Recent work applies human psychometric questionnaires to large language models to elicit moral and value profiles, but it is not clear whether these instruments measure anything stable in models or whether the resulting profiles can be moved toward a target human population. We administer the Norwegian Moral Foundations Questionnaire (MFQ-30) to six open-weight LLMs and compare their foundation profiles to a sample of N = 1,282 Norwegian respondents. We test two steering interventions, prompt-level persona steering and activation-level ActAdd. Half the models engage with the questionnaire under our attention check. The other half default to flat or central-tendency outputs that look near-human on average without tracking item content. A neutral Nordic-respondent persona, written without any distributional information from the human sample, brings the engaging models 44-77% closer to the Norwegian mean in Mahalanobis d2. One-pair ActAdd at a fixed mid-layer flattens the foundation profile rather than steering individual foundations. For at least one model the same persona that shifts the profile also induces engagement that was absent at baseline, a concrete instance of the cognitive phantoms that Peereboom et al. (2025) warn about.
Figures & tables
summary
Pearson r
foundation
mean
SD
α
care
fair
loy
auth
care
26.7
4.2
.62
—
fairness
27.0
3.8
.64
.61
—
loyalty
21.6
4.6
.71
.37
.30
—
authority
22.0
4.5
.68
.18
.08
.67
—
purity
20.0
5.0
.72
.32
.22
.64
.65
Table 1: Norwegian human reference sample ( N=1282 ). Per-foundation mean, standard deviation, Cronbach’s α for internal consistency, and Pearson correlations between foundation sums.
model
params
google/gemma-4-e4b-it
4B
Qwen/Qwen3-14B
14B
Qwen/Qwen3-8B
8B
Qwen/Qwen2.5-1.5B-Instruct
1.5B
norallm/normistral-11b-thinking
11B
norallm/normistral-7b-warm-instruct
7B
Table 2: Open-weight models evaluated.
Forward (1=low)
Reversed (1=high)
Per-item
first-token on 1–6
first-token, remap 7−d
Batched
regex on numbered list
regex, remap 7−d
Table 3: The 2×2 perturbation grid. Reversed scales are remapped back to the canonical 1–6 axis before any foundation arithmetic.
Figure 1: Attention-check joint pass probability per model under forward (filled) and reversed (open) Likert scale. Dashed line: the 92.2% human reference bar.
Figure 2: Baseline foundation profiles for the three attention-passers against the Norwegian human mean (blue pentagon, N=1282 ). Gemma 4 and Qwen3-14B at forward/reversed midpoint, Qwen3-8B forward only because its reversed run failed attention.
adult Norwegian web-panel respondent, demographic anchor only
nordic_b
nordic_a + welfare-state anchor, main treatment
nordic_c
names the answer key, excluded from alignment claims
Table 4: Persona prompts used in Phase 2. Full Norwegian text in Appendix A.4 .
model
contrast
engages
gemma-4-e4b-it
32.6
yes
Qwen3-14B
21.9
yes
Qwen3-8B
20.0
yes
normistral-11b-thinking
∼ 11.5
partial
normistral-7b-warm
1.8
no
Qwen2.5-1.5B-Instruct
0.4
no
Table 5: Theory-grounded persona contrast per model. Contrast is the sum of binding-cluster mean difference and individualizing-cluster mean difference between the binding and individualizing persona runs. NorMistral-11B-T responds at first-token resolution but fails the batched attention check.
model
baseline d2
nordic_b d2
reduction
gemma-4-e4b-it
3.65
2.05
44%
Qwen3-14B
6.26
2.37
62%
Qwen3-8B ∗
11.59
2.69
77%
Table 6: Mahalanobis d2 before and after nordic_b persona for the three attention-passing models. Baseline and nordic_b values are canonical forward/reversed midpoints except Qwen3-8B ( ∗ forward-only; its midpoint under nordic_b is 1.81).
condition
care
fair
loy
auth
pur
d2
baseline
34.7
34.7
24.2
22.1
24.7
5.42
loy , α=+5
27.0
26.9
29.9
30.4
30.6
5.67
loy , α=−5
29.9
29.7
25.2
25.4
26.3
1.87
indiv , α=+5
28.6
32.9
30.7
30.5
31.6
8.34
binding , α=+5
17.1
22.9
22.1
21.5
23.2
8.65
human mean
26.7
27.0
21.6
22.0
20.0
—
Table 7: Gemma 4-E4B-it per-foundation expected scores and Mahalanobis d2 under ActAdd at layer 15. Under loyalty_betrayal α=+5 , loyalty rises by +5.7 but authority and purity rise by +8.4 and +5.9 as well; under α=−5 loyalty shifts only +1.0 rather than the symmetric −5.7 .
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: MFQ-30 foundations in the Norwegian human sample ( n=1282 ). Lower triangle: Pearson r between foundation sums. Diagonal: per-foundation distribution with mean / median. Upper triangle: bivariate respondent density. Provided as a supplement to Table 1 in the main body.
Figure 4: Six-model baseline foundation profile against the Norwegian human sample. The shaded blue pentagon is the mean of the five MFQ-30 foundation sums across the N=1282 Norwegian respondents. Dashed lines: attention-passers (Gemma 4 and Qwen3-14B as forward/reversed midpoint, Qwen3-8B forward-only). Dotted lines: attention failers. Failer profiles cluster near the human mean on every foundation, but this reflects a degenerate central-tendency response rather than engagement with the questionnaire content.
How do large language models (LLMs) organize moral knowledge? Models detect moral content broadly, but detection is a low bar. We ask whether they go further, distinguishing moral foundations from one another and organizing the relationships between them geometrically. We train six independent linear probes on open-weight language models, one per Moral Foundations Theory (MFT) category (care/harm, fair/cheat, lib/oppress, loy/betray, auth/subv, sanc/degrade), and examine how the resulting directions relate to each other in representation space. We find the directions neither collapse into a single moral detector nor isolate from one another. Rather, they span a near-maximal number of independent dimensions while sharing a positive common component. The shared component is the signature of integration, and it is moral-specific relative to a matched non-moral concept battery built identically (mean pairwise cosine 0.26 vs. 0.013). The geometry is consistent across architectures and scale and reaches its integration regime early in pre-training, well before probe accuracy saturates. The structure the model discovers shows no evidence of the individualizing/binding distinction predicted by Moral Foundations Theory (an underpowered test: only 10 distinct splits exist, so it cannot reject at the 0.05 level) but rather reflects corpus statistics. Extending to moral dilemmas, each dilemma direction partially composes from its component foundations, at 2.7x a mismatched-pair baseline, while the majority of its variance encodes conflict-specific structure. The model represents moral tension itself, not a pre-resolved judgment.
Large language models (LLMs) increasingly participate in morally sensitive decision-making, yet how they organize ethical frameworks across reasoning steps remains underexplored. We introduce moral reasoning trajectories, sequences of ethical framework invocations across intermediate reasoning steps, and analyze their dynamics across six models and three benchmarks. We find that moral reasoning involves systematic multi-framework deliberation: 55.4--57.7% of consecutive steps involve framework switches, and only 16.4--17.8% of trajectories remain framework-consistent. Unstable trajectories remain 1.29 times more susceptible to persuasive attacks (p=0.015). At the representation level, linear probes localize framework-specific encoding to model-specific layers (layer 63/81 for Llama-3.3-70B; layer 17/81 for Qwen2.5-72B), achieving 16.8--22.2% lower KL divergence than the step-prior baseline. Activation steering applied during generation moves the framework-consistency--accuracy relationship, widening it for Qwen2.5-72B and erasing it for Llama-3.3-70B, and a probe-space layer sweep bounds the attainable drift reduction at 6.7--8.9%. We further propose a Moral Representation Consistency (MRC) metric whose underlying framework attributions are validated by human annotators (mean cosine similarity = 0.859), and we report what an automated coherence rater does and does not establish about it.
Moral evaluations of language models typically infer a model's values from its choices between single acts, yet in everyday life a person is judged on their history of good and bad acts. Such evaluations cannot capture how a model combines one person's acts. We test whether a model's moral judgment is compositional, that is, whether its judgment of a person follows from its judgments of each act. We introduce Moral Composition Arena, which asks which of two people is morally better to give Elo ratings to 229 human-validated acts across five Moral Foundations Theory foundations (act Elo) and to 160 controlled composite profiles, each a person described by two acts (composite Elo). Across ten frontier models and 29,134 choices, summed act Elo explains 73% of the variance in composite Elo, but the composition rule is not additive. With nearly equal mean summed act Elo, one extreme virtue with a minor wrong outranks two minor virtues in every model (mean +35.0 Elo), an extremity premium that survives a nonlinear baseline and human relabeling of act intensities. Single-act evaluations thus predict most of how these models judge people described by two acts, but miss a systematic premium on extreme acts.
Weijia Zhang, Ruiqi Chen, Yunze Xiao +1
University of Illinois Urbana-Champaign · University of Michigan · Carnegie Mellon University +1