Selected work

Representative Papers

Nationality and Region Prediction from Names: A Comparative Study of Neural Models and Large Language Models

Jan 13, 2026

This study addresses the challenge of generalization in name-based nationality prediction, particularly for low-frequency countries and geographically proximate regions. It presents the first systematic comparison of six neural network architectures against six large language model (LLM) prompting strategies across three granularity levels: nationality, region, and continent, employing frequency-stratified sampling and fine-grained error analysis. Results demonstrate that LLMs consistently outperform traditional neural models at all granularities, exhibiting exceptional robustness at the regional level. Notably, simpler machine learning approaches show greater resilience for low-frequency nationalities, while LLMs tend to make “neighborhood” errors—confusing geographically adjacent regions—rather than exhibiting cross-regional bias. The work underscores the importance of evaluating error types and quality beyond aggregate accuracy metrics.

1 citationsRead paper

Diversity Matters: Distributional Feature Coverage Sample Selection for Data-Efficient Backdoor Attacks

Aug 09, 2026

This work addresses the challenge of low-poisoning-rate backdoor attacks, where existing methods struggle to efficiently associate triggers with target labels due to sample redundancy or reliance on task-specific training. To overcome this limitation, the authors propose DFCS—a training-free, trigger-agnostic sample selection strategy that introduces distributional feature coverage as a core principle for low-budget dirty-label backdoor attacks. DFCS leverages fixed features extracted from a pretrained model, partitions the feature space via clustering, and selects samples closest to each cluster centroid to maximize coverage. Evaluated on CIFAR-10, Tiny-ImageNet, and Imagenette under BadNets and Blended attack settings, DFCS achieves an average attack success rate of 96.30%, outperforming the strongest baseline by 4.60 percentage points while preserving clean accuracy.

0 citationsRead paper

Decentralization of Agenda-Setting Power and Domain-Selective Bridging: Algorithm Design Beyond the Echo Chamber Debate

Aug 05, 2026

This study addresses the tendency of conventional recommendation algorithms to over-optimize user engagement, thereby exacerbating information cocoons—a problem further compounded by existing approaches that often neglect human cognitive limitations and lack practical implementability. To counter this, the authors propose the Agenda Democratization Index (ADI) and a Social Information Health (SIH) model, introducing a domain-selective bridging strategy that integrates user engagement with cross-group bridging directly into the algorithmic scoring function. Bridging weights are dynamically adjusted based on the variability of information domains and their collective influence, enabling domain-adaptive optimization of bridging intensity. This work reframes the information cocoon challenge as a computationally tractable engineering design problem. Multi-agent simulations demonstrate that, compared to uniform bridging strategies, the proposed approach effectively enhances the sharing of collective-decision-related information while preserving user experience in interest- and lifestyle-oriented domains.

0 citationsRead paper

A Bures-Wasserstein Formulation of Matrix Decomposition Structural Equation Modeling

Aug 04, 2026

This study addresses the lack of theoretical connection between matrix decomposition structural equation modeling (MDSEM) and traditional covariance-based structural equation modeling, as well as the unclear statistical properties of MDSEM estimators. By constructing a unified loss function, the authors reformulate MDSEM as a minimum discrepancy estimation problem that minimizes the squared Bures–Wasserstein distance between the observed and model-implied covariance matrices. This formulation establishes, for the first time, the equivalence of MDSEM to covariance structure equation modeling within the minimum discrepancy framework. The proposed estimator is shown to be consistent and asymptotically normal, exhibits finite-sample performance comparable to maximum likelihood estimation, achieves confidence interval coverage close to nominal levels, and demonstrates superior numerical stability under small samples or model misspecification.

0 citationsRead paper

The Context Access Divide: Interaction-Level Architecture as a Complementary Dimension of Agentic Inequality

Jul 09, 2026

This study addresses a novel dimension of AI-mediated inequality—termed the Context Access Divide (CAD)—arising from disparities in how users obtain contextual information, specifically between automatic retrieval and manual attachment. Integrating the fan effect from cognitive psychology, the authors develop a probabilistic model demonstrating that in knowledge-intensive tasks, manually attaching context leads to a combinatorial collapse in task success rates as both corpus size and task complexity increase. The work extends analyses of AI inequality beyond individual or organizational levels to the human-AI interaction layer. By leveraging the Model Context Protocol (MCP) within retrieval-augmented generation (RAG) architectures, the study shows that dynamic context retrieval effectively mitigates the exponential decline in success rates associated with larger knowledge bases, thereby uncovering a critical technical root of knowledge work stratification and platform governance.

0 citationsRead paper

Recent publications

Latest Papers

Diversity Matters: Distributional Feature Coverage Sample Selection for Data-Efficient Backdoor Attacks

Aug 09, 2026

This work addresses the challenge of low-poisoning-rate backdoor attacks, where existing methods struggle to efficiently associate triggers with target labels due to sample redundancy or reliance on task-specific training. To overcome this limitation, the authors propose DFCS—a training-free, trigger-agnostic sample selection strategy that introduces distributional feature coverage as a core principle for low-budget dirty-label backdoor attacks. DFCS leverages fixed features extracted from a pretrained model, partitions the feature space via clustering, and selects samples closest to each cluster centroid to maximize coverage. Evaluated on CIFAR-10, Tiny-ImageNet, and Imagenette under BadNets and Blended attack settings, DFCS achieves an average attack success rate of 96.30%, outperforming the strongest baseline by 4.60 percentage points while preserving clean accuracy.

0 citationsRead paper

Decentralization of Agenda-Setting Power and Domain-Selective Bridging: Algorithm Design Beyond the Echo Chamber Debate

Aug 05, 2026

This study addresses the tendency of conventional recommendation algorithms to over-optimize user engagement, thereby exacerbating information cocoons—a problem further compounded by existing approaches that often neglect human cognitive limitations and lack practical implementability. To counter this, the authors propose the Agenda Democratization Index (ADI) and a Social Information Health (SIH) model, introducing a domain-selective bridging strategy that integrates user engagement with cross-group bridging directly into the algorithmic scoring function. Bridging weights are dynamically adjusted based on the variability of information domains and their collective influence, enabling domain-adaptive optimization of bridging intensity. This work reframes the information cocoon challenge as a computationally tractable engineering design problem. Multi-agent simulations demonstrate that, compared to uniform bridging strategies, the proposed approach effectively enhances the sharing of collective-decision-related information while preserving user experience in interest- and lifestyle-oriented domains.

0 citationsRead paper

A Bures-Wasserstein Formulation of Matrix Decomposition Structural Equation Modeling

Aug 04, 2026

This study addresses the lack of theoretical connection between matrix decomposition structural equation modeling (MDSEM) and traditional covariance-based structural equation modeling, as well as the unclear statistical properties of MDSEM estimators. By constructing a unified loss function, the authors reformulate MDSEM as a minimum discrepancy estimation problem that minimizes the squared Bures–Wasserstein distance between the observed and model-implied covariance matrices. This formulation establishes, for the first time, the equivalence of MDSEM to covariance structure equation modeling within the minimum discrepancy framework. The proposed estimator is shown to be consistent and asymptotically normal, exhibits finite-sample performance comparable to maximum likelihood estimation, achieves confidence interval coverage close to nominal levels, and demonstrates superior numerical stability under small samples or model misspecification.

0 citationsRead paper

The Context Access Divide: Interaction-Level Architecture as a Complementary Dimension of Agentic Inequality

Jul 09, 2026

This study addresses a novel dimension of AI-mediated inequality—termed the Context Access Divide (CAD)—arising from disparities in how users obtain contextual information, specifically between automatic retrieval and manual attachment. Integrating the fan effect from cognitive psychology, the authors develop a probabilistic model demonstrating that in knowledge-intensive tasks, manually attaching context leads to a combinatorial collapse in task success rates as both corpus size and task complexity increase. The work extends analyses of AI inequality beyond individual or organizational levels to the human-AI interaction layer. By leveraging the Model Context Protocol (MCP) within retrieval-augmented generation (RAG) architectures, the study shows that dynamic context retrieval effectively mitigates the exponential decline in success rates associated with larger knowledge bases, thereby uncovering a critical technical root of knowledge work stratification and platform governance.

0 citationsRead paper

Who Determines the Meaning of an Emotion? Affective Sovereignty as an Epistemic Consequence of Measurement Limits

Jun 30, 2026

This study addresses the unresolved question of who holds ultimate interpretive authority over the meaning of individual emotional experiences in the context of widely deployed emotion-aware AI systems. Building on the cognitive limitations inherent in emotion measurement, the work models the distribution of meanings across annotator populations, decomposes sources of uncertainty, and analyzes cognitive coverage to distinguish reducible from irreducible components in emotion labeling. It demonstrates that high-confidence system outputs fundamentally fail to capture the incommensurable nature of personal emotional meaning. The paper introduces, for the first time, a normative principle of “affective sovereignty,” advocating that final interpretive authority over one’s emotions be procedurally reserved for the experiencing subject. It further argues that irreducible uncertainty cannot be adequately estimated under realistic annotation scales, thereby establishing affective sovereignty as a foundational ethical and regulatory principle for emotion AI.

0 citationsRead paper