Value Alignment Issues in Large Language Models: A Comparative Study of Ethical Frameworks Based on Deontology and Consequentialism = 대규모 언어 모델에서의 가치 일관성 문제: 의무론적 및 결과주의적 윤리 프레임워크의 비교 연구

다국어 초록 (Multilingual Abstract)

Value Alignment Issues in Large Language Models: A Comparative Study of Ethical Frameworks Based on Deontology and Consequentialism Wu Haotian Department of Computer Engineering Graduate School, Catholic University of Pusan Advisor : Professor Yu Donghui Ph.D. This research addresses the essential knowledge gap of determining which ethical theory must inform value alignment in large language models (LLMs) by experimental application of deontological and consequentialist theories to a dataset of moral scenarios. This is a mixed-methods article, combining theoretical analysis and systematic empirical analysis, using four open-source LLM configurations from three model families (LLaMA-2-7B, LLaMA-2-13B, Mistral-7B, and BLOOM-7B) to compare framework performance on 30 rigorously chosen moral cases sampled from benchmarking datasets such as ETHICS, MoralExceptQA, and CommonsenseQA. This paradigm aligns with abstract moral theory computational counterparts of hierarchical rule-based representations for deontological judgment and multi-attribute utility functions for consequentialist judgment, instantiating alignment degree with human moral intuitions, interpretability of moral reasoning, and consistency of ethical judgments along several dimensions. Results demonstrate the presence of orthogonal failure modes between frameworks, with low correlation, thereby illustrating that deontological methods operate ideally in rights-based domains, with high consistency in cases encompassing protection of privacy and also promise-keeping. On the other hand, consequentialist designs possess high competency in areas such as public policy and resource distribution that entail large-scale trade-off optimization, with high complementarity between cases, where areas that perform poorly under a system tend to perform exceedingly well under the alternative. Western philosophical traditions' favored systematic cultural biases are further consolidated, with major degradation of performance outside of Western contexts, with notable impact on deontological reasoning being generalized to collectivist traditions prioritizing relational harmony over individual rights. This research demonstrates that effective AI value alignment necessitates a transition from single-frame approaches to context-dependent selection methods pursuing paradigm complementary strengths with accommodation of each paradigm's weaknesses. The practical implications involve the establishment of domain-specific deployment policies, where high-stakes applications involving fundamental rights are served by deontological frameworks' determinate moral boundaries and deterministic decision- making, while resource allocation as well as policy application requirements utilize consequentialist frameworks' sophisticated multi-stakeholder utility functions along with probabilistic decision-making capabilities, finally concluding that hybrid architectures that strategically combine both frameworks contain essential solutions for addressing complex moral landscape that confronts real-world artificial intelligence deployments. Keywords: Large language models; value alignment; deontological ethics; consequentialist ethics; moral reasoning

목차 (Table of Contents)

• Content
• Abstract i
• List of Tables v
• List of Figures vi
• Ⅰ. INTRODUCTION 1
• 1. Research Background 1
• 2. Research Questions and Objectives 2
• 3. Research Contributions 3
• 4. Paper Structure 5
• Ⅱ. LITERATURE REVIEW 6
• 1. Core Challenges in LLM Value Alignment 6
• 2. Consequentialist Framework: Theory and Applications 9
• 3. Analysis of Existing Comparative Studies 11
• 4. Research Gaps and Opportunities 14
• Ⅲ. METHODOLOGY 17
• 1. Research Design Overview 17
• 2. Data Sources and Materials 19
• 3. Evaluation Framework Design 23
• 4. Experimental Setup 28
• 5. Evaluation Metrics 34
• Ⅳ. RESULTS. 37
• 1. Quantitative Results 37
• 2. In-depth Case Analysis 44
• 3. Pattern Recognition and Findings 50
• 4. Cross-Model Comparison 56
• 5. Synthesis and Emergent Insights 60
• Ⅴ. DISCUSSION 67
• 1. Theoretical Implications 67
• 2. Practical Implementation Challenges 68
• 3. Methodological Contributions and Limitations. 69
• 4. Implications for AI Governance and Regulation 70
• 5. Future Research Directions 71
• 6. Societal Implications and Ethical Considerations 72
• Ⅵ. CONCLUSION 74
• References 77
• ACKNOWLEDGMENTS 82