An Old Question in Psychology: Is Morality Unified or Categorical?
Jonathan Haidt's Moral Foundations Theory (MFT) proposes six independent moral "taste receptors": care/harm, fairness/cheating, liberty/oppression, loyalty/betrayal, authority/subversion, and sanctity/degradation. MFT groups them into:
- Individualizing foundations: care, fairness, liberty — protecting individual rights
- Binding foundations: loyalty, authority, sanctity — maintaining group order
- Paper: https://arxiv.org/abs/2608.27402
- HTML version: https://arxiv.org/html/2608.27402v1
- Code: https://github.com/orionrr/moral-geometry
- License: CC BY 4.0
This dichotomy has survey-data support in psychology. But do large language models, trained on trillions of tokens, organize moral knowledge according to MFT's structure — or something else?
Method: Six Linear Probes + Geometric Analysis
The paper's approach is straightforward:
1. Train six independent linear probes on OLMo-2 1B and 7B, plus OLMoE-1B-7B, one per MFT foundation 2. Each probe is a direction vector indicating where that foundation points in the model's representation space 3. Examine the geometric relations among the six directions
"Collapse" would mean all six merge into one generic moral detector; "separation" would mean mutually orthogonal directions. The actual result is a third pattern: integration.
Integration: Neither Collapsed Nor Separated
The six directions occupy "near-maximally independent dimensions" but share a common positive component. The mean pairwise cosine similarity among the six moral directions is 0.26. As a control, six non-moral concept directions (e.g., "food," "weather") built the same way average only 0.013.
In plain terms: the model knows "care" and "loyalty" are different things, but they resemble each other more than either resembles "food" — because both are moral.
MFT's Dichotomy Does Not Emerge
If MFT's individualizing/binding split were reflected in the model, hierarchical clustering of the six directions should recover it. Using multiple clustering methods, the paper never stably recovers the two-group structure. Instead, the model's moral geometry reflects the statistical structure of the training corpus — the co-occurrence patterns of words like "care" and "loyalty," not a psychology theory's abstraction.
Where MFT and the model agree (six separately probeable foundations), MFT captures real distinctions in the corpus. Where they diverge (no dichotomy), MFT is the psychologist's abstraction, not the corpus's structure.
Geometry Stabilizes Early in Training
The integration geometry appears early in pretraining, well before probe accuracy saturates. The model learns the geometric structure — six directions with a shared component — long before it can reliably answer moral judgment questions. Geometry precedes capability: the infrastructure is built before the application.
Moral Dilemmas: Partial, Structured Composition
Beyond single-foundation cases, the paper tests moral dilemmas where two foundations conflict (e.g., "he hit someone, but to protect a bullied child"). Dilemma directions are not simple sums of the two foundation directions, nor entirely novel — they are partially composed: foundation directions contribute part, but dilemmas retain unique components.
This partial composition is more pronounced in the MoE architecture (OLMoE-1B-7B) than the dense OLMo-2 1B — possibly because sparse activation lets different experts handle different foundations, enabling more flexible composition.
Complexity–Fragility Gradient
More complex moral concepts are more fragile: single-foundation directions are most stable under bootstrap resampling; dilemma directions are less stable. Counterintuitively, fragility cannot be assigned per foundation — it is context-dependent, not foundation-dependent.
Scale Consistency: Geometry Unchanged from 1B to 7B
The relative arrangement of the six directions, the size of the shared component, and the dendrogram structure are nearly identical at 1B and 7B. Moral geometry is not an emergent property of scale but a statistical regularity that appears given sufficient corpus — larger models only raise probe accuracy, not geometry.
What It Means
One sentence: the model's moral knowledge is integrated, not categorical. Six foundations are neither independent drawers nor one unified detector — they are six overlapping but distinct directions sharing a common "morality" component, learned from corpus statistics rather than taught by psychological theory. To change a model's moral behavior, changing the theory won't help; changing the corpus will.
The author speculates this integration pattern may generalize: emotions, causality, social roles — multi-dimensional concepts may all be integrated in model representations. Geometry precedes capability, integration precedes categorization — possibly a general law of concept organization in large models.