How Language Models Organize and Structure Moral Knowledge

Summary

This paper (arXiv:2608.27402, Orion Reblitz-Richardson, 2026-08-27) investigates how large language models internally organize moral knowledge beyond simple moral-content detection. The author trains six independent linear probes on open-weight language models, one for each Moral Foundations Theory (MFT) category: care/harm, fairness/cheating, liberty/oppression, loyalty/betrayal, authority/subversion, and sanctity/degradation, then analyzes the geometric relationships between the recovered directions in representation space. The key finding is that these directions neither collapse into a single generic moral detector nor remain fully isolated. Instead, they span a near-maximal number of independent dimensions while sharing a positive common component. This shared component appears to be a signature of moral integration: it is morality-specific relative to matched non-moral concept batteries, with mean pairwise cosine similarity of 0.26 versus 0.013. The geometry is consistent across architectures and model scales and reaches its integrated state early in pretraining. The work offers a mechanistic view of how moral foundations are structurally encoded in LLM representations.