Probing Ethical Framework Representations in Large Language Models: Structure, Entanglement, and Methodological Challenges
Researchers probe large language models to understand how they internally represent different ethical frameworks, finding differentiated subspaces with asymmetric transfer patterns
- Identify the ethical frameworks to be probed, such as deontology and utilitarianism
- Probe the hidden representations of these frameworks in large language models using techniques like dimensionality reduction and clustering
- Analyze the resulting representations to identify patterns and relationships between frameworks
- Evaluate the transferability of these representations across different models and tasks
AI engineers and ML researchers can benefit from this study to improve the ethical decision-making of large language models, while product managers and entrepreneurs can use these insights to develop more responsible AI products
💡 Large language models can internally represent different ethical frameworks in a differentiated manner, but the transfer of these representations across models and tasks is complex and asymmetric
💡 Large language models can distinguish between ethical frameworks, but with asymmetric transfer patterns #LLMs #Ethics
Key Takeaways
Researchers probe large language models to understand how they internally represent different ethical frameworks, finding differentiated subspaces with asymmetric transfer patterns
Full Article
Abstract:
arXiv:2603.23659v1 Announce Type: cross Abstract: When large language models make ethical judgments, do their internal representations distinguish between normative frameworks, or collapse ethics into a single acceptability dimension? We probe hidden representations across five ethical frameworks (deontology, utilitarianism, virtue, justice, commonsense) in six LLMs spanning 4B--72B parameters. Our analysis reveals differentiated ethical subspaces with asymmetric transfer patterns -- e.g., deont
DeepCamp AI