Selected work

Representative Papers

MOSAIC: Unveiling the Moral, Social and Individual Dimensions of Large Language Models

Feb 09, 2026arXiv.org

This work addresses the limitations of current ethical evaluations of large language models, which predominantly rely on moral foundations theory while overlooking critical dimensions such as social values and personality traits that shape human moral judgment. To bridge this gap, the authors introduce MOSAIC, the first large-scale, multidimensional ethical benchmark that integrates nine standardized scales from moral philosophy, psychology, and social theory, along with four contextualized game-theoretic tasks, to form an extensible and ready-to-use evaluation framework. Experiments across three mainstream models demonstrate that moral foundations alone are insufficient for comprehensively characterizing AI ethical behavior. The project further contributes an open-source dataset and a Python evaluation library to advance more holistic and human-aligned assessments of AI ethics.

2 citationsRead paper

Automated Compliance Mapping in Cloud Security with Domain-Adapted Sentence Transformers

Jul 07, 2026

This study addresses the inefficiency and error-proneness of manual mapping between cloud security controls and technical specifications. To overcome this limitation, the authors propose the first application of a domain-adapted Sentence Transformer model to cloud compliance semantic matching. High-quality, multi-standard training corpora are generated using back-translation and large language models, followed by task-specific fine-tuning on both control-to-metric matching and cross-standard linkage tasks. Experimental results demonstrate that the best-performing model achieves a 23-point improvement in nDCG@10 for control-to-metric matching and attains a cross-standard association performance of 0.870, confirming the critical role of domain adaptation and data augmentation in enhancing semantic matching effectiveness for cloud compliance.

0 citationsRead paper

Auditing LLM Editorial Bias in News Media Exposure

Oct 31, 2025

This study addresses the underexplored issue of *agentic editorial bias*—systematic, implicit information curation—when large language models (LLMs) function as news gatekeepers. Method: We conduct the first systematic audit of four state-of-the-art LLMs (GPT-4o-Mini, Claude-3.7-Sonnet, Gemini-2.0-Flash) against Google News, employing a multi-layered algorithmic framework integrating topic-based querying, media outlet classification, ideological positioning, and factual accuracy assessment—rigorously validated across diverse prompting strategies and reliability benchmarks. Results: All LLMs exhibit statistically significant, robust ideological skew and uneven attention allocation: they amplify ideologically aligned outlets while suppressing others, yielding lower media diversity and narrower exposure sets than conventional news aggregators. Crucially, models differ markedly in directional bias. We introduce the concept of *agentic editorial policy* to formalize LLMs’ latent, systemic filtering mechanisms—revealing their emergent role as high-stakes news intermediaries with substantial information manipulation potential. This work provides foundational empirical evidence and a theoretical framework for LLM content governance.

0 citationsRead paper

The Urban Impact of AI: Modeling Feedback Loops in Next-Venue Recommendation

Apr 10, 2025

This study investigates how next-generation location recommendation systems—via human-AI feedback loops—reshape individual mobility patterns and exacerbate collective spatial inequality and degraded accessibility. We develop a multi-agent simulation framework grounded in real-world GPS trajectories, the first to formally model closed-loop human-AI interaction in place recommendation. Our approach integrates graph neural networks, spatiotemporal trajectory modeling, and counterfactual intervention analysis. Results show that while recommendations increase individual visit diversity, they intensify concentration at popular venues, reduce aggregate spatial accessibility, and reinforce socioeconomic segregation. The core contribution is uncovering the “individual benefit–collective harm” paradox, establishing the first quantifiable paradigm for assessing AI’s urban impact—enabling ethically informed algorithm design and regulatory policy prototyping. (149 words)

0 citationsRead paper

Recent publications

Latest Papers

Automated Compliance Mapping in Cloud Security with Domain-Adapted Sentence Transformers

Jul 07, 2026

This study addresses the inefficiency and error-proneness of manual mapping between cloud security controls and technical specifications. To overcome this limitation, the authors propose the first application of a domain-adapted Sentence Transformer model to cloud compliance semantic matching. High-quality, multi-standard training corpora are generated using back-translation and large language models, followed by task-specific fine-tuning on both control-to-metric matching and cross-standard linkage tasks. Experimental results demonstrate that the best-performing model achieves a 23-point improvement in nDCG@10 for control-to-metric matching and attains a cross-standard association performance of 0.870, confirming the critical role of domain adaptation and data augmentation in enhancing semantic matching effectiveness for cloud compliance.

0 citationsRead paper

MOSAIC: Unveiling the Moral, Social and Individual Dimensions of Large Language Models

Feb 09, 2026arXiv.org

This work addresses the limitations of current ethical evaluations of large language models, which predominantly rely on moral foundations theory while overlooking critical dimensions such as social values and personality traits that shape human moral judgment. To bridge this gap, the authors introduce MOSAIC, the first large-scale, multidimensional ethical benchmark that integrates nine standardized scales from moral philosophy, psychology, and social theory, along with four contextualized game-theoretic tasks, to form an extensible and ready-to-use evaluation framework. Experiments across three mainstream models demonstrate that moral foundations alone are insufficient for comprehensively characterizing AI ethical behavior. The project further contributes an open-source dataset and a Python evaluation library to advance more holistic and human-aligned assessments of AI ethics.

2 citationsRead paper

Auditing LLM Editorial Bias in News Media Exposure

Oct 31, 2025

This study addresses the underexplored issue of *agentic editorial bias*—systematic, implicit information curation—when large language models (LLMs) function as news gatekeepers. Method: We conduct the first systematic audit of four state-of-the-art LLMs (GPT-4o-Mini, Claude-3.7-Sonnet, Gemini-2.0-Flash) against Google News, employing a multi-layered algorithmic framework integrating topic-based querying, media outlet classification, ideological positioning, and factual accuracy assessment—rigorously validated across diverse prompting strategies and reliability benchmarks. Results: All LLMs exhibit statistically significant, robust ideological skew and uneven attention allocation: they amplify ideologically aligned outlets while suppressing others, yielding lower media diversity and narrower exposure sets than conventional news aggregators. Crucially, models differ markedly in directional bias. We introduce the concept of *agentic editorial policy* to formalize LLMs’ latent, systemic filtering mechanisms—revealing their emergent role as high-stakes news intermediaries with substantial information manipulation potential. This work provides foundational empirical evidence and a theoretical framework for LLM content governance.

0 citationsRead paper

The Urban Impact of AI: Modeling Feedback Loops in Next-Venue Recommendation

Apr 10, 2025

This study investigates how next-generation location recommendation systems—via human-AI feedback loops—reshape individual mobility patterns and exacerbate collective spatial inequality and degraded accessibility. We develop a multi-agent simulation framework grounded in real-world GPS trajectories, the first to formally model closed-loop human-AI interaction in place recommendation. Our approach integrates graph neural networks, spatiotemporal trajectory modeling, and counterfactual intervention analysis. Results show that while recommendations increase individual visit diversity, they intensify concentration at popular venues, reduce aggregate spatial accessibility, and reinforce socioeconomic segregation. The core contribution is uncovering the “individual benefit–collective harm” paradox, establishing the first quantifiable paradigm for assessing AI’s urban impact—enabling ethically informed algorithm design and regulatory policy prototyping. (149 words)

0 citationsRead paper