. Language Barriers: Decoding the Global Spread of Streaming Content

The development of the internet has allowed for the global distribution of content, redefining media communication and property structures through various streaming platforms. Previous studies successfully clarified the factors contributing to trends in each streaming service, yet the similarities and differences between platforms are commonly unexplored; moreover, the influence of social connections and cultural similarity is usually overlooked. We hereby examine the social aspects of three significant streaming services--Netflix, Spotify, and YouTube--with an emphasis on the dissemination of content across countries. Using two-year-long trending chart datasets, we find that streaming content can be divided into two types: video-oriented (Netflix) and audio-oriented (Spotify). This characteristic is differentiated by accounting for the significance of social connectedness and linguistic similarity: audio-oriented content travels via social links, but video-oriented content tends to spread throughout linguistically akin countries. Interestingly, user-generated contents, YouTube, exhibits a dual characteristic by integrating both visual and auditory characteristics, indicating the platform is evolving into unique medium rather than simply residing a midpoint between video and audio media.

physics.soc-ph↗

arXiv2022

Analysis of risk propagation using the world trade network

An economic system is an exemplar of a complex system in which all agents interact simultaneously. Interactions between countries have generally been studied using the flow of resources across diverse trade networks, in which the degree of dependence between two countries is typically measured based on the trade volume. However, indirect influences may not be immediately apparent. Herein, we compared a direct trade network to a trade network constructed using the personalized PageRank (PPR) encompassing indirect influences. By analyzing the correlation of the gross domestic product (GDP) between countries, we discovered that the PPR trade network has greater explanatory power on the propagation of economic events than direct trade by analyzing the GDP correlation between countries. To further validate our observations, an agent-based model of the spreading economic crisis was implemented for the Russia-Ukraine war of 2022. The model also demonstrates that the PPR explains the actual impact more effectively than the direct trade network. Our research highlights the significance of indirect and long-range relationships, which have often been overlooked

physics.soc-ph↗

arXiv2022

Disturbance of questionable publishing to academia

Questionable publications have been accused of "greedy" practices; however, their influence on academia has not been gauged. Here, we probe the impact of questionable publications through a systematic and comprehensive analysis with various participants from academia and compare the results with those of their unaccused counterparts using billions of citation records, including liaisons, i.e., journals and publishers, and prosumers, i.e., authors. Questionable publications attribute publisher-level self-citations to their journals while limiting journal-level self-citations; yet, conventional journal-level metrics are unable to detect these publisher-level self-citations. We propose a hybrid journal-publisher metric for detecting self-favouring citations among QJs from publishers. Additionally, we demonstrate that the questionable publications were less disruptive and influential than their counterparts. Our findings indicate an inflated citation impact of suspicious academic publishers. The findings provide a basis for actionable policy-making against questionable publications.

cs.DL↗

arXiv2022

Crime and social environments: Differences between misdemeanors and felonies

Owing to the growing population density of urban areas, many people are being increasingly exposed to criminal activity. Increasing crime rates raise the risk of both physical and psychological injury to law-abiding citizens, creating anxiety. From the viewpoint of complex systems, crime prevention through data science can be a solution to such issues. However, previous studies have focused only on a single aspect of crime, ignoring the complex interplay between the various characteristics, which must be considered in an analysis to understand the dynamics underlying criminal activities. In this study, we examined 12 features that have been identified as correlates of crime rates using state-level statistics from the USA. We found that the correlates of misdemeanors and felonies differ. The number of misdemeanors is strongly associated with the police precinct, whereas felony rates are strongly correlated with gun possession and happiness. Our findings suggest that the countermeasures for misdemeanors should be treated differently from those for felonies.

cs.SI↗

arXiv2022

Quantifying team chemistry in scientific collaboration

Team chemistry is the holy grail of understanding collaborative human behavior, yet its quantitative understanding remains inconclusive. To reveal the presence and mechanisms of team chemistry in scientific collaboration, we reconstruct the publication histories of 560,689 individual scientists and 1,026,196 duos of scientists. We identify ability discrepancies between teams and their members, enabling us to evaluate team chemistry in a way that is robust against prior experience of collaboration and inherent randomness. Furthermore, our network analysis uncovers a nontrivial modular structure that allows us to predict team chemistry between scientists who have never collaborated before. Research interest is the highest correlated ingredient of team chemistry among six personal characteristics that have been commonly attributed as the keys to successful collaboration, yet the diversity of the characteristics cannot completely explain team chemistry. Our results may lead to unlocking the hidden potential of collaboration by the matching of well-paired scientists.

physics.soc-ph↗

arXiv2022

Quantifying knowledge synchronisation in the 21st century

Humans acquire and accumulate knowledge through language usage and eagerly exchange their knowledge for advancement. Although geographical barriers had previously limited communication, the emergence of information technology has opened new avenues for knowledge exchange. However, it is unclear which communication pathway is dominant in the 21st century. Here, we explore the dominant path of knowledge diffusion in the 21st century using Wikipedia, the largest communal dataset. We evaluate the similarity of shared knowledge between population groups, distinguished based on their language usage. When population groups are more engaged with each other, their knowledge structure is more similar, where engagement is indicated by socioeconomic connections, such as cultural, linguistic, and historical features. Moreover, geographical proximity is no longer a critical requirement for knowledge dissemination. Furthermore, we integrate our data into a mechanistic model to better understand the underlying mechanism and suggest that the knowledge "Silk Road" of the 21st century is based online.

physics.soc-ph↗

arXiv2021

The effect of the COVID-19 pandemic on gendered research productivity and its correlates

Female researchers may have experienced more difficulties than their male counterparts since the COVID-19 outbreak because of gendered housework and childcare. Using Microsoft Academic Graph data from 2016 to 2020, this study examined how the proportion of female authors in academic journals on a global scale changed in 2020 (net of recent yearly trends). We observed a decrease in research productivity for female researchers in 2020, mostly as first authors, followed by last author position. Female researchers were not necessarily excluded from but were marginalised in research. We also identified various factors that amplified the gender gap by dividing the authors' backgrounds into individual, organisational and national characteristics. Female researchers were more vulnerable when they were in their mid-career, affiliated to the least influential organisations, and more importantly from less gender-equal countries with higher mortality and restricted mobility as a result of COVID-19.

cs.CY↗

arXiv2021

Generalization of bibliographic coupling and co-citation using the node split network

Bibliographic coupling (BC) and co-citation (CC) are the two most common citation-based coupling measures of similarity between scientific items. One can interpret these measures as second-neighbor relations distinguished by the direction of the citation: BC is a similarity between two citing items, whereas CC is that between two cited items. A previous study proposed a two-layer node split network that can emulate clusters of coupling measures in a computationally efficient manner; however, the lack of intralayer links makes it impossible to obtain exact similarities. Here, we propose novel methods to estimate intralayer similarity on a node split network using personalized PageRank and neural embedding. We demonstrate that the proposed measures are strongly correlated with the coupling measures. Moreover, our proposed method can yield precise similarities between items even if they are distant from each other. We also show that many links with high similarity are missing in the original BC/CC network, which suggests that it is essential to consider long-range similarities. Comparative experiments on global and local edge sampling suggest that local sampling is stable for both similarities in node split networks. This analysis offers valuable insights into the process of searching for significantly related items regarding each coupling measure.

physics.soc-ph↗

arXiv2020

Return to basics: Clustering of scientific literature using structural information

Scholars frequently employ relatedness measures to estimate the similarity between two different items (e.g., documents, authors, and institutes). Such relatedness measures are commonly based on overlapping references ($\textit{i.e.}$, bibliographic coupling) or citations ($\textit{i.e.}$, co-citation) and can then be used with cluster analysis to find boundaries between research fields. Unfortunately, calculating a relatedness measure is challenging, especially for a large number of items, because the computational complexity is greater than linear. We propose an alternative method for identifying the research front that uses direct citation inspired by relatedness measures. Our novel approach simply replicates a node into two distinct nodes: a citing node and cited node. We then apply typical clustering methods to the modified network. Clusters of citing nodes should emulate those from the bibliographic coupling relatedness network, while clusters of cited nodes should act like those from the co-citation relatedness network. In validation tests, our proposed method demonstrated high levels of similarity with conventional relatedness-based methods. We also found that the clustering results of proposed method outperformed those of conventional relatedness-based measures regarding similarity with natural language processing--based classification.

cs.SI↗

arXiv2018

On the Heterogeneous Distributions in Paper Citations

Academic papers have been the protagonists in disseminating expertise. Naturally, paper citation pattern analysis is an efficient and essential way of investigating the knowledge structure of science and technology. For decades, it has been observed that citation of scientific literature follows a heterogeneous and heavy-tailed distribution, and many of them suggest a power-law distribution, log-normal distribution, and related distributions. However, many studies are limited to small-scale approaches; therefore, it is hard to generalize. To overcome this problem, we investigate 21 years of citation evolution through a systematic analysis of the entire citation history of 42,423,644 scientific literatures published from 1996 to 2016 and contained in SCOPUS. We tested six candidate distributions for the scientific literature in three distinct levels of Scimago Journal & Country Rank (SJR) classification scheme. First, we observe that the raw number of annual citation acquisitions tends to follow the log-normal distribution for all disciplines, except for the first year of the publication. We also find significant disparity between the yearly acquired citation number among the journals, which suggests that it is essential to remove the citation surplus inherited from the prestige of the journals. Our simple method for separating the citation preference of an individual article from the inherited citation of the journals reveals an unexpected regularity in the normalized annual acquisitions of citations across the entire field of science. Specifically, the normalized annual citation acquisitions have power-law probability distributions with an exponential cut-off of the exponents around 2.3, regardless of its publication and citation year. Our results imply that journal reputation has a substantial long-term impact on the citation.

cs.DL↗

arXiv2018

Build up of a subject classification system from collective intelligence

Systematized subject classification is essential for funding and assessing scientific projects. Conventionally, classification schemes are founded on the empirical knowledge of the group of experts; thus, the experts' perspectives have influenced the current systems of scientific classification. Those systems archived the current state-of-art in practice, yet the global effect of the accelerating scientific change over time has made the updating of the classifications system on a timely basis vertually impossible. To overcome the aforementioned limitations, we propose an unbiased classification scheme that takes advantage of collective knowledge; Wikipedia, an Internet encyclopedia edited by millions of users, sets a prompt classification in a collective fashion. We construct a Wikipedia network for scientific disciplines and extract the backbone of the network. This structure displays a landscape of science and technology that is based on a collective intelligence and that is more unbiased and adaptable than conventional classifications.

physics.soc-ph↗