Probing Ethical Framework Representations in Large Language Models: Structure, Entanglement, and Methodological Challenges

TL;DR

This paper investigates how large language models internally represent different ethical frameworks, revealing structured but entangled representations and highlighting methodological challenges in probing ethical reasoning.

Contribution

It introduces a systematic probing approach across multiple ethical frameworks in large language models, uncovering differentiated yet entangled ethical subspaces and discussing interpretability limitations.

Findings

Ethical representations form differentiated subspaces within models.

Probes show asymmetric transfer patterns between ethical frameworks.

Surface features influence probe outcomes, affecting interpretability.

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., deontology probes partially generalize to virtue scenarios while commonsense probes fail catastrophically on justice. Disagreement between deontological and utilitarian probes correlates with higher behavioral entropy across architectures, though this relationship may partly reflect shared sensitivity to scenario difficulty. Post-hoc validation reveals that probes partially depend on surface features of benchmark templates, motivating cautious interpretation. We discuss both the structural insights these methods provide and their epistemological limitations.